Most small commercial real estate firms don’t have an AI tools problem. They have a practice problem. JLL’s 2025 Global Real Estate Technology Survey found that roughly nine in ten CRE investors and occupiers are piloting AI, yet only 5% report hitting all of their program goals. The gap between those two numbers is not software. It is people who open ChatGPT twice, get a mediocre lease summary, and quietly go back to doing everything by hand. This playbook lays out a 90-day training program a 4–20 person firm can run without an IT department: weekly drills tied to real deliverables, data-security ground rules you can adopt in an afternoon, and concrete milestones that define what “fluent” means for a broker, a property manager, and an acquisitions lead.
Why training is the gap, not tools
The tools are already good enough. ChatGPT, Claude, and Gemini can all draft a credible LOI, summarize a 60-page lease, and turn messy call notes into a clean market write-up. None of that requires proptech subscriptions, integrations, or a data science hire. It requires someone at your firm who knows how to ask, how to check the output, and how to fold the result into a workflow they run every week.
The industry data backs this up from the institutional side. Deloitte’s 2026 Commercial Real Estate Outlook, built on a survey of more than 850 executives across 13 countries, recommends treating AI literacy as a board-level pillar with companywide learning programs, because returns lag when human adaptation lags. Institutions respond by appointing AI leads and standing up centers of excellence. A 10-person shop can’t do that, and shouldn’t try.
Here is the small-firm advantage nobody talks about: you can retrain your entire company in one quarter. At a global brokerage, rolling out AI training is a change-management program measured in years. For you it is a dozen working sessions and a shared prompt library. The small CRE firm AI manifesto makes the full argument for why small shops can out-operate institutional giants on AI adoption; this playbook is the training leg of that argument.
One caveat before the program. Training fails when it is framed as “learn AI.” It works when it is framed as “produce this Tuesday’s deliverables faster.” Every drill below is anchored to a document your firm already produces: an LOI, a lease summary, a market write-up, a tenant email. That anchoring is the whole method.
What fluency means at a 10-person firm
Fluency is not the ability to explain how a language model works. It is the ability to complete your own recurring tasks with AI assistance, faster than you could without it, while catching the errors that matter. A useful definition has three parts.
Task completion. Each person can run their three most common writing and analysis tasks through ChatGPT, Claude, or Gemini and get a usable first draft in one or two prompts, not ten.
Error detection. Each person knows the failure modes of their tasks. A broker checks invented comps. A property manager checks lease clause references against the actual document. An acquisitions analyst checks every number in a generated summary against the source rent roll.
Habit. The firm’s default for drafting work has flipped from blank-page-first to AI-first-then-edit. That flip, more than any individual skill, is what the 90 days are designed to produce.
Concrete milestones make this testable. By day 90, everyone at the firm should be able to hit targets like these:
| Milestone | Role | Target |
|---|---|---|
| First-pass LOI from deal terms | Broker / principal | Under 15 minutes, one review cycle |
| Lease summary from a PDF | Property manager | Key terms extracted and verified in under 30 minutes |
| Market write-up draft from comps and notes | Broker / analyst | Structured draft in under 20 minutes |
| Tenant or client email in firm voice | Everyone | Under 5 minutes, no template hunting |
| Deal screen memo from an OM | Acquisitions | One-page screen in under 25 minutes, numbers verified |
If those targets look aggressive, that is the point. They are achievable precisely because the underlying tasks are language tasks, and language is what these models do best.
Before day 1: ground rules for confidential deal data
Security anxiety kills more small-firm AI adoption than skepticism does, and the anxiety is legitimate. Your firm handles rent rolls, tenant financials, and deal terms under NDA. Nobody at a 10-person shop has time to read a 40-page data processing agreement, so set three ground rules before the first training session.
Rule 1: Use business-tier accounts, not personal free accounts. OpenAI, Anthropic, and Google all publish data-handling terms for their business plans, and those plans state that business inputs are not used for model training by default. Verify the current terms for the specific plan you buy at purchase time, because terms change. The practical move: one firm-wide subscription with workspace controls, not ten personal logins.
