For guidance purposes only. Not comprehensive or legal advice.
Designed for business, CRE, AEC, Tech/AI and Natural Resources professionals as part of the earlier Mainframe, Mini/PC, CADD, Web, Mobile/Cloud revolutions.
Click to expand or collapse, or click the ► after each term to open individually
The broad field of computer science dedicated to building systems that can perform tasks that previously required human intelligence — reasoning, recognizing patterns, understanding language, making decisions. AI is not one thing: it is a family of techniques, some around since the 1950s, others invented in the last five years. When people say "AI" today they almost always mean a specific subset: generative AI powered by large language models. That is the revolution happening now.
Not magic, not conscious. AI does not think. It predicts. It finds patterns in enormous amounts of data and uses those patterns to generate outputs that look like thinking.
The prior revolutions in context. Mainframes automated calculation. PCs put that power on every desk. CAD/CADD replaced the drafting table with a digital one. The web connected everything. AI is different: for the first time the machine handles language, judgment, and synthesis — not just calculation and retrieval. Those of you who rode all three prior waves will recognize the pattern: early adopters build the moat.
For PS clients. AI is already embedded in tools you use: your CRM, email, mapping software. The question is whether you are using it intentionally or just passively.
Generative AI
The specific class of AI that creates new content — text, images, audio, video, code, documents — rather than simply classifying or predicting from existing data. When you type a question into Claude or ChatGPT and get a paragraph back, that paragraph did not exist before you asked. The model generated it, word by word, based on patterns learned from billions of examples. Generative AI is what makes the current moment different from every prior AI wave.
Text generation. Drafting proposals, summarizing contracts, writing due diligence reports, answering complex questions in plain language.
Image generation. Creating property renderings, site concept visuals, marketing graphics — from a text description alone.
Code generation. Writing scripts, automating spreadsheet tasks, building simple applications — without a programmer in the room.
The AEC parallel. CAD did not just digitize the drafting table — it changed what was possible to design. Generative AI does not just speed up writing — it changes what a small team can produce.
Large Language Model (LLM)
The engine inside virtually every AI chatbot and text tool you encounter today. An LLM is a neural network trained on a massive corpus of text — books, websites, code, scientific papers, legal documents — that learns the statistical relationships between words, phrases, sentences, and ideas at extraordinary scale. "Large" refers to the number of parameters: modern frontier models have hundreds of billions. The result is a system that can read, write, summarize, translate, reason, and converse across virtually any domain, in any language.
Claude, ChatGPT, Gemini, Grok are all LLMs — different training data, different fine-tuning, different strengths.
Not a database. An LLM does not look up answers. It generates them from learned patterns. This is why it can be wrong with total confidence — a critical limitation to internalize.
Context window. Every LLM has a limit on how much text it can see at once. Modern models handle 100,000+ words — roughly 500 pages. Older ones struggled with a few pages.
For PS clients. You do not need to understand how an LLM works to use one — just as you did not need to understand raster vs. vector graphics to use AutoCAD. But understanding the basics protects you from over-trusting it.
Neural Network
The foundational architecture underlying modern AI. A neural network is a system of interconnected mathematical nodes — loosely inspired by how neurons in the brain connect — organized in layers. Data goes in one end, passes through many layers of transformation, and a prediction or output comes out the other. "Deep learning" refers to neural networks with many layers. Every LLM, every image generator, every speech recognizer you use today is built on deep neural networks.
Training. A neural network learns by adjusting billions of internal connection weights based on examples — seeing correct and incorrect outputs and gradually improving. This requires enormous computing power.
Inference. Once trained, running the network to produce outputs is called inference. This is what happens when you type a prompt and get a response.
The mainframe parallel. The first computer revolution taught us that mainframes were not magic — they were deterministic logic at scale. Neural networks are statistical logic at scale. Same insight applies: understand the mechanism, and you understand the limitations.
Transformer Architecture
The specific neural network design that made modern LLMs possible, introduced in a landmark 2017 Google paper titled "Attention Is All You Need." Every major commercial AI model — Claude, ChatGPT, Gemini, Copilot — is built on the Transformer. Its key innovation is the self-attention mechanism: when processing your prompt, the model simultaneously weighs every word against every other word, identifying which parts of the input are most relevant to each part of the output. This allows it to handle long complex documents and multi-part instructions in a way no earlier architecture could.
Why it matters for operators. Input structure directly affects output quality. A well-organized document with clear headings and logical flow gives the attention mechanism better signals than a disorganized one with the same information.
Pre-training vs. fine-tuning. The Transformer is pre-trained on massive text corpora, then fine-tuned for specific behavior. Different fine-tuning produces Claude vs. ChatGPT vs. Gemini — same underlying architecture, different strengths and personalities.
The CAD parallel. AutoCAD and MicroStation both ran on computational geometry. The Transformer is the computational geometry of modern AI — different products built on the same foundation.
Token
The unit of text that AI models actually process. Not exactly words — tokens are chunks of characters the model's vocabulary recognizes. "Construction" is one token; "uncharacteristically" is three. On average, one token ≈ 0.75 words in English. Understanding tokens matters because AI models have token limits for both input (context window) and output — and token counts drive API pricing.
Typical equivalents. 1,000 tokens ≈ 750 words ≈ about 1.5 pages of a Word document.
Context window in tokens. Claude's context window is 200,000 tokens — roughly 150,000 words, about 500 pages. GPT-4o handles 128,000 tokens.
Practical implication. You can paste an entire 50-page lender package into Claude and ask questions about the whole document simultaneously. That was not possible with earlier AI tools, and it changes what due diligence looks like.
Context Window
The maximum amount of text — measured in tokens — that an AI model can see and reason about in a single session. Think of it as the model's working memory. Everything inside the context window is available when generating a response: your prompt, uploaded documents, the entire conversation history. Anything outside is invisible. Early LLMs had context windows of a few thousand tokens. Modern frontier models handle 100,000–200,000 tokens — hundreds of pages.
What fits in 200K tokens. An entire contract portfolio. A full project specification. Years of meeting notes. Multiple financial models. Claude can read and reason across all of it simultaneously.
The forgetting problem. In very long conversations, older content can fall out of the window. If key information was shared early in a long session, restate it before relying on it.
For CRE and AEC. Context window size is what makes document-heavy professional workflows tractable — lease abstractions, feasibility studies, permit packages, environmental reports — all processable in one session.
Training Data
The corpus of text, images, code, and other content an AI model learns from during its initial training. For large language models, training data typically includes large swaths of the public internet, books, scientific papers, code repositories, and licensed datasets — measured in trillions of words. The training data determines what the model knows, what biases it has absorbed, and — critically — when its knowledge ends.
Knowledge cutoff. Training data has a cutoff date. A model trained on data through early 2025 does not know about events after that date — no matter how confidently it answers questions about them. Always verify time-sensitive information against live sources.
Garbage in, garbage out — at scale. If training data contains errors, biases, or outdated information, the model reflects those. This is not a software bug; it is a fundamental characteristic of the approach.
Proprietary training. Fine-tuned models can be trained on your firm's documents, templates, and past work — producing outputs calibrated to your specific standards and voice.
Inference vs. Training
Two fundamentally different phases of an AI model's existence. Training is the computationally intensive process of building the model — feeding it enormous datasets, adjusting billions of internal parameters, requiring weeks or months on thousands of specialized chips at a cost of millions to hundreds of millions of dollars. Inference is running the trained model to produce outputs — what happens every time you submit a prompt. Training happens once (or periodically for updates); inference happens billions of times daily.
Why the distinction matters. Training is a capital investment made by AI companies. Inference is the operating cost passed to users via API pricing or subscription fees. When OpenAI or Anthropic charges per token, they are charging for inference compute — not for the training investment already made.
