Run AI as a five-day cycle: owner and market research on Monday, a verified comps summary on Tuesday, an offering memorandum draft on Wednesday, tour notes on Thursday and follow-up messages on Friday. Give it only facts you have checked, keep confidential material out of the chat, and read every figure it returns against your source.
That order matters because the risk rises as the work gets closer to a client. Research can be wrong and cost you an hour; a published memorandum with the wrong net rentable area can cost a deal. The routine below puts the checks where the risk is highest, and each day names the job, the prompt, the time it saves and the check you cannot skip.
The ground rules before Monday
Set three rules once, in writing, so nobody argues about them mid-deal.
- One account the firm controls. A business plan from ChatGPT or Claude, or Microsoft 365 Copilot if you already run Microsoft 365. Business plans do not train on your content by default; personal accounts rely on the user switching training off. If you work alone, an individual plan with training switched off is the minimum. Read the comparison of AI tools for brokers, by job if you have not chosen yet.
- A Project per live deal. ChatGPT Projects and Claude Projects both hold files and standing instructions. Put the property fact sheet in as the single source of truth, and tell the Project to answer only from it. Custom GPTs are scheduled to retire on 11 December 2026, so do not build this on one.
- The confidentiality line. Public information and your own notes may go in. Client pricing instructions, tenant financials under an NDA and principals' personal details go in only with written client agreement and with names swapped for labels.
The fact sheet is the backbone. A one-page document with the address, size, year built, tenants and lease expiries, asking terms, parking, power and zoning that you have verified yourself. Every task below leans on it, and every error you catch in a draft is traced to a number that was not on it.
Monday: owner research and the first calls
Monday's job is to decide whom to ring. AI helps you turn scattered public information into a call list with reasons, not to find owners you have not verified. The full method is in the tutorial on prospecting owners and outreach; the weekly version is lighter.
Take the ten properties you identified from your own database or a map walk. Paste in what you hold for each: address, size, last sale year if known, and the owning entity name as it appears on public records. Then use a prompt like this:
Below are 10 commercial properties with the facts I hold. For each, suggest one plausible reason the owner might consider selling or leasing in the next 12 months, based ONLY on the facts given (hold period, lease expiry, vacancy, size). If the facts do not support a reason, say "no basis" and do not guess. Then draft a three-sentence opening for a call, with no claims about the owner's finances or intentions.A sample reply, illustrative: for a 14,000 sq ft office building owned for 11 years with two of four suites vacant, "Hold period and vacancy suggest the owner may be weighing a refinance or a sale; opening: I'm calling about the building on [street], which I have been following; I've been working with occupiers looking for around 3,000 sq ft, and wondered whether either vacant suite might suit." For a unit with no lease information, "no basis".
The "no basis" line is the entry worth keeping in the prompt. Without it, models fill gaps with confident stories about owners, and you end up opening a call with something you cannot support. In the illustrative timings further down, that is about 45 minutes back on a Monday morning that otherwise goes on writing ten different call openers. Check: every reason on the list must trace to a fact you supplied.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowTuesday: a comps summary you can stand behind
Tuesday is for a one-page comps summary, built from comps you have verified. This is where a careless broker gets burned, because language models will happily produce sale prices and rents that look right and are not. The rule is that the model does arithmetic and writing; you supply every data point. There is more on reports and checks in the guide to source-checked reports and comps.
