A ban on ChatGPT does not stop anyone from using it. It just moves the usage somewhere you cannot see. The associate who was drafting a listing description in a browser tab on the office machine now does it on the phone in their pocket, on a personal account, with whatever deal details they happen to paste in. You have not reduced your risk. You have blinded yourself to it and forfeited any say in how it happens. The numbers bear this out at every firm size: a 2025 Salesforce survey found that 55% of employees who use AI at work do so without their employer’s knowledge, and Reco’s 2025 Shadow AI Report put unapproved AI use among knowledge workers at 71%. Blocking the tool did not lower those figures. It produced them. For a small commercial real estate firm, the alternative is not a legal project or a security program you cannot staff. It is a single page that takes an afternoon to write.
Before the specifics, one bit of orientation. Writing the policy is one move inside a larger sequence for getting a lean team genuinely fluent, which we lay out in the CRE AI training playbook, and the broader argument for why small shops can out-operate larger competitors sits in the small CRE firm manifesto. A policy is not a compliance chore bolted on at the end. It is the containment step that makes everything after it safe, which is exactly why the staged adoption framework puts it second, right after mapping where the hours go and well before anyone touches real deal data.
A Ban Is Not a Policy. It Is the Absence of One.
A policy is a set of instructions for doing something well. A ban is an instruction to not do a thing, which is a different object entirely. When an owner blocks ChatGPT, the intent is usually reasonable: confidential deal data is involved, the tools are new, and the safe-seeming move is to say no until someone figures it out. The problem is that “no” is not enforceable at a small firm, and everyone involved knows it.
There is no device-management software watching the associate’s iPhone. There is no network filter on the coffee-shop Wi-Fi where half the LOIs get drafted. The ban is a request, and the people receiving it have already discovered that the tool makes their afternoon shorter. Reco’s finding that 71% of knowledge workers use AI without approval is not a story about disobedient employees. It is a story about a productivity gain too large to refuse and a rule too vague to follow. A ban gives your team a binary they will resolve in favor of getting their work done.
MIT’s 2025 study of AI in business, “The GenAI Divide,” found that more than 40% of workers use AI tools personally even where their employer has deployed nothing. Your firm’s real choice is not whether AI enters your workflows. It already has. The choice is whether it enters through a door you built or through a window you pretended was locked.
What a Ban Actually Buys You
The case against the ban is not philosophical. It is that a ban actively increases the exact risk it was meant to prevent.
Consider what changes the moment usage goes underground. On a sanctioned business-tier account, the firm controls the settings that keep inputs out of model training, and the owner can point to a rule about what data is fair game. On a personal free account reached from a phone, none of that exists. The data-privacy protections that come with business tiers are off, and there is no rule at all, because the official position is that this is not happening. IBM’s 2025 Cost of a Data Breach Report quantified the gap: incidents involving shadow AI, meaning unsanctioned tools employees adopt on their own, added an average of $670,000 to breach costs and appeared in one in five breaches studied. The same report found that 63% of organizations had no AI governance policy in place. Those two findings are the same finding. The absence of a policy is the risk.
For a real estate firm the exposure is concrete. A broker pastes a draft LOI with the buyer’s name and price into a personal chatbot to tidy the language. A rent roll goes in to get a quick summary. An unexecuted lease term sheet gets rewritten for a tenant. None of these are exotic misuses. They are ordinary Tuesday tasks, and under a ban every one of them happens on the least protected path available, with no record and no rule. The ban did not prevent the behavior. It stripped away the one layer of protection the firm could have provided.
The Five Decisions That Fit on One Page
A policy that works at a 12-person brokerage is not a legal document. It is five decisions written plainly enough that a new hire can read it in two minutes and know exactly what to do. You do not need a template mill or a security consultant to make these calls. You need an afternoon and a willingness to be specific.
Here are the five. Each is a single decision with a single output.
Decision 1: Name the Approved Tool and Its Tier
Pick one primary tool and name it. ChatGPT, Claude, Gemini, and Microsoft Copilot all handle the core text work a broker does, and for a firm that already lives in Outlook and Office, Microsoft Copilot is often the path of least resistance because it sits inside tools people already open. The specific choice matters less than making one, because a named default is what turns “AI” from an abstraction into a place people go.
The tier is the part firms skip, and it is the part that protects them. Specify a business or team tier, not personal free accounts. The paid business tiers of the major assistants offer administrative settings that keep your prompts out of model training by default, which is the baseline any firm handling confidential deal data needs. The choice between platforms is worth its own analysis, and we compare the two most common options for a brokerage in our look at ChatGPT Team versus Claude for a CRE office. For the policy, one line is enough: the firm’s approved tool is X, business tier, and personal accounts are not to be used for firm work.
