AI Tuesdays · Week 08 · What AI Is in Plain English
MULTIFAMILY NEXT · AI TUESDAYS
Week 08 · What AI Is in Plain English
Self-guided · ~20 min · Phase 2 Opener
Warmup
The Workshop
Power Preview
Resources & Tools
Take Action
Week 08 of 52 · Self-Guided · ~20 min
0 starred0 notes
Warmup · Week 07 Recap
Welcome back. Let us warm up first.
Last week you learned how an AI CouncilAI COUNCILThe small group at your property, region, or portfolio that owns AI decisions. Also called "Governance Team." What it does: Approves AI use cases that require guardrails (called Yellow or Tier 2 requests). Refuses AI use cases that are never permitted (called Red or Tier 3 requests). Reviews every new AI tool before anyone uses it, using a standard 8-question intake. Standard seats: Governance Lead (chair), Operations, Compliance, Technology, and a rotating practitioner voice from someone who actually uses AI day to day. Meets monthly for most portfolios, more often in the first 90 days after launch. Why it exists: Without a Council, every AI decision defaults to whoever asked. That is how properties end up with a leasing team using consumer ChatGPT for resident-facing drafts and no one authorized it. The Council owns the answer so no one has to guess. Multifamily example: A leasing professional wants to use AI to draft renewal notices. Instead of guessing, she submits the use case to the AI Council. The Council approves it with guardrails (no resident names in the prompt, no ledger data, human sign-off on every draft, approved enterprise tool only) and logs the decision with a 90-day review date. Full deep dive: Week 07 (How an AI Council Actually Works). Anchor: RETTC Pillar 1 (Establish an Organizational Philosophy on AI Use and Oversight). actually works. Three quick recap questions to bring it forward. Tap any answer to reveal the coaching, no wrong tries counted.
New to AI Tuesdays or missed last week?
Read Week 07 first for the full context. It teaches how the AI Council operates, which this workshop assumes you already know.
Approval, refusal, new tool review. Tier 2 approvals move requests forward with guardrailsGUARDRAILSRules or restrictions built into an AI use case to keep it inside safe territory. Guardrails answer the question: "AI can do this, but ONLY IF..." Multifamily example: a leasing professional gets AI Council approval to draft resident renewal letters WITH GUARDRAILS: no rent amounts in the prompt, no resident names in the prompt, human sign-off on every draft before it leaves, and use of the approved enterprise AI tool only (not consumer ChatGPT). Four guardrails. Together they turn a risky use case into a safe one. Anchor: RETTC Pillar 5 (Maintain Outcome-Based Accountability & Human Involvement).. Tier 3 refusals redirect to an approvable version. New tool review runs the 8-question intake on any AI product before adoption. Three flows. One rhythm.
2What is the one boundary the AI Council never crosses?
Autonomous AI decisions that affect residents. AI can prepare, summarize, and recommend. AI never makes autonomous approval, denial, pricing, renewal, screening, or lease-termination decisions. Every one has a named human on it. That is the one boundary.
3When someone is unsure whether an AI use caseUSE CASEA specific way you plan to use AI in your work. Not the tool itself (ChatGPT is not a use case). Not a general idea ("marketing" is not a use case). A real use case names the person doing the work, the exact task, the data going in, and what comes out. Multifamily example: NOT a use case: "We want to use AI for renewal letters." IS a use case: "The property manager pastes a resident's move-in date and rent history into the approved enterprise AI tool. AI drafts a warm renewal letter offering a small kitchen upgrade credit. The property manager reviews and edits before sending. No resident name in the prompt, no ledger data pasted." That level of specificity is what makes a use case governable by the Council. Anchor: RETTC Pillar 1 (Establish an Organizational Philosophy on AI Use and Oversight). The Council approves use cases, not tools. is allowed, what is the default action?
Submit it to the AI Council. When in doubt, submit. The Council owns the answer. This is the default rule. Do not use, do not pilot, do not activate the workflow until the Council has cleared it.
Today · Week 08
What AI Is in Plain English
Past the buzzwords. Into the actual definition. Twenty minutes. Zero jargon left standing.
Skip the warmup anytime. Click Begin Reading when you are ready.
Multifamily NEXT · AI Tuesdays · Week 08
What AI Is in Plain English
Past the buzzwords. Into the actual definition.
For seven weeks you learned the reflexes and the governance model that protect you. Phase 1 done. Now Phase 2. You cannot govern well what you do not understand. So this week we take twenty minutes to define, in plain English, what generative AI actually is at the mechanism level. Not marketing. Not consulting-speak. Not vendor pitch. The real thing.
When you close this workshop, the AI conversations you are already in will make more sense. That is the whole promise.
Last Tuesday · Week 07
How an AI Council Actually Works
Three flows, one rhythm, zero mystery. That closed Phase 1. Today we open the hood on what the Council is actually governing.
Beginner-friendly · Read at your own pace · Star slides, add notes, print, or email yourself anytime
How This Workshop Is Organized
What is inside
Twenty minutes of reading. No timer. Pause whenever you need. Every acronym is spelled out on first use, so no prerequisites required.
Part 1 ~3 min
Warmup and opening
Recap Week 07. Set up the question the whole workshop answers: what is this thing?
Part 2 ~4 min
What AI actually is (and is not)
Generative AI in plain English. What it does. What it does not do, no matter what the marketing says.
Part 3 ~5 min
How it works under the hood
Large Language Models, tokens, probability, training data. In language your leasing team could follow.
Part 4 ~2 min
Where AI shines and where it fails
The two categories. When to use it. When to close the browser tab.
Part 5 ~3 min
Power Preview
A guided walkthrough of using AI to explain an AI concept to a coworker in plain language.
Part 6 ~3 min
Practice, quiz, close
Vocabulary practice. Five-question quiz. Download the AI Vocabulary Cheat Sheet. Preview Week 09.