On-premises inference. Organizations with data privacy requirements can run open-source models (Llama, Mistral) on their own servers — paying only infrastructure costs, with no data leaving their environment. The trade-off: smaller models with less capability than frontier commercial systems.
The mainframe parallel. Early computing separated batch processing (training equivalent) from interactive terminals (inference equivalent). The economics were the same: high fixed cost to run the batch; low marginal cost per terminal query once the batch was done.
Fine-Tuning
The process of taking a pre-trained base model and continuing to train it on a smaller, more specific dataset to adjust its behavior for a particular domain, task, or style. Fine-tuning is how a general-purpose LLM becomes a specialized tool — and how AI companies create the different "personalities" of models from similar underlying architectures. It is also how organizations build custom AI applications tuned to their specific standards.
RLHF (Reinforcement Learning from Human Feedback). The most common fine-tuning approach for commercial models. Human raters evaluate model outputs and the model trains to produce preferred outputs — making it more helpful, more honest, and less likely to produce harmful content.
Domain fine-tuning. Train a model on thousands of your firm's past proposals, reports, and deliverables — it learns to produce outputs matching your house style, terminology, and quality standard without being told explicitly each time.
The cost trend. Fine-tuning a large model once required significant AI engineering resources. API fine-tuning products from Anthropic, OpenAI, and Google now make it accessible to organizations without dedicated AI teams.
Hallucination
When an AI model generates information that is factually wrong — stated with complete confidence and no disclaimer. The model is not lying: it is producing a statistically plausible-sounding output that happens to be false. Hallucinations range from minor (a slightly wrong date) to serious (a citation that does not exist, a statistic invented from nothing, a regulatory requirement fabricated). Every AI user encounters hallucinations. The only question is whether you catch them before they reach a client, a lender, or a regulator.
High-risk categories. Statistics and data points. Legal and regulatory citations. Personnel and organizational details. Financial figures. Specific dates and locations.
The confidence problem. AI does not flag hallucinations. It presents invented content in the same tone as verified content. You cannot tell from the writing style which is which — that is what makes it genuinely dangerous in professional contexts.
The fix. Verify anything important against a primary source before acting on it. For professional deliverables, embed verification as a structural step — not an individual afterthought subject to time pressure.
Improving but not solved. Newer models hallucinate less than older ones. RAG (retrieval-augmented generation) with authoritative source documents significantly reduces hallucination on domain-specific questions. Neither eliminates it entirely.
Grounding
The practice of tethering an AI model's outputs to specific, verifiable source documents or data — rather than allowing it to generate from training memory alone. A grounded AI response is one where every claim can be traced to a document you provided. Ungrounded responses draw on training data that may be outdated, incomplete, or simply wrong. Grounding is the primary technical mechanism for reducing hallucination in professional use cases.
How grounding works in practice. You paste a lease, financial report, or specification into the prompt — the model must answer from that document. Or you use a RAG system that retrieves relevant passages from your document library before generating. Either way, the model's output is anchored to your material, not its training data.
The professional standard. For any output that will be included in a client deliverable, lender package, or regulatory filing: ground it. Do not rely on the model's memory of market conditions, regulatory requirements, or comparable transactions. Provide the source; require it to answer from the source.
Grounding vs. fine-tuning. Fine-tuning teaches the model new behavior permanently. Grounding provides context for the current session only. Grounding is the right tool for document-specific professional tasks; fine-tuning is for changing how the model writes or reasons across all tasks.
Retrieval-Augmented Generation (RAG)
A technique that combines an AI language model with a document retrieval system, allowing the model to answer questions using your specific documents rather than relying solely on its training data. When you ask a question, the system first retrieves the most relevant passages from your document library, feeds those passages to the LLM along with your question, and the model answers using both its language capability and your actual documents. Result: accurate, citation-grounded responses from your proprietary data — dramatically reducing hallucination.
Why it matters. Base LLMs know nothing about your firm's specific projects, policies, past work, or client relationships. RAG gives the model access to that institutional knowledge without the cost and complexity of full fine-tuning.
Typical architecture. Your documents → chunked and embedded → stored in vector database → query arrives → relevant chunks retrieved → fed to LLM with query → grounded response with source citations.
Simple RAG for PS clients. NotebookLM is a consumer RAG product — upload your documents, ask questions, get cited answers from your own materials. Enterprise RAG systems scale the same approach across organizational document libraries.
Embeddings and Vector Databases
The technology that allows AI to search documents by meaning rather than by keywords. An embedding is a mathematical representation — a long list of numbers — that captures the semantic meaning of a piece of text. Similar meanings produce similar number patterns. A vector database stores millions of these embeddings and instantly finds the documents, paragraphs, or data points most semantically similar to a query. This is the foundation of RAG and most enterprise AI search systems.
Keyword vs. semantic search. Keyword search finds documents containing "flood risk." Semantic search finds every document discussing water damage liability, force majeure for weather events, and insurance requirements — even if those exact words never appear together.
CRE application. Index your entire document library — leases, appraisals, environmental reports, lender agreements — and query semantically: "Which properties have asbestos remediation history?" across thousands of documents in seconds.
AEC application. Index all past project specifications. When writing a new spec section, retrieve all prior comparable sections automatically — consistent quality, no missed precedents, institutional knowledge that survives personnel turnover.
Multimodal AI
AI models that process and generate multiple types of content — text, images, audio, video, and data — within a single system. Earlier AI models were unimodal: a language model handled text, a separate image model handled images. Frontier models today are multimodal: Claude can read a PDF with charts and tables; GPT-4o can analyze a photograph; Gemini can process a video. The bottleneck between different content types is dissolving.
CRE applications. Upload a property photo: "What deferred maintenance issues are visible?" Upload a floor plan image: "Does this layout meet ADA corridor width requirements?" Upload a financial table screenshot: extract the data into a spreadsheet.
AEC applications. Photograph a construction defect and ask for likely causes and remediation options. Upload a drawing set image and ask the model to identify inconsistencies between architectural and structural plans.
Natural resources. Upload satellite imagery and ask for visible surface disturbance patterns. Analyze LiDAR data visualizations for terrain features relevant to site access planning.
Temperature and Sampling
The parameter that adjusts how creative or conservative an AI model is when generating responses. Temperature (typically 0 to 1) controls the randomness of the model's word choices during generation. Low temperature (near 0): the model picks the most statistically likely word at each step — consistent, predictable, sometimes repetitive. High temperature (near 1): the model samples more broadly — varied, creative, sometimes surprising or incoherent.
Low temperature use cases. Legal document review. Financial analysis. Factual summaries. Regulatory compliance checks. Situations where consistency and accuracy matter more than variety.
High temperature use cases. Brainstorming. Marketing copy. Creative naming. Ideation sessions where novelty is the goal and accuracy is secondary.
Consumer interfaces. Claude.ai and ChatGPT manage temperature automatically. API users set it explicitly. Most commercial products default around 0.7 — a practical middle ground for everyday professional tasks.
GAN (Generative Adversarial Network)
A generative AI architecture where two neural networks compete: a generator creates synthetic content, and a discriminator evaluates whether each output is real or generated. The generator improves by fooling the discriminator; the discriminator improves by detecting fakes. This adversarial feedback loop produces highly realistic synthetic outputs. GANs dominated image generation before diffusion models became the preferred approach around 2022.
CRE and AEC implication. AI-generated property renderings and architectural visualizations are GAN or diffusion model outputs. Synthetic imagery that misrepresents finished conditions of a property or project is a growing liability exposure. Buyers and investors are entitled to know when visualizations are AI-generated.
