Broker tools
Five jobs AI can take off a leasing broker’s desk
Most of what is written about AI in commercial real estate is about the industry. This is about your Tuesday.
The useful question is not whether AI changes brokerage. It is which forty minutes of today it can give back. Below are five places it reliably does, with instructions you can copy and adjust, what to check each time, and where each one stops. Then three jobs to keep away from it.
None of this needs a pilot project or a committee. It needs a browser and the discipline to check the output. One rule runs under all five: hand over a job only where checking the answer is quicker than doing the work yourself.
Before any of it: what you may paste
Whatever you paste goes to someone else’s server, and consumer tools may keep it and train on it [1, 2]. A client’s rent roll, a tenant’s financials, an unexecuted offer: those are not yours to hand over. Check your brokerage’s policy first: BC’s regulator tells brokerages to set clear policies on AI tools and on how client information is handled [2].
Where you are unsure, redact. Replace tenant names with “Tenant A”, drop the civic address, keep the structure and the arithmetic. Every instruction below works just as well on an anonymised version. Personal information in your contact records carries obligations under federal privacy law, and under provincial law in Quebec, British Columbia and Alberta, which generally applies instead of the federal law inside the province [3], and those obligations do not stop at the door of a chatbot. Quebec’s Law 25 carries the heaviest penalties, up to C$25 million or 4% of worldwide turnover [4], and is the one most likely to reach a bilingual book.
The shape of an instruction that works
Most disappointing results come from asking for a thing rather than describing a job. Every instruction in this post uses the same five parts, and you can build your own from the same skeleton:
Job: [what you want done, in one line] Input: [the document, notes or table below] Output: [the exact shape you want back] Rules: - Use only what is in the input. Do not add facts. - Where the input does not say, write "not stated". - Quote figures and dates exactly as written. - Where something is ambiguous, flag it instead of choosing.
The rules are the part people leave out, and they are the part that does the work. “Do not add facts” is what stops a tidy answer with invented numbers in it. Keep your own copy of this skeleton in a note and start from it every time.
1. Abstracting a lease or an offer
Forty-three pages, and a client who wants to know about renewal, restoration and assignment before a call at four.
Job: abstract this commercial lease into a term table. Input: the lease text below. Output: a two-column table with one row for each of: commencement, expiry, rentable area, net rent by period, additional rent treatment, renewal option including the notice window and how the rent is set, assignment and sublet consent standard, restoration and make-good, parking, use restriction or exclusivity. Rules: - Against each row, cite the section number you took it from. - Where the lease does not say, write "not stated". Do not infer. - Quote dates, rates and areas exactly as written. - Where two clauses disagree, list both and say so.
Adjust it. Swap the row list for whatever this deal turns on. On an offer or a letter of intent, add a row for the conditions and their deadlines, and one for who prepares the lease.
Check. Read every clause it cites. The citation rule is there to make that quick and to suppress invention, because a model that cannot cite is guessing. Defined terms are where it slips, and rentable area is often defined in a schedule that contradicts the body.
Where it stops. It will not tell you that the restoration clause and the landlord’s work letter contradict each other in practice. The last rule catches only the obvious collisions.
2. Cleaning up the list you already keep
Your expiry list came from three sources. Dates are in four formats, sizes read “10,000 sf” and “10000” and “10.0k”, and half the confirmed column is blank.
Job: normalise this lease list. Input: the table below, pasted from a spreadsheet. Output: columns tenant, building, unit, rentable_sf, expiry, confirmed. Rules: - rentable_sf as a whole number, no commas, no "sf". - expiry as YYYY-MM-DD. - Where a date could be day-first or month-first, write AMBIGUOUS and repeat the original text beside it. - Where only a month and year are given, use the last day of that month and set confirmed to no. - Where only a year is given, use 31 December of that year and set confirmed to no. - Never guess a missing value. Leave the cell empty. - Return the table and nothing else.
Adjust it. Add the columns you actually keep. If your dates are reliably day-first, say so in the rules and drop the AMBIGUOUS line.
Check. Every row it flagged, then sort by expiry and look at both ends of the list. Outliers are where a misread date lands.
Where it stops. For a list you clean every week this is slower than a formula. Ask it once to write the formula, then reuse that. Or paste the cleaned list into the free lease expiry check to get the twelve-month call calendar. It reads a half-known date exactly the way the instruction above does, and the sweep behind it sets out which call to make when.
3. Turning tour notes into something you would send
Four tours, notes on your phone, and a client expecting a summary tonight.
Job: turn my tour notes into a client-facing summary. Input: the notes below. Output: one paragraph per space, in the order given, each covering size, condition, what works, what does not, and the one open question. Rules: - Use only the facts in my notes. - Do not add specifications, dimensions or rents. - Do not rank the spaces and do not recommend one. - Where I did not record something, write "to confirm". - Plain sentences, no marketing adjectives.
Adjust it. The last rule is the one to change. Drop it for brochure copy; keep it for anything a client will forward to their finance director.
Check. The specifications. Left to itself it will supply a ceiling height nobody measured, and it will sound certain about it.
Where it stops. The recommendation. Write the closing paragraph yourself, because that judgment is the part the client is paying for.
4. Turning a call into a record and a follow-up
Twenty minutes with a landlord, and you are already driving to the next appointment.
Job: turn this call into a file note and a follow-up email. Input: the transcript below. Output: two parts. Part one, a file note: who was on the call, the date, what each party said they would do, and every date, rate or figure mentioned. Part two, a follow-up email from me confirming only what was actually agreed. Rules: - Quote figures and dates exactly. Do not round anything. - Include nothing that was not said. - Do not propose next steps that were not discussed. - Keep the email under 150 words and skip the pleasantries.