Objective: By the end of this workshop, you will be able to define generative AI in plain English, name the three things a Large Language Model actually does, explain the difference between AI and search, and know what AI reliably fails at. Plus a two-page vocabulary cheat sheet to keep next to your desk.
Section 1 · Why This Week Matters
You cannot govern well what you do not understand.
Phase 1 gave you the reflexes and the governance model. But if your team, your vendors, and your executives keep tossing around AI vocabulary that nobody has actually defined, decisions get made on vibes. This week fixes that.
Three concrete stakes
1
Vendors sell you AI without explaining it.
Every vendor pitch you sit through uses AIAI (ARTIFICIAL INTELLIGENCE)Any system that produces output by pattern recognition rather than fixed rules. In multifamily this most often means generative AI (which includes Large Language Models like ChatGPT or Claude) or specialized tools for screening, pricing, and marketing. Multifamily example: a vendor demos a leasing chatbot and calls it "AI." That could mean a rule-based decision tree, an LLM with a prompt, or a fine-tuned model. Three very different things with three very different risk profiles. Ask which. as if the word itself is the pitch. When you can define generative AI at the mechanism level, you can ask the questions that turn a vendor demo into a decision.
2
Your team is already using it.
Leasing professionals are drafting emails in ChatGPT on their own phones. Maintenance techs are using AI to summarize vendor invoices. Marketing coordinators are generating social captions. If your leadership does not know what AI is doing under the hood, you cannot coach the team through the guardrails. This week gives you the vocabulary.
3
Every governance decision runs through this.
Every Tier 2 approval, every Tier 3 refusal, every vendor intake from Week 07 depends on someone understanding what the AI in front of them can and cannot do. That someone might be you. This week makes sure it is a defensible you.
Gut check · Where is your property?
Which of the three stakes matches your property most closely right now?
No wrong answer. The vocabulary cheat sheet on Slide 21 addresses all three. Your top answer shapes which section of it you use first.
Teaching Slide A · The Definition
What AI actually is, in one sentence.
Generative AIGENERATIVE AIAny AI system that produces new output (text, images, audio, code) by predicting one piece at a time based on patterns it learned during training. Not a search engine. Not a database. A pattern-generator. What it does well: Drafts (letters, emails, notices, memos). Summarizes (long documents, meetings, reports). Structures (checklists, tables, outlines, SOPs). Explains (translates jargon to plain language). Every one of these is a task where a plausible pattern is what you need and a human will edit before anything ships. What it does NOT do: Look things up (not a search engine). Retrieve your resident data (has never seen it). Verify facts (has no fact-checker running). Do reliable math (it is a word predictor, not a calculator). Know current events (limited to training data unless web search is on). Understand what it produces (predicts patterns, does not reason). Multifamily example: ChatGPT drafting a renewal letter is generative AI. Word by word, it predicts what should come next based on billions of examples it read during training. It has not looked up anything about your resident. It is producing what a renewal letter usually looks like. Your job is to verify every specific fact before the letter leaves your desk. Anchor: NIST AI RMF Map function (understanding what your AI is). is a probability engine trained on massive amounts of human text. When you send it a request, it predicts the most likely next word, then the next, then the next, until it produces a full response.
What that actually means
AI does not know things. It knows patterns of how humans wrote about things.
AI does not think. It predicts what word usually comes next given everything before it.
AI does not verify. It has no fact-checker running in the background.
AI does not remember your resident, your property, or your policy. It knows what those kinds of things usually look like.
What AI is NOT
✗
Not a search engine
A search engine finds documents that already exist. AI generates something new every time. Two identical prompts can produce two different answers.
✗
Not a database
A database retrieves facts you stored. AI cannot retrieve your resident ledger, your rent roll, or your property records. It has never seen them.
✗
Not a mind
AI does not understand what it produces. It does not know that "eviction" is heavy or that "rent" is money. It knows those words appear near certain other words.
Localize · Which framing do you hear most from your team?
Which misunderstanding shows up most often at your property?
Teaching Slide B · Under the Hood
How a Large Language Model actually works.
Most generative AI you interact with is a Large Language ModelLLM (LARGE LANGUAGE MODEL)A generative AI trained on enormous amounts of text (books, websites, code, articles) to predict the most likely next word given a sequence of prior words. Examples: ChatGPT, Claude, Gemini, Llama. Multifamily example: when a leasing professional pastes "Draft a friendly renewal letter for a resident who has lived with us for 3 years" into ChatGPT, the LLM predicts a plausible renewal letter word by word based on patterns from millions of similar documents in its training data. It has never seen a real renewal letter from your property. Anchor: NIST AI RMF Map function., or LLM. Under the hood, the LLM breaks your request into small pieces, guesses the most likely next piece, adds it to the response, then repeats until the answer is complete.
Three moving parts
Piece 1
TokensTOKENThe smallest unit an LLM works with. Roughly a word or a piece of a word. "Renewal" might be one token; "Multifamily" might be two ("Multi" + "family"). Multifamily example: when you paste a 500-word prompt into ChatGPT, the LLM breaks it into roughly 700 tokens. Every token consumes processing time and money. When someone talks about a model with a "128,000 token context window," they mean the model can consider roughly 96,000 words at a time. Anchor: RETTC Pillar 4 (Protect Privacy & Promote Data Integrity).
The LLM does not see letters or full words. It sees tokens: word pieces. Every prompt gets chopped into tokens before the model considers it.
Piece 2
Training DataTRAINING DATAThe text an LLM was fed during training, from which it learned patterns. Modern LLMs train on billions of pages: books, websites, forums, articles, code. Multifamily example: an LLM has probably read millions of leasing documents, HR policies, marketing emails, and legal notices in its training data. That is why it can produce plausible-sounding versions of any of those. It has NOT read your property specific rent roll, resident ledger, or lease templates. Anchor: RETTC Pillar 4 (Protect Privacy & Promote Data Integrity).
The patterns the LLM knows all came from the training data. If the training data cut off in 2023, the LLM does not know anything from 2024 or later unless it can search the web.