Rule 2: Classify before you paste. A simple two-bucket rule works better than a policy document. Green bucket: anything you would put in an email to a counterparty (listing details, public comps, your own draft language). Red bucket: anything under NDA or containing tenant personal information, which either stays out of the tool or gets anonymized first (strip names, addresses, and identifying figures before pasting).
Rule 3: Name a data owner. One person, usually the ops director or a principal, owns the question “can this go in the tool?” When someone is unsure, they ask, and the answer gets added to a running one-page FAQ. After a month, the FAQ covers nearly every case that comes up and the questions taper off.
That is the whole policy. Firms that start with a 20-page AI governance document written by outside counsel tend to still be reviewing it when their competitors are on drill week six.
The 90-day program
The program runs in four phases. Total structured time commitment: about 90 minutes per person per week, most of it applied to work they were going to do regardless. That framing matters when you pitch it to the team, because “90 minutes a week doing your own work with a new method” lands differently than “training program.”
Days 1–14: baseline and setup
Start by measuring where you are, because the before-and-after is what sustains the program when enthusiasm dips around week five.
First, run a task audit. Each person lists their five most frequent writing or analysis tasks and estimates the time each takes today. For a typical small firm the list converges fast: LOIs and proposals, lease reviews and summaries, market and property write-ups, client and tenant email, deal screening memos. This audit becomes your drill syllabus, so don’t skip it.
Second, make the tooling decision. Pick one primary general-purpose tool (ChatGPT, Claude, or Gemini) for the whole firm rather than letting everyone choose independently. A shared tool means shared prompts, shared lessons, and one bill. If your firm lives in Microsoft 365, Microsoft Copilot is a reasonable primary because it sits inside Outlook and Word, though in our workshop experience the standalone chat tools are better for the drill phase because they force people to learn prompting rather than clicking suggested actions.
Third, decide pilot versus full-team. We recommend training everyone at once when the firm is under about eight people; the social effect of the whole firm drilling together outweighs the efficiency of a pilot. Above that, start with a pilot group of three or four volunteers, let them get two weeks ahead, then use their wins to pull in the rest. Never assign the pilot to the most senior skeptic or to the office tech enthusiast alone. The best pilot is a respected mid-tenure producer who is busy, mildly curious, and known for judgment.
By day 14 you should have: the task audit on one page, business accounts provisioned, the three data rules circulated, and a scheduled weekly 45-minute working session for the next ten weeks.
Days 15–45: fundamentals through drills
This phase teaches prompting, but never as an abstraction. Each weekly session takes one task from the audit and turns it into a drill with the same five-step shape:
- Bring a real artifact. Last week’s actual LOI, an executed lease from a closed file, a genuine market write-up.
- Prompt cold. Everyone attempts the task with the AI, no guidance, ten minutes.
- Compare outputs. Read three or four results aloud. The gap between the best and worst output in the room is the lesson, and it is usually a gap in context provided, not cleverness.
- Build the firm prompt. Together, write the reusable prompt: role framing, the firm’s format conventions, what to include, what to flag as uncertain. Save it in a shared document. This shared prompt library becomes a real asset; by day 90 it is the closest thing your firm has to documented process.
- Assign the week’s reps. Everyone uses the new prompt on live work at least three times before the next session and notes where it failed.
Three fundamentals get taught through these drills rather than as lecture material. Context beats phrasing: pasting in the deal terms, the comp set, and a prior example improves output more than any clever wording. Iteration beats one-shotting: the second prompt (“tighten the rent escalation paragraph, cite the clause number”) is where the value is. And verification is part of the task: every drill ends with checking the output against source documents, so checking becomes muscle memory, not an afterthought.
Weeks one through four of this phase cover, in order: email and client correspondence (easiest win, builds confidence), LOI and proposal drafting, lease summarization, and market write-ups. That ordering is deliberate. Start where the stakes are lowest and the time savings are most obvious.