Decision 2: Draw the Data Line for CRE
This is the decision the generic templates never make for you, and it is the entire value of the page. “Do not share confidential information” is not a rule; it is a feeling. A broker cannot act on it, so they either freeze and paste nothing useful or shrug and paste everything. You have to draw the line in terms of the documents that actually cross a CRE desk.
The safe side is public and hypothetical: a listing that is already on the market, a marketing description, a general market question, a first draft written from invented numbers. The unsafe side is anything that identifies a party tied to a live financial position: client or investor names attached to figures, unexecuted LOI and lease terms, rent rolls and tenant financials, purchase prices before they are public, and anything under an NDA. State it as two short lists. A broker reading them should never have to guess which side a given task falls on. When something genuinely sits in the middle, the rule is to strip the identifying details first and paste the sanitized version, which is a skill worth practicing on real work rather than a paragraph to memorize.
Decision 3: Set the Disclosure Rule
Decide when AI involvement has to be disclosed and to whom. This protects the firm’s credibility, which for a brokerage is the whole asset. The workable standard is that AI can draft, but a human owns every word that leaves the firm, and material client-facing analysis notes its tools where a reasonable client would want to know. You are not asking anyone to footnote a cleaned-up email. You are drawing a line at the point where an undisclosed machine-written analysis could embarrass the firm or mislead a client. One sentence in the policy handles it: AI may assist internal drafting; final client deliverables are reviewed and owned by a named person.
Decision 4: Set the Human-Review Rule
Language models produce fluent text that is sometimes wrong, and in real estate a confident wrong number has consequences. The review rule closes that gap. Nothing an AI produces that touches a number, a legal term, a date, or a factual claim goes out without a human checking it against the source. This is not a knock on the tools; it is the same discipline you already apply to a junior associate’s first draft. Write it as a hard rule rather than a hope, because the failure mode is not malice but speed: the output reads so well that no one thinks to verify it. The rule makes verification the default instead of the exception.
Decision 5: Name the Owner
A policy with no owner is a suggestion. Name one person, almost certainly the owner or a principal at a firm this size, who owns the document, answers questions about it, and updates it when the tools change. This is not a large job. It is an hour a quarter to confirm the approved tool’s data settings still hold and to fold in whatever the team has learned. What it prevents is the slow rot where the policy written in August is quietly irrelevant by spring because a vendor changed a default and nobody noticed. The owner is the difference between a living rule and a PDF nobody has opened since onboarding.
How to Enforce a Policy Without an IT Department
Here is where most advice falls apart for a small firm. Enterprise guidance assumes you can enforce a policy with technology: device management, network filtering, data-loss prevention. You have none of that, and buying it would cost more than the risk it addresses. So the honest question is how you enforce a rule you cannot technically enforce.
The answer is to make the sanctioned path the easy path. People route around bans because the ban makes the safe option harder than the unsafe one. Flip that. Provide the business-tier account, set it up on everyone’s actual devices, put the two-minute policy where people see it, and make the approved tool genuinely faster to reach than a personal login. When the compliant path is also the convenient path, compliance stops requiring enforcement. Deloitte’s 2026 Commercial Real Estate Outlook, surveying more than 850 executives, still found roughly a quarter of firms blocked by change resistance and expertise gaps even with real budgets; a small firm closes that gap not with a bigger budget but by removing the friction that pushes people to the shadow path in the first place.
The second lever is culture, and at a firm this size the owner sets it in a sentence. If the message is “use the approved tool, paste nothing from the unsafe list, and ask me when you are unsure,” and the owner visibly uses the tool the same way, the policy holds because it is normal, not because it is policed. The reason bans generate shadow AI is that they make honest use impossible; a clear policy makes honest use the obvious choice.
What the Page Actually Unlocks
The reframe that matters most is that a policy is not a brake. It is the thing that lets you press the accelerator without flinching. The reason so many small firms sit at the wrong end of the adoption gap, where NAR’s 2025 survey found 68% of Realtors using AI but only 17% reporting a significant positive impact, is not that the tools are weak. It is that fear caps how far anyone will push. People who are unsure what is allowed use the tool timidly, on toy tasks, and get toy results.