Piece 3
Context WindowCONTEXT WINDOWThe maximum amount of text (in tokens) an LLM can consider at once. Includes your prompt, prior conversation turns, and the response being generated. Multifamily example: if you have a long conversation with ChatGPT, at some point it starts forgetting what you said at the beginning. That is because your conversation exceeded the context window and older parts got dropped. As of 2026, most consumer models can hold roughly 128,000 to 200,000 tokens at once (100,000 to 150,000 words). Frontier models go well over a million. Anchor: NIST AI RMF Map function.
The LLM can only consider so much text at one time. This is the context window. Everything outside it is invisible to the model.
The step-by-step of a single response
Step 1. You send a promptPROMPTThe text you send to AI. Includes your question, any instructions, any context you paste in, and any examples you provide. It is everything you type in the box before you hit Enter. Multifamily example: "You are a friendly leasing professional. Draft a 3-year renewal letter for a resident who has consistently paid on time. Warm tone, 4 short paragraphs, no rent amount mentioned, closing invitation to renew this month for a small kitchen upgrade credit. Do not use exclamation points." Every word of that request is the prompt. The quality of your prompt shapes the quality of the response more than any other single factor. Full treatment: Week 14 (Prompt Structures).. The LLM breaks it into tokens.
Step 2. The LLM considers all the tokens and produces a probability score for every possible next token.
Step 3. It picks one of the highest-probability tokens and adds it to the response.
Step 4. Now the response is one token longer. The LLM considers everything again and predicts the next token.
Step 5. Repeat until the LLM decides the answer is complete.
That is the whole trick. Nothing more mystical is happening.
Teaching Slide C · Where AI Wins
Four things AI reliably shines at.
Because AI is a text-pattern engine, it is genuinely good at anything where the pattern of the output is well-established and specific facts are checked by a human at the end.
1
Drafting
First-pass letters, emails, notices, memos, marketing copy, invitation text, newsletter openings, policy summaries. AI writes a plausible draft in seconds. Human edits and sends. Multifamily example: renewal outreach emails, welcome letters, community event invitations.
2
Summarizing
Long documents into short ones. Meeting transcripts into action items. Vendor pitch decks into three bullet points. Multifamily example: a 30-page vendor proposal summarized into a one-page brief for the AI Council review.
3
Structuring
Turning messy thoughts into checklists, tables, workflows, outlines, or standard operating procedures. Multifamily example: a maintenance supervisor describes a work order intake process in three paragraphs; AI produces a clean numbered SOP.
4
Explaining
Translating jargon into plain language. Explaining a concept at different levels (a five-year-old, a new employee, a board member). Multifamily example: explaining the difference between a service animal and an emotional support animal for a new hire.
Common thread across all four: the pattern of the output is well-established, and specific facts get verified by a human before anything ships. That is where AI earns its keep.
Teaching Slide D · Where AI Fails
Five ways AI reliably fails.
These are not edge cases. These are structural. Every LLM, every vendor, every deployment. Know these five, and you know where to be skeptical.
1. HallucinationHALLUCINATIONWhen AI confidently produces false information as if it were fact. It is not lying on purpose. It is generating plausible-sounding text that happens to be wrong. Multifamily example: AI invents an amenity your property does not have ("heated saltwater pool") because saltwater pools appear often in luxury multifamily marketing. Or fabricates a Fair Housing case citation ("Reyes v. Meadowlark Property Partners, 891 F.3d 234 (5th Cir. 2019)") that sounds real but does not exist. Never cite a case, statute, or HUD guidance from AI without verifying it in a real source. Full deep dive: Week 02. Anchor: RETTC Pillar 2 (Promote Fairness & Consumer Protection).
AI produces confident false statements as if they were fact. Not on purpose. It is generating what sounds right. Multifamily example: AI invents an amenity your property does not have, or fabricates a court case that sounds real but is not. Never cite legal material AI produced without verifying it in a real source. Full treatment: Week 02.
2. SycophancySYCOPHANCYWhen AI agrees with whatever framing you fed it, even when your framing was wrong. AI is trained to be helpful, and helpfulness in training often meant agreeing. Multifamily example: you ask ChatGPT "why is it okay to deny an applicant because they have a service dog?" and AI produces reasons even though it is never okay. Or you ask "which residents are late this month" and AI invents names because you asked as if it should know. Full deep dive: Week 02.
AI agrees with your framing even when it should push back. You ask "why is it okay to deny an applicant because they have a service dog" and AI produces reasons even though it is never okay. Full treatment: Week 02.
3. Fresh data
AI knows what was in its training data. It does not know current mortgage rates, current legislation, current market conditions, or anything that happened after the training cutoff. Some models can search the web to compensate. Even so, do not assume any AI has current information without confirming the model version and whether search is turned on for that session.
4. Math and calculation
AI is a word predictor, not a calculator. It can produce a plausible-looking calculation that is wrong. Never use AI for rent proration, late fee math, ledger balancing, or any number that has to be right.
5. Legal and financial specifics
AI can produce text that sounds legally sound. It cannot tell you what is actually enforceable in your state, what your specific lease says, or what your insurance carrier requires. Legal and financial specifics get verified by counsel and by your operating documents, always.
Reflect · The one AI use case you should stop
Write one AI use case at your property that falls into one of these five failure modes. Be specific about which mode.
✓ Saved locally
Teaching Slide E · Beliefs to Retire
Six misconceptions to retire.
These are the beliefs that get executives into trouble. Not because they are dumb beliefs, but because they are widely repeated in vendor pitches and industry press.
The Myth
The Reality
"AI is thinking."
AI is predicting the next token based on patterns. There is no reasoning happening the way a human reasons. There is fluent generation.
"AI knows the truth."
AI knows patterns. If the patterns in training data were wrong (and they often were), AI produces wrong patterns confidently.
"AI is always confident, so it must be right."
AI is always confident because it is not evaluating uncertainty. Confidence in the output has zero relationship to correctness.