Days 46–75: workflow drills by role
Now the drills split by role, because a 10-person CRE firm is three or four firms sharing an office. The weekly session stays joint (the cross-pollination is valuable) but the reps diverge:
| Role | Drill focus | Example weekly rep |
|---|---|---|
| Brokerage | Prospecting and marketing copy, comp narratives, OM sections | Draft three property narratives from raw comp data and photo notes |
| Property management | Lease abstraction, tenant communications, maintenance triage summaries | Abstract two leases into the firm’s summary template and verify against source |
| Acquisitions / investment | Deal screening memos, OM analysis, investor update drafts | Screen one live OM into a one-page memo with a verified numbers table |
| Ops / admin | Meeting notes to action items, document formatting, CRM hygiene notes | Convert one weekly meeting recording’s notes into assigned action items |
Two practices distinguish firms that get fluent from firms that stall in this phase. First, chain tasks: the lease summary produced on Tuesday becomes the input for the tenant email on Wednesday, which teaches people that AI output is a working material, not a final product. Second, log failures in the shared library next to the prompts. “Invented a comp on Main St, caught it because the cap rate looked wrong” is worth more to your firm than any generic prompting guide, because it trains everyone’s eye for your failure modes on your document types.
This is also the phase where deeper workflow questions surface: should lease abstraction be a repeatable pipeline instead of a chat session? Those questions are the on-ramp to the automation playbooks: the document intelligence playbook for lease stacks, and the deal analysis playbook for screening at volume. Park them in a running list rather than chasing them mid-program; the list becomes your automation backlog for day 91.
Days 76–90: habits, measurement, and closing the loop
The last two weeks convert practice into default behavior. Three moves.
Re-run the baseline. Repeat the day-1 task audit with the same people and the same task types, and compare times. This is where the program pays for itself in a way everyone can see, and the numbers you collect are yours, not a vendor’s benchmark.
Set AI-first defaults. For every task type covered in drills, the firm’s stated default becomes: start from the shared prompt, then edit. Make it explicit in a one-page “how we work” note. Defaults are what survive after the weekly sessions end.
Name a prompt librarian. Someone owns the shared library going forward: pruning stale prompts, adding new failure notes, and onboarding the next hire with it. Fifteen minutes a week. Without an owner, the library rots in a quarter.
Close the program with a working session on what stays manual. Every firm ends the 90 days with a short list of tasks where AI assistance was tried and rejected, and writing that list down is as valuable as the wins. It stops the tool-pushing, and it defines the boundary you will re-test in six months when the models improve.
Getting buy-in from skeptical senior brokers
The standard failure mode: the youngest person at the firm gets excited, demos a tool at a Monday meeting, and the two most senior producers (who control most of the revenue and most of the culture) nod politely and never touch it. Adoption at a small firm is a social problem before it is a skills problem.
What works is respect for what the skeptics are protecting. A 25-year broker’s skepticism usually decodes to one of three real objections. “My name goes on this” is a quality concern, answered by the verification discipline built into every drill, never by “the AI is usually right.” “I don’t have time to learn software” is answered by the format: 45 minutes a week doing their own live deals, not tutorials. And “this will make my skills worthless” deserves a straight answer: the drafting is what gets automated, and the judgment (pricing instinct, negotiation, relationships) is what gets more time. In the workshops we’ve run, the moment that flips a senior skeptic is rarely a demo. It is watching a peer produce a first-pass LOI on a live deal in minutes, in the room, on a deal the skeptic knows.
Two tactical rules. Make session attendance mandatory but tool use optional for the first month; forced usage breeds quiet sabotage, while watching colleagues win breeds fear of missing out. And put a principal in the drills as a participant, not a sponsor. If ownership won’t spend 45 minutes a week on this, the firm has answered its own question about priority.
What AI should not do in a CRE shop
A training program earns trust by being honest about limits, so teach the exclusion list as explicitly as the drills.
No unverified numbers leave the building. Language models invent plausible figures: comps, cap rates, square footage, clause numbers. Every number in AI-drafted output gets checked against a source document before a client sees it. This rule has no exceptions, and it is the first thing taught in week one.