Remove the ambiguity and the ceiling lifts. Once a broker knows exactly which data is fair game and which is not, they stop self-censoring on the safe eighty percent of their work and start getting real value from it. The policy is what converts nervous, occasional, underground use into confident daily use on the tasks that actually move the firm. That is why the sequence matters: the containment the policy provides is what makes the training that follows worth paying for. A team that has been told “no” cannot practice. A team that has been given a clear “here is how” can. The one-page policy is the smallest document with the largest return in the entire adoption effort, and it is the one almost every firm skips on the way to a ban that quietly makes things worse.
Frequently Asked Questions
Is banning ChatGPT ever the right call for a real estate firm?
Rarely, and almost never as a lasting position. A short, explicit pause can make sense while you write the policy, but a standing ban does not stop use; it relocates it to personal devices where you have no visibility and no data protection. A 2025 Salesforce survey found 55% of employees using AI at work do so without their employer knowing, typically by switching to personal accounts. If the concern is confidential data, a ban makes it worse, because underground use runs on free personal accounts with none of the privacy settings a business tier provides. The right call is a clear policy that channels the behavior, not a prohibition that hides it.
What should a small CRE firm’s AI policy actually contain?
Five decisions, each one line. The approved tool and its tier, so people know where to go and on what kind of account. The data line, drawn specifically for CRE, separating public and hypothetical material from anything tied to a named party’s live financials. The disclosure rule for client-facing work. The human-review rule for anything touching a number, date, or legal term. And a named owner who keeps the page current. That is the whole document, and it fits comfortably on one page a new hire can absorb in two minutes.
What client data is safe to put into ChatGPT?
Public and hypothetical data is generally safe: listings already on the market, marketing copy, general market questions, and drafts built from invented numbers. Unsafe is anything identifying a party tied to a live financial position, including client and investor names attached to figures, unexecuted LOI and lease terms, rent rolls, tenant financials, and pre-public purchase prices. For anything in between, strip the identifying details and paste only the sanitized version. Drawing this line explicitly is the single most valuable part of the policy, because a broker can act on it where a vague “keep it confidential” only produces guesswork.
Do the business tiers really keep our data private?
The paid business and team tiers of the major assistants offer administrative settings that keep your prompts out of model training by default, which the free personal tiers generally do not. That is the baseline protection a firm handling confidential deal data needs, and it is a concrete reason to specify the tier in your policy rather than leaving people on personal logins. These terms do change, so the policy owner should confirm the current settings for your approved tool at least once a quarter rather than assuming a default set months ago still holds.
How do we enforce a policy with no IT department?
You make the sanctioned path the easy path instead of trying to police the unsanctioned one. Provide the business-tier account, set it up on the devices people actually use, and make the approved tool faster to reach than a personal login. People route around bans because bans make the safe option the harder option; when the compliant path is also the convenient one, compliance stops needing enforcement. The owner reinforces it by using the tool the same way and answering questions when they come up. Culture and convenience do the work that device-management software does at a larger firm.
Who should own the AI policy at a 10-person firm?
The owner or a principal, almost always. At this size there is no committee and no compliance function, which is an advantage: the person who can change the rule is the same person who set it. The job is small, roughly an hour a quarter to confirm the approved tool’s data settings still hold and to incorporate what the team has learned. What the named owner prevents is the drift where a policy quietly goes stale because a vendor changed a default and no one was watching for it.
How long should the policy be?
One page, and shorter is better. The failure mode of AI policies is not that they are too brief; it is that they are long, generic, and never read. A two-minute document that a broker can hold in their head beats a ten-page acceptable-use agreement that lives unopened in a shared drive. If a rule cannot be stated in a sentence someone will remember at the moment they are about to paste something, it will not change behavior. Write for the moment of decision, not for a binder.
What does it cost to get this right?
The policy itself costs an afternoon, not a budget line. The business-tier accounts are a modest per-seat subscription. If you want help drawing the data line and training the team to use the approved tool well, structured LLM-fluency training for a small team generally runs in the low thousands to low tens of thousands of dollars as a market range, far less than a single mishandled deal. The expensive path is the accidental one: an underground data exposure that scares a firm off AI for a year while competitors pull ahead.
Where to Start
The first move is not writing the document. It is finding out what your team is already doing so the policy fits reality instead of a fantasy of total control. That is exactly what a free AI-readiness assessment produces: a working session that surfaces which tools your people already reach for, which confidential-data situations actually come up in your deals, and where the one-page policy needs to draw its lines to be both safe and usable. You will leave with the data line drawn for your firm’s real workflows and a clear read on whether you need a full training engagement, a lighter course, or simply a sanctioned account and a page of clear rules. Book a free AI-readiness assessment and replace the ban you cannot enforce with the page that actually works.
Arthur Wandzel