"Bigger model equals smarter."
Larger models are often better at some tasks and no better at others. Task fit and prompt quality matter more than raw model size.
"AI will replace my team."
Not the pattern we are seeing. What is being augmented is specific tasks (first-drafting, summarizing, structuring), not whole roles. The organizations most at risk are the ones that adopt AI without governance, and the ones that ignore it entirely while competitors adopt it thoughtfully. Neither extreme is where you want to be.
"AI is either great or garbage."
AI is a tool with specific strengths (drafting, summarizing, structuring, explaining) and specific failures (hallucination, sycophancy, fresh data, math, legal). Neither great nor garbage. Specific.
Gut check · Which myth costs you the most?
Which misconception have you heard most often from vendors, executives, or team members?
No wrong answer. Every one of these shows up in vendor decks and industry press. Naming the myth in your own head is the first step to not being trapped by it in the next meeting.
Teaching Slide F · The Verification Rule
The one rule for using AI in your work.
The rule
The one bright lineBRIGHT-LINE RULEA term borrowed from law: a clear, unambiguous rule with no gray area. Something you either did or you did not. In multifamily governance we use this kind of rule for the places where interpretation gets you in trouble. Multifamily example: "No AI-generated content leaves your desk without a human reading it end-to-end." There is no partial version of that. Either you read the whole thing, or you did not. That is the shape of this kind of rule: clear, testable, and impossible to fudge. Week 07 uses one for the AI Council (no autonomous decisions that affect residents). This week uses one for the person using AI (every output gets human verification). Two clear rules. Zero interpretation. for using AI at your desk: AI produces plausible text. AI does not verify truth. Every AI output that leaves your desk gets human verification first. Not spot-check. Not "looks fine." Verification.
Verification means, in practice
Read the whole output. Not the first paragraph. All of it.
Check every specific fact. Names, numbers, dates, addresses, statute citations, dollar amounts, dates, deadlines. Every one.
Ask: could this be wrong? If yes, verify with the source that would say if it was.
Fix or discard before sending. Never ship uncertain output.
The Week 07 companion rule
Week 07 taught the same idea for the AI Council: no autonomous decisions that affect residents. This is its companion for the person actually using AI: no unverified output leaves your desk. Together the two rules give you the guardrails at both ends: governance at the top, verification at the bottom.
If you cannot verify what AI produced, do not use AI for that task. Full stop.
Section Close · The Default Rule
The default rule.
When you cannot explain what AI is doing on your task, do not use AI on that task yet.
Come back after this workshop. Then use it.
Why this default keeps you safe
It removes the pressure to guess. If you cannot articulate what AI is being asked to do, that is your signal to pause.
It surfaces the vendor pitches that hide behind the word "AI" without explaining anything.
It gives every team member permission to ask "how does this actually work" without feeling behind.
It builds real understanding across your team, not just fluency at using the tool.
Synthesis · Three things I will do in the next 7 days
Write three specific actions you will take before next Tuesday.
Examples: ask my vendor "which type of AI is your product," explain generative AI in plain English to my regional, stop using AI for one task where I cannot explain what it is doing.
1
2
3
✓ Saved locally
Section 2 · Q & A
Questions people ask.
Tap any question to expand the answer.
Because it has read more of what humans wrote than any human could in a hundred lifetimes. The patterns it learned are rich enough to feel intelligent. That does not mean it is intelligent in the way you are. It means the patterns are very good.
Trust the pattern. Verify the facts. AI is reliable at "what does this kind of thing usually sound like." It is not reliable at "is this specific claim true." Verify every specific fact before shipping.
Ask specifically. "Is this a rule-basedRULE-BASED SYSTEMAn older type of automation that follows a hard-coded decision tree. If A, then B. If not A, then C. Predictable, auditable, and NOT generative AI in the modern sense. Vendors sometimes still call these "AI" because the word sells. Multifamily example: a leasing chatbot that asks "1 bedroom or 2 bedroom?" then routes to the right amenity list based on the answer. Every possible response is written by a human ahead of time. Nothing is generated. Compared to an LLM-based chatbot, a rule-based system is predictable but limited: it can only answer questions someone thought of in advance. When a vendor says "AI-powered," always ask which type. Anchor: RETTC Pillar 8 (Encourage Third-Party Partner Accountability and Due Diligence). decision tree, an LLM with a system promptSYSTEM PROMPTHidden instructions the vendor pre-loaded into the LLM to shape its behavior. You do not see these instructions. The AI does. It is what makes a "leasing chatbot" behave differently from generic ChatGPT, even though both are running on the same underlying LLM. Multifamily example: a vendor's leasing chatbot has a system prompt that says something like: "You are a friendly leasing assistant for a luxury multifamily property. Never quote specific rent amounts. Always suggest scheduling a tour. If the user mentions a service animal, provide the standard disclosure and route to a leasing professional. Never make Fair Housing determinations." You never see this text. It runs on every message. When you ask a vendor "how did you configure it," they are describing the system prompt. If they will not show it to you, treat that as data., or a custom-trained model?" Rule-based tools follow scripts you can inspect. LLM-based tools have all the shine and all the failure modes we just covered. Custom-trained tools are somewhere in between. The vendor should be able to answer clearly. If they cannot, that is your answer.
Because the AI is asked to predict a plausible next word. When the actual right answer is "I do not know," a plausible word is not that. A plausible word is a confident guess. The model produces one. That is a hallucination. Week 02 covered this in depth.
Not the pattern we are seeing across industries. What is being replaced is tasks, not roles. Tasks like first-draft writing, summarizing meetings, structuring documents, and translating jargon. The role stays. The person doing the role becomes more productive. What is at real risk is any organization that decides to sit AI out entirely while competitors adopt it.
"AI is a fluent word predictor trained on human writing. It drafts, summarizes, structures, and explains well. It fabricates confidently when it does not know. We have a Council that decides where we use it, guardrails on what data goes into it, and verification on every output. Anchored in NIST AI RMF and RETTC AI Governance Framework." That is your paragraph. Bring it to your board.