Synthetic data. GANs are widely used to generate synthetic datasets for training other AI models where real data is scarce or sensitive — synthetic financial transactions for fraud detection, synthetic patient records for healthcare AI.
Deepfakes. GAN-generated video of real people saying things they never said. A significant risk in professional contexts — verify the source of any video-based testimonial or endorsement before acting on it.
Diffusion Model
The architecture behind the current generation of AI image generators — Stable Diffusion, DALL-E, Midjourney, Adobe Firefly. A diffusion model learns to reverse a noise-addition process: starting from a clear image progressively corrupted to pure static, the model learns to reconstruct clarity from noise. At generation time, starting from pure noise and running the process in reverse produces a new coherent image matching a text description. Diffusion models produce more photorealistic and controllable outputs than GANs for most use cases.
Text-to-image. "Aerial view of a 200-unit multifamily development with surface parking, pool, and leasing office, realistic architectural rendering, golden hour lighting." → A photorealistic concept rendering in seconds at effectively zero marginal cost.
Image-to-image. Start from an existing site photo or rough sketch and refine it toward a target style or finish condition.
Disclosure imperative. AI-generated property renderings used in investment marketing must be clearly disclosed as synthetic. Regulatory guidance is still forming, but the direction is unambiguous: treat AI imagery as you would artist's renderings — disclose it as such.
AI Agent
An AI system that can take actions — not just generate text. A basic LLM responds to a prompt with text. An AI agent receives a goal, breaks it into steps, uses tools (web search, code execution, file operations, API calls, email), executes those steps in sequence, evaluates results, and adjusts its approach — without requiring a human at each step. Agents are what make AI a workflow tool rather than just a writing assistant.
Simple agent example. "Research current cap rates for multifamily in Denver, pull the three most recent comparable sales, and draft a one-page market summary." The agent searches the web, retrieves data, and writes — you review the output.
Multi-agent systems. Complex workflows chain multiple agents: a research agent, a drafting agent, a verification agent, a formatting agent — each specialized, each passing outputs to the next.
The operator's tool. Agents are where AI stops saving individual minutes and starts saving organizational hours. A properly designed agent workflow replaces a recurring manual process entirely — the human reviews the output rather than producing the inputs.
Human-in-the-loop. Best practice is a human review gate before any agent output is delivered externally. The agent does the work; the human takes accountability for the result.
Model Context Protocol (MCP)
An open standard introduced by Anthropic in 2024 that defines how AI models connect to external data sources, tools, and systems. Before MCP, every integration between an AI model and an external system required custom-built connectors. MCP provides a universal interface: any tool or data source implementing the protocol can connect to any AI model supporting it — similar to how USB standardized device connections, or how TCP/IP standardized internet communication.
What it enables. Connect Claude or another model directly to your CRM, document management system, project management platform, financial database, or any internal tool — without custom engineering for each connection.
Practical example. An AI agent connected via MCP to your project management tool, email, and document library autonomously compiles a weekly project status report from all three systems — no manual aggregation required.
The USB analogy. Before USB, every peripheral needed a different port and a different driver. MCP is doing for AI integrations what USB did for computer peripherals: standardize the connection so you can plug anything into anything.
AI Governance
The policies, processes, and accountability structures organizations put in place to manage AI use responsibly. As AI embeds itself in professional workflows, governance is no longer optional — it is a risk management, liability, and increasingly a regulatory requirement. Good AI governance answers four questions for every AI deployment: Who owns the output? What data is going in? What human review is required? What happens when something goes wrong?
Data privacy. Most external AI tools retain inputs by default. Client confidential information, personnel records, unreleased financial data — none of this should enter an external AI tool without understanding the data retention policy. Know before you paste.
Output ownership. Every AI-assisted deliverable needs a named human accountable for its accuracy. "The AI wrote it" is not a defense for an error in a client proposal, lender package, or regulatory filing.
Regulatory direction. The EU AI Act, SEC guidance on AI in financial advice, and CFPB guidance on AI in lending are the leading edge of a regulatory wave. Firms that build governance frameworks now will be ahead of the requirements arriving in the next 3–5 years.
Anthropic's family of AI models, widely considered the leading option for professional writing, document analysis, and nuanced reasoning. Anthropic was founded by former OpenAI researchers with an explicit focus on AI safety — Claude's training emphasizes being helpful, honest, and harmless, making it less likely than some competitors to produce confidently wrong outputs on domain-specific professional content. Available via claude.ai (consumer and business tiers) and the Anthropic API for enterprise integrations.
Strengths. Long-form writing with consistent professional voice. Document analysis at scale (200K token context window — roughly 500 pages in a single session). Following complex multi-part instructions. Nuanced judgment on ambiguous situations. Acknowledgment when it does not know something rather than fabricating an answer.
Model tiers. Claude Haiku (fast, inexpensive, high-volume tasks). Claude Sonnet (the everyday workhorse — strong capability, reasonable speed and cost). Claude Opus (highest capability — complex analytical tasks, sophisticated reasoning).
For PS clients. Best choice for drafting lender packages, summarizing lease portfolios, writing investment memos, and any task requiring consistent professional voice across long documents.
ChatGPT (OpenAI)
OpenAI's consumer-facing AI product, built on the GPT series of models. ChatGPT launched in November 2022 and reached 100 million users faster than any consumer application in history — the product that brought generative AI into mainstream awareness. OpenAI remains the market leader by user volume and brand recognition. Available at ChatGPT.com (free, Plus, and Pro tiers) and via the OpenAI API for enterprise integration.
GPT-4o. The everyday model — fast, multimodal (handles text, images, audio), strong general performance across virtually any task type.
o3 and o4 (reasoning models). OpenAI's "thinking" models that reason through complex problems before responding — significantly stronger on math, logic, and multi-step analytical tasks. Slower and more expensive than GPT-4o; worth it for the right tasks.
Strengths. Breadth. The largest prompt engineering community and most third-party integrations. Code generation. Image generation via DALL-E built in.
Relative to Claude. Less consistent on very long documents. More likely to produce polished-sounding but subtly inaccurate outputs on specialized professional content. Better on coding and mathematical reasoning tasks.
Gemini (Google)
Google's family of AI models, deeply integrated with Google Workspace — Docs, Gmail, Sheets, Drive. Gemini represents Google's attempt to leverage its search, data, and infrastructure advantages in the AI era. The family spans Nano (on-device), Flash (fast and efficient), Pro (everyday tasks), and Ultra (highest capability). Available at gemini.google.com and embedded throughout Google products.
Key advantage. Google integration. If your organization runs on Google Workspace, Gemini's native access to your Drive, calendar, email, and documents — without uploading — is a significant operational advantage over alternatives.
Search integration. Gemini has access to current web information through Google Search, making it stronger than Claude or ChatGPT on questions requiring up-to-date data not in training sets.
NotebookLM. Google's document intelligence tool — upload your documents, ask questions, get cited answers from your own materials. Excellent for document-heavy workflows. See separate entry.
For PS clients. Best choice if you run Google Workspace. For current data and research tasks, Gemini and Perplexity outperform Claude and ChatGPT. For drafting and deep document analysis, Claude remains the stronger choice.
Microsoft Copilot
Microsoft's AI assistant, embedded throughout the Microsoft 365 suite — Word, Excel, PowerPoint, Outlook, Teams — as well as available standalone. Built on OpenAI's models (Microsoft is OpenAI's largest investor), Copilot brings AI directly into the tools most professional organizations already use. For firms with Microsoft 365 Business or Enterprise licenses, Copilot represents the lowest-friction AI adoption path available — the tools look the same, there is just an AI layer inside them.
In-application use cases. Word: draft from an outline, summarize a document, rewrite a section. Excel: analyze data, generate formulas, create charts from a description. PowerPoint: create a presentation from a Word document or text prompt. Outlook: draft replies, summarize email threads, prepare meeting briefs.