Adjust it. For a note that stays internal, drop part two. For a call with three people on it, add a rule to attribute every commitment by name.
Check. The figures, every time. Transcripts mangle numbers and names, and a moving car is the worst place to record one.
Before you record anything. The rules on recording a conversation differ depending on who is on the call and where they are. The Criminal Code allows a recording with one party’s consent [5], but a business covered by federal privacy law must tell the caller it is recording, say why, and ask for consent [6]. Your brokerage’s policy and the other party’s expectations matter more than the legal minimum. Ask first.
Where it stops. Tone. It will not know that this landlord’s “we’ll see” means no.
5. Producing the other language
A Quebec listing, or a bilingual client, and your written French is functional rather than marketing-grade.
Job: translate this listing into Quebec French for a commercial real estate audience. Input: the English text below. Output: the French text, same structure and headings. Rules: - Keep these terms consistent: net rent = loyer net; additional rent = loyer additionnel; rentable area = superficie locative; improvement allowance = allocation pour améliorations locatives; sublease = sous-location; net effective rent = loyer net effectif. - Do not translate building names or civic addresses. - Keep every figure, unit and date in the format given. - Where the English is ambiguous, flag the sentence instead of choosing for me.
Adjust it. Keep your own glossary line and reuse it on every listing. Consistency from one listing to the next matters more than any single word choice, and a model will not remember what it chose last week.
Check. Have somebody who actually works in French read it before it goes out. In Quebec this is not only a quality question, because French-language obligations apply to commercial advertising and to contracts [7].
Where it stops. It is a first draft. It is never a certified translation.
Three jobs to keep away from it
- Interpreting a clause you will be held to. It will give you a confident reading of an assignment clause that is wrong in a way you will not notice. Use it to find the clause, not to tell you what it means.
- Market data from memory. Ask what industrial rents are doing in your node and it will answer with a straight face and no idea. Give it your comps; never take comps from it, and never let it fill a gap in a rent roll.
- Anything where the accountability does not transfer. You cannot tell a client, a landlord or your regulator that the model said so [2]. It all goes out under your name and your licence.
The test
Not whether the output sounds impressive. Whether it saved time on work you were going to check anyway. All five above pass, because the checking is quick and the failure modes are visible. Anywhere verifying the answer takes as long as doing the job, do the job.
The same idea runs through the arithmetic: put the numbers in one shape and the comparison becomes obvious. That is all the net effective rent calculator does, and it needs no model to do it. The same goes for the bill: what CRE software actually costs when you are the only broker is arithmetic, not a demo.
Questions
Is it safe to paste a client’s lease into an AI tool?
Not by default. Consumer tools may retain what you send and use it for training. Check your brokerage’s policy, use whatever tool it has approved, and redact names and addresses where you can. An anonymised lease abstracts just as well.
Which tool should I use?
The one your brokerage allows. The instructions here are deliberately plain and work in any of the mainstream assistants, because the result comes from how the job is described rather than from the brand.
Will it get the numbers right?
It copies numbers more reliably than it reasons about them, and the copy is only ever as good as the source: a transcript recorded in a car has the figure wrong before the model sees it. Ask it to quote figures exactly and to show where each one came from, then do the arithmetic yourself or in a calculator built for it.
Can I use it to write listing copy?
Yes, as a first draft from facts you supply. The risk is not bad prose, it is invented specifications. Tell it to add nothing you did not give it, and read the result against the actual space.
Do I have to tell a client I used AI?
Treat it as your own work either way, because the accountability sits with you and your licence. Brokerage policy and regulator guidance are both moving on this: BC’s regulator published an AI guideline in February 2024. Check yours rather than assume.
Disclosure, and the same questions asked of us. We build Frontage, a leasing desk for independent Canadian brokers. Its paid tier includes an AI engine that drafts, summarises and answers questions across your own book, and a phone assistant that answers your line; the free tier has no AI at all. The rules at the top of this post apply to us too, so here are our answers. What you put in Frontage is not used to train AI models. The drafting, the answers and call transcription run on Microsoft Azure OpenAI; what Frontage stores stays in Canada, but those requests may be processed in any region where Microsoft runs the model [8]. The phone assistant’s reasoning between calls is processed in Canada. The phone assistant's live conversation with a caller runs on OpenAI in the United States, and the assistant tells every caller, before they speak, that it is an AI and that the call is recorded. The privacy notice names those processors and sets out how long we keep your records. How long a provider keeps a prompt or a transcript is a separate question, and one worth putting to any vendor, ourselves included.
Sources
Checked in September 2026.
- OpenAI Help Center, How your data is used to improve model performance, read Sep 2026.
- BC Financial Services Authority, Artificial intelligence guideline, Feb 22, 2024.
- Office of the Privacy Commissioner of Canada, Provincial laws that may apply instead of PIPEDA, read Sep 2026.
- Osler, Law 25: a new enforcement scheme for the protection of personal information in Quebec, Jan 16, 2024.
- Criminal Code of Canada, section 184, current.
- Office of the Privacy Commissioner of Canada, Recording of customer telephone calls, updated Mar 6, 2018.
- Office québécois de la langue française, Langue du commerce et des affaires, updated Jun 3, 2026.
- Microsoft Learn, Data, privacy, and security for Azure OpenAI models, May 18, 2026.
Written for independent leasing brokers in Canada. Nothing here is legal advice, and the notes on privacy, call recording and French-language obligations are general; check your brokerage’s policy and, where it matters, counsel.