Say you are advising a landlord on asking rent for 12,000 sq ft of light industrial space. You paste in six comps from your data subscription and your own records: size, rent per square foot per year, term, concessions and date. Illustrative figures:
| Comp | Size (sq ft) | Rent ($/sq ft/yr) | Term (yrs) | Free rent (months) |
|---|---|---|---|---|
| A | 10,500 | 9.25 | 5 | 2 |
| B | 13,000 | 9.75 | 7 | 3 |
| C | 11,200 | 8.90 | 3 | 0 |
| D | 15,000 | 10.10 | 10 | 4 |
| E | 9,800 | 9.00 | 5 | 1 |
| F | 12,400 | 9.60 | 5 | 2 |
The prompt: "Using only the six comps above, calculate the average and median face rent, then the net effective rent for each comp after free rent spread over its term, and say which comps are most similar to a 12,000 sq ft, five-year lease. Show your arithmetic. Do not add any comps." The summary comes back with, for example, an average face rent of $9.43 and a median of $9.425 (the midpoint of $9.25 and $9.60), and a net effective rent for comp F of $9.60 less 2 months of free rent over 60 months: $9.60 × (58 ÷ 60) = $9.28. Illustrative arithmetic, but you can see the point: the figure that matters to your landlord is the net effective, not the face rent.
Verify every figure with a calculator or a spreadsheet. A wrong average slips past because it looks plausible, so recompute at least two numbers by hand. Then ask for two sentences of commentary you can adapt, such as "Comps with terms of five years or more show rents between $9.00 and $9.75; the three-year comp at $8.90 shows what a shorter term costs." In the illustrative timings below, the job drops from two hours to 75 minutes including the check, and the saving sits in the write-up and the net effective table.
Wednesday: the offering memorandum draft
An offering memorandum (OM) is where an invented detail does the most damage. Treat the model as a drafter working from your fact sheet, not as a researcher. Open the Project for the deal, confirm the fact sheet is current, and ask for sections one at a time rather than a full document.
Using ONLY the property fact sheet in this Project, draft the "Property overview" section of an offering memorandum in a plain, factual tone. Rules: no superlatives, no claims about future rent growth, no statements about the neighbourhood or demand. If a detail I would normally include is missing from the fact sheet, list it at the end under MISSING instead of writing around it.Here is an illustration of what the reply tends to look like, and what to fix. It writes a clean 180-word overview, then lists MISSING: ceiling clear height, loading dock count, last roof replacement. That list is the best part: it tells you what the memorandum cannot yet say. The draft may still contain one soft claim, "well-positioned for logistics users", which you remove because the fact sheet does not support it.
Do the financial summary separately, and keep the model away from the tenant schedule unless you have swapped tenant names for labels. For each section, add a final step: "List every number in this section and the fact-sheet line it came from." Anything with no source line is deleted. If a lease abstract feeds the schedule, the tutorial on abstracting a lease with page citations shows how to demand the page for every clause, which is the same discipline.
A realistic failure: a draft states "ample on-site parking" where the fact sheet said "42 spaces", and the building is 28,000 sq ft. That is 1.5 spaces per 1,000 sq ft, which many occupiers would not call ample. Replace adjectives with the number and let the reader judge. In the illustration, drafting the sections takes 1.5 hours and checking them another 45 minutes, against four hours by hand.
Thursday: tour notes that become a record
Thursday is tour day. The tool does not walk the building; you do. The gain comes afterwards, from turning your scraps into a record that you and the client can use.
In the car after the tour, dictate into your phone for three minutes: who attended, what they said about space, power, loading, parking and price, and what they asked that you could not answer. Paste the transcript into the Project with this prompt:
Turn these dictated tour notes into: (1) a five-line summary, (2) a table of the occupier's requirements and whether the unit meets each one (yes, no, unknown), (3) open questions I owe the occupier, (4) open questions I owe the landlord, (5) one recommended next step. Use only what is in the notes. Mark anything the notes do not state as "not stated".A sample result, illustrative: the occupier needed 2 dock doors and 18 ft clear height; the notes say the unit has 1 dock door and "high ceilings" so the table reads "Dock doors: no (1 of 2)", "Clear height: unknown". Those two lines are what you ring the landlord about on Friday morning, and the occupier now sees from your summary what you heard. The most common failure is the model turning "high ceilings" into a number. The phrase "not stated" in the prompt stops it, but still read the table with your own notes beside it.