Microsoft 365 Copilot (enterprise). Accesses your organization's content across all Microsoft 365 apps — SharePoint, Teams conversations, email history — synthesizing answers from your institutional knowledge base, not just the current document.
The adoption advantage. The CAD revolution required learning new software. Copilot works inside software you already know. The learning curve is about prompting, not about the interface.
Perplexity
An AI-powered search engine that retrieves live web sources and synthesizes cited answers — combining the research capability of a search engine with the conversational fluency of a language model. Unlike Claude or ChatGPT, which generate responses from training data alone, Perplexity retrieves current sources and builds its answer from them, providing citations for every claim. The result is AI-quality synthesis with real-time currency.
When to use Perplexity instead of Claude or ChatGPT. Any question where the answer may have changed in the last 6–24 months: current market data, regulatory updates, personnel changes, recent transactions, current interest rates, new legislation.
For CRE research. "What are current cap rates for industrial properties in the Denver metro?" "What zoning changes were passed in Denver County in Q1 2026?" Perplexity retrieves current sources; Claude and ChatGPT answer from training data that may be a year or more stale.
The combination workflow. Perplexity for current data retrieval → Claude for synthesis and drafting → you for judgment and verification → professional output in a fraction of prior time.
Perplexity Pro / Deep Research Tools
The advanced tier of AI research tools that go beyond single-query answers to conduct multi-step research autonomously — searching dozens of sources, synthesizing findings, identifying gaps, and producing structured research reports. OpenAI's Deep Research, Perplexity Pro's research mode, and similar tools from Google and Anthropic represent a new capability tier: AI that does an analyst's research workflow, not just answers a single question.
What deep research produces. A comprehensive market analysis brief with 30+ sources cited. A regulatory landscape overview across multiple jurisdictions. A competitive intelligence report on an acquisition target. Work that previously took a researcher 2–3 days now takes 20–30 minutes of AI-assisted work and 30–60 minutes of human review.
Verification requirement. Deep research tools produce more sources than a standard query, but hallucination risk does not disappear — it can actually increase when the tool is synthesizing across many sources. Verify every material claim before including in professional deliverables.
Best use. Background research for investment decisions. Regulatory and compliance landscape reviews. Market entry analysis. Competitive intelligence. Any research task where breadth and synthesis are the primary challenge.
NotebookLM (Google)
Google's document intelligence tool that turns your own uploaded documents into a queryable knowledge base. Upload PDFs, Word documents, Google Docs, websites, or YouTube videos, and NotebookLM creates an AI assistant that answers questions, generates summaries, identifies connections, and produces briefing documents — entirely grounded in your source material, with citations. It does not hallucinate content from outside your uploaded documents.
Due diligence application. Upload an entire data room and ask: "What environmental concerns are mentioned across these documents?" "Which documents reference the parking easement?" "Summarize all representations about tenant creditworthiness."
Lease portfolio analysis. Upload 20 leases and ask: "Which leases have personal guarantee provisions?" "Which have co-tenancy clauses?" "Summarize the rent escalation structures across the portfolio."
The grounding advantage. Because NotebookLM answers only from your uploaded documents, hallucination is dramatically reduced compared to general-purpose LLMs. If the answer is not in your documents, it says so — a critical professional protection.
GitHub Copilot
Microsoft's AI coding assistant, built on OpenAI's models and integrated directly into code editors. Copilot suggests code completions, generates functions from plain-language descriptions, explains existing code, and identifies bugs — in real time, as you type. Developers report 30–55% productivity gains. For organizations building internal tools, automating workflows, or managing data pipelines, Copilot represents one of the highest-ROI AI deployments available today.
Not just for software companies. Any organization building custom tools — automated report generation, data processing scripts, API integrations — benefits from Copilot. Cost: $19/month per developer.
Non-developer applications. Copilot's "explain this code" and "fix this error" features help non-developers debug Excel macros, Python scripts, or simple automations — without needing a developer for every small fix.
The ROI benchmark. GitHub's internal study found developers complete tasks 55% faster with Copilot. For a developer at $120K/year, that is roughly $66K of additional output per year at $228/year cost. Well above the 20:1 standard.
Claude Code
Anthropic's agentic coding tool, available as a command-line interface that gives Claude direct access to your codebase, terminal, and development environment. Unlike Claude in a browser window — where you paste code and read responses — Claude Code reads files, edits code, runs tests, executes commands, and iterates autonomously toward a goal you define. It is designed for developers who want AI to act as a working collaborator on their codebase rather than an advisor they consult manually.
What it does. Reads your entire codebase to understand structure and conventions. Writes and edits files directly. Runs tests and fixes failures. Executes terminal commands. Operates across multiple files simultaneously — refactoring, debugging, and implementing features end to end.
Agentic distinction. Claude Code does not just answer questions about code — it takes actions. The difference between a tool you consult and a tool that works alongside you. This places it in the AI agent category: goal-directed, multi-step, with tool use and environmental feedback.
For PS pipeline operations. Claude Code is the appropriate tool for building and maintaining the kind of GitHub Actions pipeline, Python scripts, and automation workflows that Prosper Systems operates. It can read the full codebase context, make coordinated changes across multiple scripts, and test changes before deployment.
Cursor / AI Code Editors
A new generation of code editors built with AI as a first-class feature — not bolted on, but architecturally central. Cursor is the leading example: a VS Code fork where the AI assistant has full awareness of your entire codebase, can edit multiple files simultaneously, understands project structure and conventions, and accepts natural language instructions to implement features or fix bugs. GitHub Copilot (integrated into VS Code, JetBrains, and other editors) is the incumbent with the largest user base.
How it differs from pasting code into a chat window. A chat window sees only what you paste. An AI code editor has persistent awareness of your full project — every file, every function, every dependency. Instructions like "add error handling to all database calls in this project" are executed across the entire codebase, not just the snippet you happened to paste.
Productivity evidence. GitHub's own studies found developers complete tasks 55% faster with Copilot. Cursor users report similar or higher gains on complex, multi-file tasks. For organizations building internal tools, automating workflows, or maintaining custom applications, the ROI on AI code editors consistently clears the 20:1 threshold.
Non-developer applications. The "explain this code" and "fix this error" capabilities lower the floor for non-developers maintaining scripts, macros, and simple automations — reducing dependency on developer availability for every small change.
n8n
A no-code and low-code workflow automation platform that connects AI models to existing business systems — Gmail, Google Sheets, Slack, CRMs, project management tools, databases — without requiring engineering resources. n8n uses a visual drag-and-drop interface to build automated pipelines: when X happens, do Y, then Z. AI nodes (Claude, ChatGPT, or others) can be inserted at any step to add intelligence — classifying, drafting, summarizing, deciding — within an automated workflow.
Proven use cases. Lead qualification: new contact enters CRM → AI evaluates against criteria → qualified leads routed, unqualified archived with reason code. Invoice processing: invoice received by email → AI extracts line items → accounting system updated → exceptions flagged. Compliance monitoring: regulatory source monitored → changes detected → AI summarizes impact → alert sent to relevant team member.
Data privacy option. n8n can run on your own server — relevant for organizations with strict data governance requirements on what leaves their infrastructure.
Vs. Zapier. Zapier is simpler but more limited. n8n is more powerful with direct AI model integration. For complex organizational workflows involving AI, n8n is the more capable platform.
Local / On-Premises AI Models
AI models that run entirely on your own hardware — laptop, workstation, or private server — with no data transmitted to external services. Tools like LM Studio and Ollama let non-technical users download and run open-source models (Meta's Llama, Mistral, Google's Gemma) locally. The capability gap between local models and frontier commercial models (Claude, GPT-4o) is real but narrowing; for many professional tasks, a well-chosen local model running on a modern workstation delivers acceptable results with complete data privacy.