No final legal language. AI can summarize a lease and flag unusual clauses for attention. It does not produce final contract language, and lease summaries used for decisions get spot-checked against the document. Treat AI as a first-pass reader, never as counsel.
No confidential data in unapproved tools. The three data rules from the setup phase apply forever, not just during training. The most common breach vector is a well-meaning employee pasting a rent roll into a free personal account, which is why the firm-wide business subscription comes first.
No relationship delegation. Drafts of sensitive communications (a tenant dispute, a busted deal, an investor apology) can start with AI, but a human rewrites in their own voice before sending. Clients can smell fully generated empathy, and in this business the relationship is the product.
Measuring whether the training worked
Keep measurement light enough that it happens. Four numbers, collected monthly, on one shared page.
Weekly active use. How many people used the primary tool on real work this week? At day 90 the answer should be everyone, unprompted. This is the single best health metric.
Task time on the audit set. The before-and-after from the baseline re-run, tracked per task type. Expect the steepest gains on drafting-heavy tasks and modest gains on judgment-heavy ones.
Library growth. Prompts added and failure notes logged per month. A growing library means people are engaged enough to contribute; a frozen library predicts relapse to old habits.
Throughput, not just speed. The strategic payoff is capacity: more deals screened per week, more listings marketed per person, faster lease turnaround. JLL’s survey found firms chasing an average of five AI use cases at once with thin results, and the small-firm counter is to go deep on the handful of workflows that gate your revenue, then count those workflows’ output. One number a brokerage can use: proposals out the door per month. One for an investment shop: OMs screened per analyst per week.
Where a facilitated workshop fits
Everything above is runnable self-serve, and this playbook is written so a motivated principal can be the facilitator. The honest case for bringing in outside help is speed and stall-prevention: an experienced facilitator has seen where week-five energy dips, knows the failure modes of CRE document types, and compresses the trial-and-error out of the drill design. Market pricing for hands-on AI team workshops generally runs $2K–15K depending on depth and team size, against which the relevant comparison is the billable time your principals would spend designing the program themselves.
For transparency about what we sell, since this playbook is the method we use: SFAI Labs runs LLM fluency workshops for small CRE teams. The scope is prompting applied to your firm’s real work: LOIs, lease summaries, market write-ups, and email handling, with your documents and your team in the room. That is the entire workshop offer. Broader needs like tool-stack audits, custom GPT builds, or written AI policies are real projects some firms take on, but they are separate work, not what the workshop covers.
If you want a read on whether your firm is ready for training, automation, or neither yet, the free AI-readiness assessment is the right entry point.
Book your free AI-readiness assessment →
After fluency: the automation roadmap
Fluency changes what your firm can see. Around day 60, teams stop asking “what can AI do?” and start asking “why are we still doing this part by hand?” That question list is your automation backlog, and each item leads to a deeper playbook:
- Lease stacks, abstraction pipelines, and structured data: the CRE document intelligence playbook
- Screening and underwriting more deals with the same headcount: the CRE deal analysis playbook
- Inbox, CRM, and listing marketing workflows: the CRE communications playbook
- Rent rolls, CAM reconciliation, and investor reporting: the CRE back-office automation playbook
- Deciding what to buy off the shelf versus build: the CRE AI buy-vs-build playbook
Sequencing matters: fluency first, automation second. A firm that automates lease abstraction before its people can evaluate an AI-generated summary has built a machine nobody can quality-check. The 90 days are what make everything after them safe.
FAQ
How long does it take a small CRE team to get fluent with AI?
About 90 days of weekly practice, at roughly 90 minutes per person per week. Individual basics come faster; a motivated broker writes useful prompts within two weeks. Firm-level fluency, where AI-first drafting is the default and everyone can verify outputs on their own document types, is a habit-formation timeline, and habits take a quarter. Six-week courses exist, but they train individuals, not firms.
What does an AI workshop for real estate teams cost?
Hands-on AI team workshops generally run $2K–15K in the current market, depending on length, depth, and team size. Self-paced online CRE AI courses are cheaper per seat but train individuals rather than changing how a firm operates. For an exact number for your team, an assessment call is more useful than a rate card, because scope drives the price.