If you want to record the tour itself, only do so with everyone's agreement, and read the terms of the transcription tool. For most brokers, dictating afterwards is simpler and avoids the question. If you use a meeting-notes tool, the note-taker comparison for professional firms explains what to check on data retention.
Friday: follow-up and the pipeline view
Friday turns the week into messages. Three kinds of follow-up usually pay: a note to the occupier after the tour, a short update to the landlord, and a nudge to anyone who went quiet. Draft all three from the same Thursday summary so they do not contradict one another.
Before and after, as an illustration. A rushed message: "Thanks for coming, let me know what you think." After the prompt "Draft a 90-word email to the occupier summarising what the unit meets and what is outstanding, with the two open items and a proposed day for a second call, in a warm, brief tone, no pressure", the message names the dock door shortfall, says you are checking whether a second door could be added, and proposes Tuesday. It is better because it moves the conversation forward and shows you were listening. For a longer sequence, automating buyer and vendor follow-up covers when to schedule messages and when to stay manual.
End the week with a pipeline paste. Give the Project a list of your open deals with stage, next action and last contact date, and ask: "Which three deals have gone longest without contact, and what is the one-line next step for each? Do not infer stages I did not give." This takes five minutes and finds the deal you were about to forget. Review the output against your own diary before acting.
What the week looks like in numbers
Here is the sum, using illustrative timings for an independent broker working three live instructions. Your own times will differ; log them for two weeks before you believe any of these.
| Day | Task | Before (hrs) | With AI (hrs) | Check cost (hrs) |
|---|---|---|---|---|
| Mon | Call list and openers | 1.5 | 0.5 | 0.25 |
| Tue | Comps summary | 2.0 | 0.75 | 0.5 |
| Wed | OM section drafts | 4.0 | 1.5 | 0.75 |
| Thu | Tour write-up | 1.0 | 0.25 | 0.25 |
| Fri | Follow-ups and pipeline | 1.5 | 0.5 | 0.25 |
| Total | 10.0 | 3.5 | 2.0 |
That is 10 hours down to 5.5 once checks are counted, a saving of 4.5 hours a week. If your billing or your pipeline depends on that time, it is real; if you simply fill it with more admin, it is not. The checking column is what keeps the routine honest, and it should never be what you cut when you are busy.
A requirements matrix for the occupier you represent
If you act for the tenant, Wednesday looks different. Instead of a memorandum you need a shortlist the occupier can read in two minutes. Start by writing the brief as numbers, not wishes. Illustrative brief for a small distribution business: 10,000 to 14,000 sq ft, at least 2 dock doors, 18 ft clear height, 15 parking spaces, a budget ceiling of $10.00 per sq ft per year, and a start date within four months.
Give the model that brief and the verified details of four candidate units, then ask it to build a table with one row per unit and one column per requirement, marking each cell as meets, misses or unknown, with a final column that lists what you must still confirm. Do not ask it to rank the units. Ranking requires weights that belong to the occupier, so ask the occupier to score each requirement from 1 to 5 for importance, and then multiply yourself. A unit that meets everything except the dock doors scores differently for a business that ships twice a week and one that ships forty times.
The table is useful because the unknown cells become your Thursday agenda. In a realistic run, three of four units have "unknown" for clear height because the listing sheets are silent, and a phone call to each landlord's agent answers it in ten minutes. Without the table you tour all four buildings and find out on site.
Keeping a prompt library across deals
By the end of a month you will have rewritten the same prompts several times. Save them. A one-page document with five entries is enough: call openers, comps summary, memorandum section, tour notes and follow-up email. Under each, record the prompt as it works today, one line on what it gets wrong and the check you run afterwards. Store the library where your colleagues can reach it, and date each change.
A short example entry, filled in: Tour notes. Prompt: the five-part conversion in Thursday's section. Known fault: turns "high ceilings" into a height. Fix: "not stated" instruction plus a manual look at the clear-height row. Check: read the requirements table beside my dictation. Entries like that are what make a prompt trustworthy to someone who did not write it, and they are the first thing to hand a new colleague.