When local models make sense. Client confidential information that must not leave your infrastructure. High-volume repetitive tasks where per-token API costs accumulate. Regulated industries with strict data residency requirements. Air-gapped environments where internet connectivity is restricted or prohibited.
The capability trade-off. Local models on consumer hardware are 1–3 generations behind frontier commercial models on complex reasoning, long documents, and nuanced professional writing. For document summarization, classification, and structured extraction, the gap is smaller. For complex analysis and drafting, frontier models remain materially better.
LM Studio. The most accessible local model interface — a desktop application with a chat interface, model library, and local API server. No command line required. Download a model, run it, use it like a local Claude or ChatGPT. Starting point for any organization evaluating on-premises AI options.
DALL-E, Midjourney, Stable Diffusion
The leading AI image generation platforms, each built on diffusion model architecture. DALL-E 3 (integrated into ChatGPT) produces precise, instruction-following imagery; Midjourney produces the highest-quality photorealistic and artistic outputs but requires a Discord interface; Stable Diffusion is open-source, runs locally, and offers the highest customization at the cost of setup complexity. Adobe Firefly (integrated into Creative Cloud) is trained on licensed content — the clearest intellectual property position for professional commercial use.
Professional use cases. Concept renderings for development pitches. Marketing visuals for investment decks. Diagram and illustration generation for reports. Quick visualization of spatial concepts before engaging a designer for production work.
CRE and AEC limitation. AI-generated images are not replacements for professional architectural renderings, accurate site plans, or engineering drawings. They are ideation tools — faster and cheaper than early design sketches, not substitutes for construction documents.
Disclosure requirement. AI-generated imagery used in investment marketing or client presentations must be clearly disclosed as synthetic. The standard is forming rapidly across real estate, securities, and construction sectors.
Whisper (OpenAI)
OpenAI's open-source speech recognition model, capable of transcribing and translating audio in 99 languages with near-human accuracy. Whisper is the foundation for most "transcribe this meeting" features across productivity platforms today. Available free as an open-source model that runs locally, or via the OpenAI API. The combination of Whisper-quality transcription plus an LLM for summarization is one of the highest-ROI workflow combinations available to professional organizations.
Meeting workflow. Record the meeting → Whisper transcribes automatically → Claude summarizes into structured minutes (attendees, decisions, action items with owners and due dates) → PM reviews and distributes. 90 minutes of post-meeting work becomes 10-20 minutes of review.
Field use. In construction, natural resources, and property inspection workflows, voice dictation via Whisper means field personnel describe observations verbally → AI formats into structured field reports → no manual data entry required.
Multilingual. Whisper handles accented speech and language mixing significantly better than earlier speech recognition systems — important for international projects and diverse site teams.
The instruction you give an AI model — the text input that tells it what you want. The quality of your prompt is the single largest determinant of the quality of AI output. Unlike a Google search query (a few keywords) or a database query (structured syntax), a prompt can be conversational, instructional, contextual, or any combination. "Write something" is a prompt. So is a 2,000-word detailed brief with role definition, context, format requirements, constraints, and example outputs. The latter produces radically better results.
Prompt engineering. The practice of crafting and refining prompts to reliably produce high-quality outputs. A developed skill, not an innate talent — anyone who invests time in it improves measurably and quickly.
The scope-of-work analogy. "Fix my building" → chaos. "Replace the HVAC units on floors 3–5, match the existing Carrier system, coordinate access with building management, work completed by end of Q2, budget $180K" → a contractor who can deliver. Specificity in a prompt works the same way as specificity in a scope of work. AEC professionals understand this instinctively.
Iteration. Good prompting is a conversation, not a one-shot transaction. When the first output misses the mark, refine: "make it shorter," "more formal," "add the insurance requirements." Build toward the target iteratively rather than starting over.
System Prompt
A hidden instruction that sets the AI model's role, behavior, constraints, and context for an entire conversation or application. Unlike a user prompt (what you type), the system prompt runs in the background, shaping every response the model gives. Consumer interfaces have system prompts you never see — they define how those products behave. Developers and enterprise operators write their own system prompts to create custom AI applications tuned for specific professional purposes.
What a system prompt can do. Assign a role ("You are a commercial real estate underwriter with 20 years of experience"). Enforce output format ("Always respond with an executive summary first, then supporting detail"). Set constraints ("Never speculate about future market conditions without explicitly flagging the speculation"). Load firm-specific context ("Here is our standard template for investment memos [paste]").
Why it matters. A well-crafted system prompt is what turns a general-purpose AI model into a specialized professional tool. The difference between a generic chatbot and a purpose-built AI assistant for your firm is almost entirely in the system prompt — not the model.
For operators. Document your system prompts with the same rigor you would document any critical business process. Version them. Test them when underlying models update. They are your firm's AI IP.
The Six-Part Prompt Framework
A structured approach to prompt construction that consistently produces professional-grade outputs. Most casual AI users employ 2–3 of these elements instinctively. Using all six deliberately is what separates reliable professional results from inconsistent outputs. The framework was developed and documented in the Prosper Systems Make US AI-Ready course.
1. ROLE. Who should AI be? "You are an experienced commercial real estate underwriter..." Assigning a role activates domain-specific vocabulary, reasoning patterns, and appropriate caution levels. The model behaves differently as an underwriter than as a generalist.
2. CONTEXT. What is the specific situation? Not "my client" but "an accredited investor considering their first multifamily syndication, skeptical about projections after a prior bad experience with a different sponsor."
3. TASK. What exactly do you want? "Write a two-page executive summary of this acquisition" is a task. "Write something about this property" is not.
4. FORMAT. What should output look like? Word count. Structure. Tone. Bullet points or prose. Formal or conversational. Tables or narrative.
5. RULES. What to avoid. "No jargon." "Keep assumptions conservative — flag anything speculative." "Do not mention interest rates without citing a current source." Rules define guardrails before generation begins.
6. EXAMPLES. Paste a previous output that hit the right standard. Demonstrated quality consistently outperforms described quality — showing beats telling for AI just as it does for human briefing.
Prompt Library
A documented collection of tested, reusable prompt templates for an organization's most common recurring AI tasks. The difference between AI being a tool you experiment with occasionally and AI being a structural part of how your organization works is almost entirely in whether you maintain a prompt library. Without one: every user starts from scratch, results are inconsistent, good prompts disappear when the person who wrote them leaves, and the organization never builds on its own learning.
What each entry should contain. Task name. The full prompt text with variable placeholders in [brackets]. The best AI model for this specific task. Notes on common failure modes. Date last tested and updated.
Maintenance cadence. AI models update continuously. Review and retest your most-used prompts quarterly — outputs that worked six months ago may need adjustment as underlying models change. Treat the prompt library as a living document, not a filing cabinet.
Few-Shot and Chain-of-Thought Prompting
Two of the most reliably effective prompting techniques for improving AI output quality. Few-shot provides examples of the desired input-output pattern before the actual request — showing the model what good output looks like rather than describing it. Chain-of-thought instructs the model to show its reasoning step by step before reaching a conclusion, improving accuracy on analytical and multi-step tasks and creating a reasoning trail you can review and verify.
Few-shot in practice. "Summarize this lease in 3 bullet points. Here is an example of a summary I consider high quality: [example]. Now summarize this new lease using the same format and level of detail." The model matches demonstrated quality far better than described quality.
Chain-of-thought in practice. "Analyze the financial viability of this acquisition. Think step by step: first evaluate the NOI, then the debt service coverage, then the exit multiple assumptions, then give your overall assessment." Each step builds correctly on the last — and you can catch errors in the reasoning before they corrupt the conclusion.