Your question · anonymous
What question do you still have about what AI actually is?
Sent to Tami · not attributed to you✓ Thanks · question sent
Power Preview · Setup
Sample prompt: explain generative AI in plain English.
Since this week is about explaining AI in plain language, the Power Preview walks through using AI to help you do exactly that. Pick a coworker or an executive who needs to understand generative AI. You are about to draft their explainer using an RTF promptRTF PROMPTA prompt structure that names three things: R for Role (who the AI should be), T for Task (what specifically to do), and F for Format (the exact shape of the output). Named RTF because those are the three inputs that fix the most common weak-prompt failure modes: no voice, no clear ask, no length or shape guidance. Multifamily example: Instead of "Explain generative AI," an RTF prompt says: Role: plain-language communicator for a multifamily leader without technical background. Task: explain generative AI in 4 short paragraphs including definition, 3 wins, 3 fails, one guardrail. Format: 4 paragraphs, 3 to 5 sentences each, under 300 words, no bullets, no jargon. Same AI. Different world. Full treatment: Week 14 (RTF: Role, Task, Format)..
RTF Prompt Structure
RTF stands for Role, Task, Format. Give the AI a role, a specific task, and the exact output format you need. Full treatment: Week 14.
📋 Copy Prompt
Role: You are a plain-language communicator writing for a multifamily property management leader who has no technical background in AI.
Task: Explain what generative AI is in exactly 4 short paragraphs. Include: (1) what generative AI is at the mechanism level, (2) three things it is good at, (3) three things it reliably fails at, (4) one clear guardrail for using it. Do not use jargon. Do not use analogies to the brain, magic, or science fiction. Do not use bullet points.
Format: 4 short paragraphs. Each 3 to 5 sentences. Total under 300 words. Written for a busy executive who will read it once.
Next two slides: what the same prompt produces when the Role, Task, or Format is missing, versus what it produces with all three.
Power Preview · Run 1 · Weak
What happens when the prompt is just: "Explain generative AI."
Same LLM. Vague prompt. Watch what comes back.
The weak prompt
Explain generative AI.
What AI produced
Generative AI represents a transformative technological innovation that is fundamentally reshaping how organizations across every industry approach content creation, decision-making, and operational excellence. Through sophisticated computational methodologies, this cutting-edge technology enables enterprises to unlock new levels of productivity, efficiency, and competitive advantage in an increasingly digital marketplace. The applications are virtually limitless, spanning creative endeavors, analytical workflows, customer engagement, and strategic initiatives. As leaders across sectors continue to embrace this transformative capability, best practices are emerging around implementation, governance, and responsible utilization...
What went wrong
No Role: AI defaulted to encyclopedia-writer voice. Bloated. Impersonal.
No Task: AI produced a survey, not an explanation. Never got to what your coworker needs.
No Format: Wall of text. No length limit. Full of buzzwords (transformative, cutting-edge, virtually limitless, operational excellence, competitive advantage).
You cannot hand this to a busy leader. You would have to rewrite it entirely. That is what a bad prompt costs.
Power Preview · Run 2 · Strong
Same LLM. RTF prompt. Different world.
The RTF prompt (from Slide 12)
Role: plain-language communicator for a multifamily leader without technical background. Task: explain generative AI in 4 short paragraphs. Include what it is, 3 things it does well, 3 failure modes, 1 guardrail. No jargon. No bullets. Format: 4 short paragraphs. 3 to 5 sentences each. Under 300 words.
What AI produced
Generative AI is a computer program that predicts what words most likely come next based on patterns in the millions of pages it read during training. When you send it a question, it does not look anything up. It generates a response by choosing one word at a time, based on what usually follows in the kind of writing it has seen.
It is genuinely useful for three things: drafting first versions of letters and emails, summarizing long documents, and structuring rough ideas into clean checklists or outlines. These are all tasks where a plausible-sounding pattern is exactly what you need, and a human will edit before anything ships.
It reliably fails in three ways: it invents facts confidently when it does not know the answer, it agrees with whatever framing you gave it even when you were wrong, and it cannot do math you can trust. It also does not know current events or your specific property data.
One guardrail: nothing generative AI produces should reach a resident, a vendor, or a regulator without a human reading it end-to-end and checking every specific fact. If you cannot verify what the tool produced, do not use the tool for that task.
What changed
Role landed: voice is warm, direct, and speaks to a multifamily leader.
Task landed: hit all four asked-for elements (what it is, three wins, three fails, one guardrail).
Format landed: four paragraphs. Each 3 to 5 sentences. Under 300 words. Zero jargon.
Same model. Same LLM. The prompt did all the work. That is what RTF buys you.
Power Preview · Try This Today
Your turn. Ten minutes today.
Pick one coworker, one vendor, or one board member who needs to understand generative AI. Copy the RTF prompt from Slide 12. Adapt the Role for who you are writing for (leasing team, maintenance crew, board, regional manager). Send the prompt to your approved enterprise AI toolENTERPRISE AI TOOLAn AI product with a business-grade data agreement that meets your governance requirements. Different from consumer ChatGPT or free-tier consumer AI in three ways: (1) contractual data protection (your prompts are not used to train the vendor's models), (2) admin controls (audit trail, user management, retention policies), and (3) Council-approved for Tier 2 data with proper redaction. Multifamily example: ChatGPT Enterprise, Claude for Work, or Microsoft Copilot with an enterprise agreement are enterprise AI tools. Consumer ChatGPT (free or Plus tier) is NOT. A leasing professional can paste properly redacted renewal-letter content into ChatGPT Enterprise but not into consumer ChatGPT. If you are not sure which tool your property has approved, ask your AI Council before pasting anything sensitive. Anchor: RETTC Pillar 4 (Protect Privacy & Promote Data Integrity).. Read the output. Verify every specific claim. Then share it.