One more habit is worth building: when a draft contains an error you only caught by luck, add a line to the prompt that would have prevented it, and add the check to the library. Over a quarter that turns a handful of near-misses into rules, which is a far better use of an error than a quiet fix.
Five ways the routine goes wrong, and how it shows
The invented comp. Asked for "recent comps in the area", the model supplies three. Rule: it never sources data. If a comp is in a draft and not in your records, delete it.
The stale fact sheet. The landlord signs a new tenant on Tuesday and the Wednesday memorandum still says 40% vacant. Update the sheet the day anything changes and put the date at its top.
Averaging across unlike comps. A 100,000 sq ft distribution warehouse averaged with 8,000 sq ft units gives a number that suits neither. Ask the model to group comps by size band first.
The over-friendly follow-up. A drafted email promises a landlord response by Monday that you have not agreed. Read every date and commitment aloud before sending.
Confidential leakage. A tenant's name sits in a pasted rent roll on a personal account. Fix the process, not the person: label swaps, business account, and a saved prompt that begins with a reminder.
Adapting the week for different instructions
The cycle is a pattern, not a calendar. A tenant representation week swaps Wednesday's memorandum for the requirements matrix above. An investment sale week puts more weight on Tuesday and Wednesday and on a first-pass model, for which the underwriting first-pass tutorial explains the limits.
A leasing-only week with no tours shifts the weight to Monday and Friday: more owner research, more follow-up, and a longer pipeline review. A quiet week is the time to test the saved prompts against old material, such as last quarter's tour notes, before the next busy spell exposes a weak one.
If you would like to map these jobs against your own deal flow, that is what my AI implementation consultation covers.
Questions brokers ask before adopting the routine
Can I paste a landlord's rent roll into ChatGPT?
Only if your listing agreement and the landlord's confidentiality terms allow it, and only on a business plan that does not train on your content. Safer still, remove tenant names and replace them with Tenant 1, Tenant 2. Ask the client in writing the first time.
Will AI write my comps for me?
It will summarise comps you supply, and it will do the arithmetic you ask for, but it must not source them. Models invent plausible-looking sale prices. Give it your own verified comp set from your data subscription or records and ask it to use nothing else.
How long does the routine take to set up?
A half day. Write four saved prompts, set up one Project per live deal with its facts file, and decide your confidentiality rule. After that each task takes minutes. Expect to refine the prompts over the first three deals.
Does AI belong in the tour itself?
Only after the tour. Record only if everyone present has agreed, and otherwise dictate your own notes into your phone afterwards. Then have AI turn them into a structured summary with open questions for each party.
What should never go to a chatbot?
Unpublished pricing instructions from a landlord, tenant financial statements covered by an NDA, personal details of principals, and anything a client has told you in confidence. Summarise to a neutral description before pasting, or keep that task manual.
Further reads
- AI Listing Mistakes That Can Mislead Buyers, and How to Check — How AI errors in listings mislead, and how to catch them.
- AI Virtual Staging for Property Listings: Costs and Disclosure — Costs and disclosure rules for staged listing images.
- How Estate Agents Use AI to Prepare Valuation Appointments — Preparing for a pricing meeting with AI.
- AI Automation Agency for Real Estate: Scope, Fees and Red Flags — What agencies charge to automate property workflows.
- Real Estate AI Chatbot Features: A Checklist Before You Buy — Checklist for a website enquiry chatbot.
- AI for Real Estate Brokers: Compliance File Review Step by Step — A checklist-driven way to review transaction files with AI: exception report, human sign-off and the data-handling rules that keep agent files safe.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: OpenAI ChatGPT Business pricing page, OpenAI help on Projects and the custom GPT retirement FAQ, Claude pricing page and Claude help on Projects; your own listing agreements and data-subscription terms.