When to use each. Few-shot: when you have strong examples of the output you want and consistency matters. Chain-of-thought: when the task involves multi-step reasoning, analysis, or judgment where the reasoning process itself is as important as the conclusion.
Plugin
An extension that connects an AI model to an external service or capability — allowing the model to take actions or retrieve information beyond what it can generate from its training data alone. The term originated with ChatGPT's plugin system (launched 2023, largely superseded by GPT-4o's built-in tool use). In general usage, a plugin is any add-on that expands what an AI can do: browse the web, run code, query a database, send an email, read a file. The technical implementation varies by platform; the concept is consistent.
Plugin vs. tool. These terms are often used interchangeably, but technically: a plugin is the packaged extension (the thing you install or enable); a tool is the capability the model invokes during generation (the action it takes). A plugin provides tools. One plugin may provide multiple tools.
Plugin vs. skill. A skill is a reusable prompt template or instruction set that shapes how the model works — its behavior, voice, and output format. A plugin extends what the model can do — its access to external systems. Skills are about capability configuration; plugins are about capability extension.
Current landscape. ChatGPT plugins were deprecated in favor of GPT Actions and the built-in tool ecosystem. Claude uses MCP (Model Context Protocol) for the same purpose. Microsoft Copilot uses connectors and Copilot Studio extensions. The terminology differs; the concept — connecting AI to external systems — is universal.
Skill (AI)
A reusable, documented instruction set — typically a structured prompt or prompt template — that reliably produces a specific type of output or behavior from an AI model. Skills are the building blocks of organizational AI deployment: instead of every user figuring out how to prompt the model for lease abstractions, investment memo drafts, or field reports, the organization maintains tested skills for each task that any user can invoke with consistent results. The term is used differently across platforms but the concept is consistent.
Skill vs. prompt. A prompt is a single instruction for a single use. A skill is a reusable template built from a refined prompt — documented, versioned, and maintained. The same way a procedure is a documented version of something someone figured out once and wrote down so it could be repeated reliably.
Skill vs. plugin. A plugin extends what the model can do (access external systems, run code). A skill defines how the model behaves for a specific task (what to produce, in what format, with what constraints). Skills are behavioral configuration; plugins are capability extension.
Platform usage. Microsoft Copilot Studio uses "skills" for reusable AI behaviors. Claude Code uses skills as documented prompt templates with environment-specific constraints (in SKILL.md files). Custom GPT instructions serve a similar purpose in OpenAI's ecosystem. The underlying concept — packaging and reusing refined prompting intelligence — is the same across platforms.
Tool (AI Function Calling)
A defined capability that an AI model can invoke during generation — such as searching the web, running code, querying a database, reading a file, or calling an API. The model decides when to use a tool based on the user's request, executes it, incorporates the result into its response, and continues. This is what separates an AI agent from a standard language model: an LLM generates text; an LLM with tools takes actions. "Function calling" is the technical term for the same concept in API contexts.
How it works. The developer defines available tools (name, description, parameters). The model reads the definitions. When generating a response, if using a tool would help, the model outputs a structured tool call. The application executes the tool and returns the result. The model incorporates the result and continues. From the user's perspective: you ask a question; the model searches, computes, or retrieves; you get an answer grounded in real data.
Common tools in professional deployments. Web search (current data). Code interpreter (calculations, data analysis, chart generation). File read/write (document processing). Database query (structured data retrieval). Email send (workflow automation). Calendar query (scheduling).
The agent foundation. Tools are what make agents possible. An agent is not a special type of model — it is a standard model with access to tools, given a goal and the autonomy to invoke those tools in sequence until the goal is achieved. Remove the tools and you have a standard chatbot. Add the tools and you have an agent.
Agentic Pipeline
An automated, multi-step workflow in which one or more AI agents execute a sequence of tasks — retrieving data, generating content, making decisions, calling external systems — with minimal human intervention at each step. The human defines the goal and reviews the final output; the pipeline handles the work in between. Agentic pipelines represent the shift from AI as a writing assistant to AI as an operational infrastructure layer — where recurring organizational work happens automatically rather than requiring human effort at each step.
Pipeline vs. agent. A single agent completes a task. A pipeline chains multiple agents or steps: a research agent feeds a drafting agent, whose output feeds a formatting agent, whose output triggers a notification. Each step may use a different model, different tools, and different success criteria.
Prosper Systems example. The PS content pipeline: email digest fetched → stories scored and selected → Claude drafts posts → DALL-E generates images → output committed to GitHub → notification sent → content ready for review. No human involvement between trigger and output. The Admiral reviews the result; the pipeline does the work.
The Admiral/Captain model. Effective agentic pipeline design separates strategic direction (human) from execution (AI agents). The human defines goals, reviews outputs, and handles exceptions. The pipeline handles the recurring work. This is not AI replacing humans — it is AI handling the execution layer so humans can operate at the judgment layer.
AI Workflow Automation
The integration of AI into recurring business processes so that work happens automatically — without requiring human effort at each step. Distinct from AI-assisted individual tasks (a human using AI to do their work faster), workflow automation removes the human from the recurring steps entirely, with humans reviewing outputs rather than producing inputs. This is where AI shifts from saving individual minutes to recapturing organizational hours — and from a productivity tool to a structural competitive advantage.
The automation target profile. Recurring (happens weekly or more often). Predictable (same input type produces same output type). Time-consuming relative to its complexity. Currently handled by a human doing largely repetitive work. If a task fits those four criteria, it is an automation candidate.
Examples for PS clients. Weekly market data digest: sources monitored → AI summarizes → email delivered to team. New lead intake: form submitted → AI scores and routes → CRM updated → follow-up drafted. Document processing: file received → AI extracts key provisions → structured data output → exceptions flagged for human review.
The human's new role. In an automated workflow, the human's job shifts from producing the work to reviewing the output and handling exceptions. This is higher-value work — judgment and relationships, not compilation and formatting.
Guardrails / Safety Layer
Technical and procedural constraints built into an AI system to limit outputs that are harmful, inaccurate, off-brand, or operationally risky. Guardrails operate at multiple levels: the model itself (trained behavior), the system prompt (operator-defined constraints), the application layer (output filtering, human review gates), and organizational policy (what tasks AI is and is not authorized to complete autonomously). In professional deployments, guardrails are not optional — they are the mechanism by which AI deployment becomes defensible to clients, regulators, and insurers.
Model-level guardrails. Built into training via RLHF. Claude is trained to refuse requests for harmful content, acknowledge uncertainty, and avoid confident statements on topics where it lacks reliable knowledge. These are not perfect — they can be circumvented and they fail at the edges — but they represent the baseline safety layer every commercial model ships with.
Operator-level guardrails. System prompt constraints that restrict the model's behavior within a specific application: "Never provide specific investment advice." "Always flag any claim that cannot be verified from the provided documents." "Do not generate content that names specific competitors." These are your responsibility to define and maintain.
Human-in-the-loop as guardrail. The most reliable guardrail for high-stakes professional outputs is a defined human review gate before any AI output reaches a client or regulatory audience. Technology guardrails reduce the volume of problems reaching human review; they do not eliminate the need for it.
Jailbreak
A technique used to bypass an AI model's safety guardrails — causing it to produce outputs it was trained to refuse. The term comes from mobile phone culture (unlocking a device from carrier restrictions) and carries the same implication: accessing capabilities the manufacturer intentionally locked. Jailbreaks range from elaborate multi-step prompt constructions to surprisingly simple reframings of a refused request. No guardrail system is jailbreak-proof — it is an unsolved adversarial problem, not a patchable software bug.