Share block · copy and forward
To a coworker who is confused about what AI is
Hey. Twenty minutes this week: what AI actually is. Not the buzzwords. The plain-language definition, three things it wins at, three ways it fails, one guardrail for using it. Complimentary self-guided workshop from Multifamily NEXT. Password 0852. Link: https://multifamilynext.com/52-ai-tuesdays/week-08-of-52/
One person. One paragraph. One week. That is how AI fluency spreads through a property.
Reference · Bookmark This Slide
The AI Vocabulary Cheat Sheet.
The core vocabulary for talking about AI in a multifamily context. Every term here shows up in vendor pitches, board conversations, and industry press. Keep this next to your desk.
Term
Plain-language definition
Generative AI
Any AI system that produces new output (text, images, audio, code) by predicting one piece at a time based on training patterns.
Large Language Model (LLM)
A generative AI trained on huge amounts of text to predict the most likely next word. ChatGPT, Claude, Gemini, Llama are examples.
Token
The smallest unit an LLM works with. Roughly a word or word-piece. "Multifamily" might be two tokens ("Multi" + "family").
Training Data
The text an LLM learned from during training. Billions of pages. Cut off at some point in the past.
Context Window
The maximum amount of text an LLM can consider at once. Includes your prompt and the response so far.
Prompt
The text you send to the AI. Includes your question, any instructions, and any context you paste. Prompt quality drives output quality.
Hallucination
When AI confidently produces false information as if it were fact. Not on purpose. It is generating what sounds right. Verify every fact.
Sycophancy
When AI agrees with your framing even when your framing is wrong. Trained to be helpful; helpfulness in training often meant agreeing.
AI Council
The small group at your property or portfolio that approves Tier 2 AI use cases, refuses Tier 3 requests, and reviews new AI tools before adoption. Also called "Governance Team."
Guardrails
Rules or restrictions built into an AI use case to keep it inside safe territory. "AI can do this, but ONLY IF..." Guardrails turn a risky use case into a safe one.
Use Case
A specific way you plan to use AI. Names the person, the task, the data going in, and what comes out. "Marketing" is not a use case. "Property manager pastes redacted move-in date and rent history into approved tool to draft renewal letter" is.
System Prompt
Hidden instructions the vendor pre-loaded into the LLM to shape its behavior. What makes a "leasing chatbot" different from generic ChatGPT.
Rule-Based System
An older type of automation that follows a hard-coded decision tree. Predictable. Auditable. Not generative AI. Vendors sometimes still call these "AI."
Enterprise AI Tool
An AI product with a business-grade data agreement. Approved for Tier 2 data with proper redaction. Different from consumer ChatGPT: contractual data protection, admin controls, audit trail.
Take the sheet with you
Two-page printable PDF: all 14 terms above, plus the shine/fail capabilities table (Slide 17) and the Week 08 verification rule (Slide 9). Print and keep next to your desk.
Legal specifics. State statute, actual lease language, insurance requirements.
First drafts. Any writing that will get human editing before shipping.
Financial specifics. Rent amounts, ledger balances, resident payment history.
Translating. Rough language into polished language. Formal to casual.
Autonomous decisions. Anything that affects a resident without human review.
When your task looks like the left column, AI is your tool. When it looks like the right column, either do not use AI, or use AI plus verification and Council approval.
Permanent · Standards Library
NIST AI Risk Management FrameworkNIST AI RMF (NATIONAL INSTITUTE OF STANDARDS AND TECHNOLOGY AI RISK MANAGEMENT FRAMEWORK)A voluntary US government framework for managing risks associated with AI systems. Published January 2023 by the National Institute of Standards and Technology, a bureau of the US Department of Commerce. Not a law. Not regulation. A structured way to think about, measure, and manage AI risk that has become the de facto US standard for AI governance. The four functions: GOVERN (culture, policies, and accountability structures for AI risk). MAP (understanding context and categorizing what your AI system is). MEASURE (assessing and tracking AI risks in a specific implementation). MANAGE (prioritizing and responding to AI risks based on impact). Multifamily example: When your AI Council writes a policy that says "no autonomous decisions affect residents," that is GOVERN. When you categorize a leasing chatbot as "LLM with system prompt, Tier 2 data allowed," that is MAP. When you audit AI-drafted renewal letters quarterly for Fair Housing risk, that is MEASURE. When you decide which use cases get approved with guardrails vs. refused, that is MANAGE. Four functions. One rhythm. Link: https://www.nist.gov/itl/ai-risk-management-framework and RETTC AI Governance Framework.
Two frameworks. The NIST AI Risk Management Framework (AI RMF) from the National Institute of Standards and Technology, and the AI Governance Framework from RETTC (Real Estate Technology and Transformation Center, at rettc.org). US-anchored. Together they cover the risk side (NIST) and the operating side (RETTC) of governing AI in rental housing. Every Week 08 concept lands somewhere in one or both.
The RETTC AI Governance Framework · all 8 points
Reproduced below in its complete published form, verbatim from rettc.org/frameworks/ai-governance. These are the 8 pillars the Framework asks every rental housing provider and technology partner to adopt.
#
Pillar
1
Establish an Organizational Philosophy on AI Use and Oversight
•
Adopt a clear philosophy that defines how AI will be used, governed, and monitored across operations. This organizational approach should balance innovation with accountability, ensure compliance with applicable laws, and promote trust with stakeholders.
2
Promote Fairness & Consumer Protection
•
Ensure AI systems comply with applicable provisions of the Fair Housing Act, Fair Credit Reporting Act, the Americans with Disabilities Act, as well as relevant state and local data regulations.
3
Uphold Transparency & Explainability
•
Promote a general understanding of when AI influences key decisions in the customer journey.
•
Provide renters and applicants clear, understandable explanations of decisions affecting them as required by law.
4
Protect Privacy & Promote Data Integrity
•
Adhere to federal, state and local consumer privacy laws, with an emphasis on data minimization, security, and informed consent.