Will AI replace brokers or property managers?
AI replaces drafting, not dealmaking. The tasks it absorbs are first-pass writing and document review: LOIs, lease summaries, market write-ups, routine email. Pricing judgment, negotiation, tenant relationships, and local market instinct stay human, and they get more hours once the drafting load drops. The realistic risk for a small firm is not replacement by AI; it is losing listings and deals to the firm across town that got fluent first.
Is confidential deal data safe in ChatGPT, Claude, or Gemini?
On business-tier plans, the major providers state that your inputs are not used for model training by default; on free personal accounts, protections are weaker. The safe pattern for a small firm: one firm-wide business subscription, a two-bucket classification rule (NDA material gets anonymized or stays out), and one named person who answers edge-case questions. Verify the data-handling terms of the specific plan at purchase time, because terms change.
Which AI tools should a 10-person CRE firm start with?
One general-purpose assistant (ChatGPT, Claude, or Gemini) on a firm-wide business plan. That is the entire day-1 stack. Microsoft Copilot is worth considering if your firm lives in Outlook and Word, though standalone chat tools teach prompting fundamentals better during the drill phase. Hold off on CRE-specific point solutions until after the 90 days, when you can evaluate them from fluency rather than from a sales demo.
How do we get skeptical senior brokers to adopt AI?
Put them in a room where a respected peer produces real work on a live deal, fast. Demos and articles do not move 25-year producers; peer proof does. Make the weekly session mandatory but tool use optional for the first month, anchor every drill to their own deals rather than tutorials, and answer the quality objection with the verification discipline, not with claims that the AI is usually right.
Do we need an AI use policy before training starts?
You need three rules, not a policy document: business-tier accounts only, classify data before pasting (anonymize NDA material), and one named owner for edge-case questions. That is enough to start training safely at a firm without an IT department. A fuller written policy can come later, informed by 90 days of real usage, and it will be a better policy for it.
What tasks should a CRE team never hand to AI?
Final legal language, unverified numbers, and sensitive relationship communications. AI output is a first draft everywhere it is used: every figure gets checked against a source document, lease language goes to counsel, and messages involving disputes or bad news get rewritten in a human voice. Teach the exclusion list as explicitly as the skills; it is what makes the rest trustworthy.
How do we measure whether AI training worked?
Four numbers on one page: weekly active users of the firm’s primary tool, task time on a fixed audit set (measured at day 1 and day 90), prompt-library contributions per month, and throughput on one revenue-gating workflow such as proposals sent or OMs screened. If weekly active use is not at or near 100% by day 90, the program stalled, and the fix is usually social (buy-in), not technical.
Should we train everyone at once or start with a pilot group?
Under about eight people, train everyone together; the shared momentum outweighs pilot efficiency. Larger than that, run a three-to-four person pilot for two weeks, then expand using their results as internal proof. Choose pilots for credibility, not enthusiasm: a busy, respected mid-tenure producer converts more colleagues than the office early adopter ever will.
Key takeaways
- The CRE AI gap is training, not tools: most firms are piloting, few hit their goals, and the difference is practiced people, not better software.
- Define fluency as testable milestones (a 15-minute first-pass LOI, a verified 30-minute lease summary), then drill toward them weekly on real work.
- Set three data rules before day 1: business-tier accounts, classify-before-pasting, one named data owner. Skip the 20-page policy.
- Run the 90 days in four phases: baseline and setup, fundamentals through drills, role-specific workflow reps, then habits and measurement.
- Adoption is social: peer proof on live deals converts senior skeptics, mandatory sessions with optional usage beat forced rollouts, and a participating principal signals that it counts.
- Fluency first, automation second. The firms that get value from lease-stack pipelines and deal-screening automation are the ones whose people can already judge AI output.
Start with the task audit this week; it takes an hour and tells you exactly what your 90 days should contain. And if you want an outside read on where your firm stands before you begin, book your free AI-readiness assessment →.
Arthur Wandzel