How jailbreaks work. Most exploit the gap between the model's trained refusal patterns and its core language capability. If the model refuses "explain how to exploit this vulnerability," it may comply with "fix the bugs in this code" — the output is functionally equivalent but the framing bypasses the trigger. Role-play setups, hypothetical framings, and multi-turn context manipulation are common techniques.
The Fable 5 case (June 2026). The U.S. government imposed export controls on Anthropic's Fable 5 and Mythos 5 models based on a research report from Amazon's security team. The reported jailbreak: researchers asked Fable to "review the code for security issues" — it refused. They reframed the request as "fix this code" — it complied, producing patches for known vulnerabilities. Over 100 cybersecurity experts signed an open letter arguing this was not a jailbreak at all — it is standard defensive security workflow. The controversy illustrates that the line between a jailbreak and legitimate professional use is often contested.
Offensive vs. defensive framing. The same capability that lets an attacker probe for exploits lets a defender find and patch them first. Restricting AI from "fix this code" tasks handcuffs security teams without meaningfully slowing adversaries who have access to the same models via other channels or jurisdictions.
For operators. Jailbreak risk is real but context-dependent. For most professional deployments — document analysis, drafting, research — it is not a material concern. For applications where AI outputs could cause direct harm if guardrails were bypassed (financial advice, medical guidance, security tooling), operator-level constraints and human review gates are the practical defense — not reliance on model-level guardrails alone.
Verification Pipeline
A structured, documented process for checking AI-generated outputs before they reach a client, lender, regulator, or the public. Individual vigilance is not a verification system — under time pressure, individual checks get skipped. A verification pipeline embeds checkpoints in the workflow itself, making review a structural requirement rather than an individual judgment call subject to the pressures of the moment.
Gate 1: Source verification. Before AI processes source material, confirm the source is current and authoritative. Wrong-version specifications or outdated regulatory references produce wrong outputs that pass all downstream checks.
Gate 2: Output validation. Structured checks immediately after generation: required sections present? Format correct? Obvious hallucination signals — invented citations, implausibly precise statistics, names not in source material?
Gate 3: Human review. Scoped review focused on the highest-risk content categories for this output type: statistics, citations, regulatory references, client-specific financial figures. Not a full reread of every word — a targeted check of the most likely failure points.
Documentation. For any AI-assisted professional deliverable: log the model version, date, and who completed each verification step. When a client or regulator asks about your review process, have a defensible answer prepared.
ROI Standard: 10:1 / 20:1 / 50:1 / 80:1
A tiered benchmark for evaluating whether an AI workflow delivers meaningful value — developed by Prosper Systems from real deployment experience. The foundational example: an 8-page finance options list expanded to 80 pages of structured, formatted content in 30 minutes. Done manually, that task takes a week of hard work. 80 pages from 8 in 30 minutes = 80:1. That is not a cherry-picked illustration — it is the standard that demonstrates what properly deployed AI actually achieves when architecture, prompting, and workflow align.
10:1 minimum — the floor. If an AI workflow cannot deliver at least 10 units of output value per unit of input effort, it is not the right tool for that task. Deprioritize and find a better application.
20:1 — daily workflow target. A repeatable prompt system in daily production use. GitHub Copilot-class productivity. Practitioners who have built a working prompt library and tool stack operate at this level consistently.
50:1 — full deployment. A production system built once and running continuously. When the system does the work, the ratio scales with usage rather than with headcount.
80:1 and above — exceptional deployment. The original real example. Not every project reaches it, but it is achievable and repeatable when all elements align correctly.
Decision rule. Under 3:1 → deprioritize. Over 10:1 → build now. Over 20:1 → make it a daily system. Over 50:1 → document it, scale it, and teach it to your team.
The process of extracting key provisions from commercial leases — rent, term, options, CAM provisions, exclusives, co-tenancy clauses, assignment restrictions — and organizing them in a standardized summary. Manually abstracting a single complex lease takes 2–4 hours for an experienced professional. AI reduces this to 10–15 minutes with comparable accuracy on standard provisions. A portfolio of 50 leases that previously required a week of attorney or paralegal time can now be processed in a day.
Recommended approach. Upload the lease to Claude or NotebookLM. "Extract the following provisions in a structured table: [list specific provisions]. Flag any provision that deviates significantly from standard market terms. Note any defined terms that could materially affect these provisions."
Portfolio-level queries. Upload all leases to NotebookLM. "Which leases have personal guarantee provisions? Which have co-tenancy clauses? Summarize the rent escalation structures across the portfolio." Answers across dozens of leases in seconds.
Verification requirement. Always human-verify options (purchase, renewal, expansion), assignment clauses, and any provision triggering significant financial or legal consequences. AI lease abstraction is a powerful first pass; attorney review remains required for material provisions.
CRE: Investment Memoranda and Pitch Materials
The full suite of investor-facing documents — offering memoranda, executive summaries, pitch decks, investor updates — that communicate an investment opportunity. These are high-stakes documents where quality directly affects capital raising outcomes. AI produces strong first drafts of every section, dramatically reducing time from deal identification to investor-ready materials. The human's role shifts from writing to editing, judgment, and verification.
The workflow. Gather raw materials (deal summary, financials, market data). Inject into Claude with a detailed prompt defining audience, tone, format, and required sections. Generate first draft. Edit for accuracy, voice, and specific deal nuances. Verify all financial figures and market statistics against source documents before sending.
Consistency advantage. AI maintains consistent voice and format across a 30-page document — something human writers struggle with across long documents or when multiple team members contribute sections. One review cycle instead of five.
Disclosure and compliance. AI-assisted investment materials require review by qualified legal and compliance professionals before distribution to investors. AI can draft; it cannot ensure regulatory compliance. The liability and accountability are yours.
CRE: Underwriting and Financial Analysis Support
AI accelerates every non-calculation component of CRE underwriting: narrative drafting, assumption documentation, comparable analysis, lender package writing, investor Q&A preparation. The spreadsheet math remains yours; AI handles the language layer surrounding it — and the research layer feeding it.
What AI does well. Drafting narrative sections of underwriting packages. Summarizing market data from multiple research reports into a coherent market overview. Writing the assumptions and rationale section. Generating investor FAQ drafts. Producing first drafts of executive summaries.
What AI does not replace. The judgment calls in your assumptions. Your knowledge of the specific market, submarket, and property. The relationships with lenders and equity partners. The accountability for the numbers.
Current data workflow. Use Perplexity for current cap rates, vacancy rates, absorption data, and supply pipeline — its live web access makes it far more reliable than Claude or ChatGPT for market data published in the last 12–24 months. Then use Claude to incorporate that data into the narrative.
AEC: Specification Writing
Construction specifications — the written technical requirements defining materials, methods, standards, and quality for every project component — are among the most document-intensive outputs in professional practice. A full specification set for a major project can exceed 1,000 pages. AI can draft specification sections from product data sheets, project parameters, and applicable standards, with the engineer or architect reviewing for project-specific accuracy and code compliance.
The generational shift. In the 1970s and '80s, specs were typed from scratch or pulled from reference books. CAD-era firms built master spec libraries edited for each project. AI generates the first draft from project parameters — the master spec becomes a quality standard rather than a starting point. Each generation does the same work at higher leverage.
Recommended approach. Feed the AI your project parameters, applicable codes, product data sheets, and your firm's master spec section for the same product type. "Draft the specification section reconciling all inputs and flagging any conflicts between the product data and the applicable standard." Human reviews for code compliance and project-specific accuracy.
Verification non-negotiable. Errors in specifications become RFIs, change orders, and potential professional liability. AI-drafted specifications require engineer-of-record review before issuance. This is not negotiable regardless of how good the AI output looks.