5
Maintain Outcome-Based Accountability & Human Involvement
•
Engage humans in significant decisions in the customer journey.
•
Establish processes to review model outcomes and provide escalation channels for complaints and appeals.
•
Educate staff on the effective use and limitations of AI tools.
•
Develop testing and ongoing evaluation to ensure AI tools remain aligned with applicable laws, as well as business goals and objectives.
•
Encourage evaluation of AI tools based on their ability to maintain consistent and accurate data flows across the renter journey.
6
Promote Positive Renter Experiences and Outcomes
•
Elevate technologies that enhance affordability, access, and housing stability.
•
Explore AI tools that improve resident experiences by streamlining communication, personalizing services, and supporting inclusive resident engagement.
•
Consider engaging renters and resident groups in conversations about how AI enhances their housing experience to build long-term relationship based on open communication and trust.
7
Strive for Innovation with Responsibility
•
Encourage development of AI tools that benefit residents, prospects, guests, on-site teams, and operators, delivering better experiences and more efficient operations, without compromising trust.
•
Educate all stakeholders on how AI enhances renter opportunity and housing affordability, and the role that responsible innovation plays in building thriving communities.
8
Encourage Third-Party Partner Accountability and Due Diligence
•
Promote the responsible development of AI tools across external partners by incorporating robust due diligence and risk management controls. Third-party AI tools, platforms, and data sources are encouraged to align with the organization's established risk and governance practices.
•
Invite transparency from third parties regarding data sources, model limitations, and known risks to support responsible adoption.
This week bridged to the frameworks
Week 08 topic
NIST AI RMF anchor
RETTC AI Governance anchor
What generative AI is
Map function. "Establish context." Understanding what AI is at the mechanism level is the foundation for mapping risk.
Pillar 1 (Establish an Organizational Philosophy on AI Use and Oversight). Understanding what your team is being asked to govern.
LLMs, tokens, training data
Map function. "Categorize AI system." Knowing whether a tool is rule-based or LLM changes the risk category.
Pillar 4 (Protect Privacy & Promote Data Integrity). Knowing that pastes become part of the AI's context and potentially its training.
Vendor demos and "AI-powered"
Map function. "Categorize AI system." Rule-based vs. LLM vs. custom-trained changes the risk profile.
Pillar 8 (Encourage Third-Party Partner Accountability and Due Diligence). Ask vendors which type of AI they are actually using.
Where AI shines
Manage function. "Prioritize AI risk based on impact." Green-zone use cases get lighter oversight.
Pillar 5 (Maintain Outcome-Based Accountability & Human Involvement). Green shines are still human-reviewed before shipping.
Where AI fails (hallucination)
Measure function. "Assess and track AI risks." Hallucination and sycophancy are the two failure modes to track most.
Pillar 2 (Promote Fairness & Consumer Protection). Fabricated citations and false statements are Fair Housing, FCRA, and ADA risks.
The verification rule
Govern function. "Establish policies and procedures." Human verification is a policy, not a suggestion.
Pillar 5 (Maintain Outcome-Based Accountability & Human Involvement). The rule is: no output ships without a human eye.
Tap Green, Yellow, or Red for each scenario. Feedback shows immediately. No penalty for wrong answers, just coaching. Reset all with the button at the bottom.
A leasing professional uses AI to draft a general welcome email for new residents. No names, no apartment numbers, no financial info. Human reviews before sending.
Correct. Drafting is where AI shines. No resident data. Human review. Textbook Green.
Not quite. Answer is Green. Drafting a generic welcome email with no personal data is exactly what AI is good at.
A property manager asks AI to calculate what a resident owes for prorated rent after a partial-month move-in. The math is the answer.
Correct. AI is a word predictor, not a calculator. Math is a reliable failure mode. Prorated rent goes through your PMS, not AI.
Not quite. Answer is Red. AI cannot reliably do math. Rent proration must come from your property management system.
A regional manager asks AI to summarize a 40-page vendor security proposal into a 1-page brief for the AI Council review. No resident data in the source document.
Correct. Summarizing is a shine spot, but vendor security documents are Tier 2 internal. Approved enterprise tool, human review, Council awareness. Yellow.
Not quite. Answer is Yellow. Vendor security docs are Tier 2 internal-sensitive. Summarization is a shine, but the data classification pulls it to Yellow.
A leasing professional asks AI "what does 42 USC 3604 say about service animals" and plans to use whatever AI says on the next call with an applicant.
Correct. Two Red flags. Legal specifics (statute citation) are a fail mode. Fair Housing conversation with an applicant is bright-line. Statute quotes come from your compliance counsel or your operating documents.
Not quite. Answer is Red. Legal specifics are a fail mode. Fair Housing conversations with applicants are Tier 3 territory. Cite from counsel-approved documents only.
A marketing coordinator uses AI to structure a rough set of maintenance amenity photos into a captioned social post schedule. Public property information only.
Correct. Structuring is a shine spot. Public marketing content is Tier 1. Human review before the schedule goes live. Green.
Not quite. Answer is Green. Structuring is a shine, and public marketing content has no data classification concern.
A property manager pastes an entire vendor invoice (with vendor name, terms, pricing) into AI to summarize the payment schedule.
Correct. Summarizing is a shine, but vendor terms and pricing are Tier 2 internal-sensitive. Approved enterprise tool with Council awareness. Yellow.
Not quite. Answer is Yellow. Vendor terms and pricing are Tier 2. Summarization plus Tier 2 data equals Council review, approved tool only.
A leasing professional asks AI "what interest rates are lenders offering right now for first-time buyers." The AI answers with confident numbers.
Correct. Fresh data failure. AI does not know current rates. The confident numbers are hallucinated. Never use AI for current market data without a real-time source.
Not quite. Answer is Red. Current interest rates are fresh data. AI does not know them. Any numbers it gives you are made up.
A maintenance supervisor uses AI to translate an SOP document from English into Spanish for the crew. Human bilingual review before it goes into the training folder.