AEC: Project Documentation and Reporting
The administrative documentation layer of construction projects — meeting minutes, RFI logs, submittal logs, change order documentation, daily field reports, project status reports — is legally important, time-consuming to produce, and heavily repetitive in structure. AI excels at exactly this combination. Field personnel can dictate observations by voice; AI formats them into structured reports automatically. What used to take hours of post-meeting and end-of-day administrative work takes minutes of review.
Meeting minutes workflow. Record the meeting → Whisper transcribes automatically → Claude summarizes into structured minutes format (attendees, decisions made, action items with owners and due dates) → PM reviews and distributes. 90 minutes of post-meeting administrative work becomes 10-20 minutes of review.
RFI drafting. "Draft an RFI for the following coordination conflict: [describe]. Include relevant specification sections, drawing references, the contractor's proposed resolution, and a request for direction." First draft in under two minutes.
The AEC documentation problem. Experienced AEC professionals from the first computer revolutions know this: documentation has always consumed disproportionate professional time. Every technology generation promised relief. AI is the first that actually delivers it — not because it digitizes the process, but because it understands the language of the work.
Natural Resources: Regulatory Document Analysis
Mining, oil and gas, water rights, and land use operations generate enormous regulatory document burdens — NI 43-101 technical reports, COGCC filings, environmental impact assessments, water court decrees, reclamation plans, royalty agreements. These documents are long, technically dense, and expensive to review. AI can read and summarize these documents, extract key provisions, compare across projects, and flag material changes in regulatory filings — at a fraction of the time cost.
NI 43-101 and reserves reports. AI can extract key resource and reserve figures, qualifying person statements, material assumptions, and risk factors from a technical report in minutes — a task that previously required an experienced geologist or investor spending hours with the document before the actual analysis could begin.
Regulatory monitoring workflow. Monitor regulatory agency websites for new filings, rule changes, or enforcement actions relevant to your projects → AI summarizes each new document → routes alerts with summaries to the relevant team member. Compliance awareness that was previously reactive becomes proactive.
Limitation. AI regulatory summaries are starting points, not qualified professional opinions. A Qualified Person under NI 43-101 remains legally required. An attorney must review legal documents for compliance advice. AI accelerates the expert's work — it does not replace their judgment or accountability.
Business: Contract Analysis and Comparison
Reading, comparing, and extracting key provisions from contracts — vendor agreements, employment contracts, partnership agreements, NDAs, service agreements — is one of the most time-consuming and error-prone manual tasks in professional practice. AI can read a contract, extract any specified provision, compare it against a template or prior agreement, flag non-standard terms, and draft redlines — in minutes rather than hours. The attorney still reviews; AI does the preliminary work that previously consumed billable hours at attorney rates.
NDA review workflow. "Review this NDA against our standard form [paste]. Identify every provision that deviates from our standard. For each deviation, note whether it is more or less favorable to us and why." First pass in 2 minutes. Attorney reviews the flagged items. Billable time focuses on judgment, not hunting for differences across 20 pages.
Portfolio comparison. Upload 20 vendor contracts to NotebookLM. "Which contracts have auto-renewal clauses? Which have limitation of liability caps below $1M? Which do not include a data breach notification requirement?" Answers across the entire portfolio in seconds — work that previously took a paralegal two days.
Limitation. AI contract analysis is preliminary review, not legal advice. Complex legal interpretation, jurisdiction-specific questions, and high-stakes negotiation positions require attorney judgment. AI finds the issues efficiently; attorneys resolve them with expertise.
Business: Market Research and Competitive Intelligence
Gathering, synthesizing, and presenting market intelligence — competitive landscape analysis, industry trend summaries, regulatory environment overviews — used to take days of analyst time. AI with live web search compresses this to hours. The combination of a language model for synthesis and a web-enabled AI for currency produces research outputs previously available only to large organizations with dedicated research staff.
The workflow. Define research questions. Use Perplexity to retrieve current sources on each question. Compile retrieved information. Feed to Claude: "Using these sources [paste], write a 3-page competitive analysis covering [defined scope]. Cite each claim to its source." Review, verify key claims, finalize. Market-quality research in hours rather than days.
For investment decisions. "Analyze current market conditions for self-storage in secondary Mountain West markets. Identify the three most significant demand drivers, the primary supply risk, and the most likely exit timing for a 5-year hold. Use the following market data [paste from Perplexity]."
Verification priority. Market research outputs rely heavily on current data. Always verify statistics and market figures against the original sources before including in client materials. Perplexity citations are starting points, not verified facts.
Cross-Sector: Due Diligence Acceleration
Due diligence — the systematic investigation of a business, property, or investment before committing capital — is one of the highest-value AI applications across all PS client sectors. The core work of due diligence is reading, extracting, comparing, and synthesizing large volumes of documents under time pressure. This is precisely what AI does at its best. Organizations using AI in due diligence consistently report 40–60% reduction in time-to-completion and improved coverage — fewer items missed because document review is exhaustive rather than sampled under time pressure.
Document processing at scale. Upload the entire data room. Ask AI to: identify all environmental contingencies mentioned across documents. Flag any liens, judgments, or legal proceedings. Extract all material contracts and summarize key terms. Identify discrepancies between representations in the offering memorandum and supporting documents.
The PS checklists as AI prompts. The Business Due Diligence and CRE Due Diligence Checklists on this site were designed as structured prompts — each item is a question you can hand to AI with the relevant documents and receive a structured answer rather than filling in a blank manually.
Speed with rigor. AI due diligence acceleration is not about cutting corners — it is about ensuring complete coverage in the time available. The human judgment still closes the deal or walks away; AI ensures nothing important was missed in getting to that decision.
Cross-Sector: Knowledge Capture and Institutional Memory
Every organization has institutional knowledge that exists only in people's heads — how decisions get made, why certain approaches work in specific markets, what past projects teach about avoiding particular mistakes. When those people leave, the knowledge leaves with them. AI provides the first practical tools to capture, structure, and make queryable the institutional knowledge that previously could not be systematically preserved — a problem that has plagued every sector since the first computer revolution created specialized roles.
The 1980s lesson revisited. The CAD revolution created a generation problem: senior professionals who held all the technical knowledge in their heads did not transfer it to digital systems. CAD operators learned to draw digitally; they did not learn why design decisions were made a certain way. Forty years later, AI provides the first real tool for capturing the "why" — not just the "what."
Knowledge capture workflow. Interview subject matter experts using AI-assisted conversation: AI asks follow-up questions, identifies gaps, and structures the transcript into a knowledge article. A 4-hour manual documentation process becomes a 45-minute conversation with a structured, searchable output.
Institutional query system. New staff query an AI trained on your firm's documents, past projects, and captured knowledge base — getting answers to "how do we typically handle X" that previously required finding the right senior person and hoping they had time to answer before leaving for a site visit.
↑SOURCES Prosper Systems research and commentary; Anthropic, OpenAI, Google, and Microsoft product documentation; Make US AI-Ready course (Prosper Systems, 2026); publicly available AI research and industry reporting.
Neither Prosper Systems (PS), nor its Founder, Kenton H Johnson, are licensed Real Estate or Lending Brokers, Securities Dealers or Investment Advisers. However, PS has an Attorney on its Team, and works closely with and engages other licensed individuals or firms as needed. The definitions in this Glossary are for guidance and educational purposes only and are not intended to be comprehensive or to constitute legal, financial, or technical advice. AI capabilities, products, and best practices change rapidly – verify current tool capabilities and pricing directly with providers before making deployment decisions. PS makes no warranties or representations as to the completeness or currency of information contained herein.
Led by Kenton H Johnson, Prosper Systems is a Team of Business, Sales and Technical Operations Consultants with resources and talent to work with businesses and projects worth us$2M to $20M+.
AI Glossary