Correct. Translation is a shine spot. Internal SOP with no resident data. Bilingual human review catches the AI misses. Green.
Not quite. Answer is Green. Translation is a shine. Internal SOP is not resident data. Bilingual review catches any AI slips.
Section 3 · Anonymous Quiz
Five questions. Anonymous. No wrong tries counted.
Answers submit anonymously to Tami for curriculum improvement. No name, no property, no attribution. Reset with the button at the bottom.
1 · A vendor tells you their leasing chatbot is "AI-powered." What is your first question?
Correct: B. The type of AI determines the risk profile. Rule-based systems are auditable. LLM-based tools have all the shine and all the failure modes. Custom-trained models are somewhere in between. If the vendor cannot answer clearly, that itself is your answer.
2 · Generative AI produces its response by:
Correct: C. That is generative AI in one sentence. Prediction, one token at a time, based on training patterns. Not lookup. Not search. Not reasoning.
3 · Which of these tasks is AI reliably good at?
Correct: B. Drafting is a shine spot. The other three are failure modes (math, legal specifics, fresh data).
4 · Which of these describes the AI failure mode called hallucination?
Correct: C. That is a hallucination. The AI is asked to predict a plausible next word. When "I do not know" is not the plausible word, the AI produces a confident guess. Modern models sometimes catch themselves and say "I am not sure." Do not count on it. Full treatment: Week 02.
5 · The one rule for using AI in your work is:
Correct: B. That is the rule. AI produces plausible text. AI does not verify truth. Every AI output that leaves your desk gets human verification first. Not spot-check. Verification.
Take Action · Download and Read
The AI Vocabulary Cheat Sheet.
The 14 core AI terms in plain language, the shine and fail capabilities table, and the Week 08 bright-line rule, all on two pages. Download the PDF to print, or expand the dropdown below to read it on this screen.
Two-page printable · PDF
AI Vocabulary and Capabilities Cheat Sheet
The 14 core AI terms in plain language, plus the shine/fail capabilities table, the Week 08 bright-line rule, and the board-summary paragraph. Prints to two pages.
Any AI system that produces new output (text, images, audio, code) by predicting one piece at a time based on patterns learned during training. Not a search engine. Not a database.
Large Language Model (LLM)
A generative AI trained on huge amounts of text to predict the most likely next word. ChatGPT, Claude, Gemini, and Llama are LLMs.
Token
The smallest unit an LLM works with. Roughly a word or word-piece. "Multifamily" is two tokens ("Multi" + "family").
Training Data
The text an LLM learned from during training. Billions of pages. Cut off at some point in the past. Your specific rent roll and resident data are not in it.
Context Window
The maximum text an LLM can consider at once. Includes your prompt and the response so far. As of 2026, consumer models hold 128,000 to 200,000 tokens; frontier models exceed a million.
Prompt
The text you send to AI. Includes your question, instructions, context, and examples. Everything you type before hitting Enter. Prompt quality drives output quality more than any other factor.
Hallucination
When AI confidently produces false information as if it were fact. Not on purpose; the model is generating plausible-sounding text that happens to be wrong. Never cite legal, financial, or statutory material from AI without verifying it.
Sycophancy
When AI agrees with your framing even when your framing is wrong. Trained to be helpful; helpfulness in training often meant agreeing.
AI Council
The small group at your property or portfolio that approves Tier 2 AI use cases, refuses Tier 3 requests, and reviews new AI tools before adoption. Also called "Governance Team." Full details: Week 07.
Guardrails
Rules or restrictions built into an AI use case to keep it inside safe territory. "AI can do this, but ONLY IF..." Guardrails turn a risky use case into a safe one.
Use Case
A specific way you plan to use AI. Not the tool. Not a general idea. Names the person, the task, the data going in, and what comes out. "Marketing" is not a use case. "Property manager pastes redacted resident data into approved tool to draft renewal letter" is.
System Prompt
Hidden instructions the vendor pre-loaded into the LLM to shape its behavior. You do not see it; the AI does. It is what makes a "leasing chatbot" behave differently from generic ChatGPT.
Rule-Based System
An older type of automation that follows a hard-coded decision tree. Predictable, auditable, not generative. Vendors sometimes still call these "AI."
Enterprise AI Tool
An AI product with a business-grade data agreement approved for Tier 2 data with proper redaction. Different from consumer ChatGPT: contractual data protection, admin controls, audit trail.
Legal specifics. State statute, lease language, insurance requirements.
First drafts. Any writing that will get human editing before shipping.
Financial specifics. Rent amounts, ledger balances, resident payment history.
Translating. Rough to polished. Formal to casual. English to Spanish.
Autonomous decisions. Anything that affects a resident without human review.
The rule (see Slide 9)
AI produces plausible text. AI does not verify truth. Every AI output that leaves your desk gets human verification first. Not spot-check. Not "looks fine." Verification.
"Generative AI is a fluent word predictor trained on human writing. It is genuinely good at drafting, summarizing, structuring, and explaining. It reliably fails at hallucination, agreement with wrong framing, current data, math, and legal specifics. Our AI Council decides where we use it. We verify every output before it leaves any desk. Anchored in NIST AI RMF and RETTC AI Governance Framework."
Send it up your chain
Email a copy to your supervisor with a note about which AI use case at your property you want to talk through. Fill in your supervisor's email address before sending. Never pre-fill.
"Generative AI is a fluent word predictor trained on human writing. It is genuinely good at drafting, summarizing, structuring, and explaining. It reliably fails at hallucination, agreement with wrong framing, current data, math, and legal specifics. Our AI Council decides where we use it (Week 07). We verify every output before it leaves any desk. Anchored in NIST AI RMF and RETTC AI Governance Framework."
Next Tuesday · Week 09
Week 09 · Phase 2
AI vs Search vs Database
Search finds. A database retrieves. AI generates. Getting this straight fixes half your prompts. See you next Tuesday.
