Lesson objective

Keep a transaction AI use record stage by stage as the work happens, so a broker can see what the tool did, what was checked and which licensee approved it.

Plain-language explanation

Why this matters now

A month after closing, a buyer says the listing promised a finished basement that turned out to be unpermitted storage. Your broker asks three questions: who wrote those remarks, what were they based on, and who checked them? If AI helped, a fourth: what did the tool produce, and what did a person change?

Without a record made at the time, nobody can answer. With one, the answers take two minutes.

Professional guidance (California): the California Department of Real Estate's Winter 2026 Real Estate Bulletin lists, among its best practices for AI, written broker policies and training on AI use, human review of outputs before publication, and keeping records of how AI outputs were reviewed. DRE presents these as recommendations; the binding duties are the statutes and regulations the article cites, including a broker's duty to reasonably supervise, which DRE says "extends to the tools used to conduct licensed or unlicensed activities, including AI-powered software." This is California guidance; check your own commission.

Law / rule (Texas): 22 Tex. Admin. Code §535.2 requires a Texas broker to keep, for at least four years from closing, termination or the end of the transaction, records including disclosures, commission agreements, substantive communications with parties, offers and contracts, and appraisals, broker price opinions and comparative market analyses. It does not mention AI or an AI use record. But when AI helped produce a CMA or a substantive communication, the record of how that document was made is the natural companion to a document the broker must already keep.

A question to sit with

If your broker opened one of your files tomorrow, could they tell which sentences a tool wrote, which facts you checked against what, and when you approved the result?

What goes wrong

Writing it up later from memory. A Friday-afternoon summary records what the agent thinks happened, not the prompt that was run or the version that was sent.

Recording the tool but not the review. "Used AI for remarks" says nothing a broker can rely on. The review is what matters.

Copying client data into the record. A record that pastes the buyer's pre-approval letter or the seller's disclosure creates another copy of confidential information in another place.

Keeping the record apart from the file. A personal notes app or chat history is not where a broker, an auditor or a commission investigator will look, and it may vanish when the agent leaves.

Recording only the successes. A draft that was rejected, or a fair housing flag that was caught, is exactly the evidence that supervision works.

The practical answer

Write the record stage by stage, as each step of the Lesson 2 workflow finishes. The record format is an AI Coalition Network recommendation; Module 14 develops the full record, incident logging and retention in depth.

Field Written at Example (fictional)
Transaction or listing reference Stage 1 Fictional listing TC-40517
Task and triage line Stage 1 Listing remarks; Route B; sign-off J. Ortiz under broker M. Lane
Tool and prompt version Stage 4 Approved assistant; "MLS remarks v3"
Sources supplied Stage 2 Verified property file dated 2026-09-10; HOA statement
Data removed Stage 3 Sellers' reason for moving
Verification counts Stage 5 14 statements: 11 supported, 1 changed, 1 unsupported, 1 fair housing flag
Fair housing review result Stage 6 Flag rewritten to describe floor plan
Reviewer, decision, date Stage 6 J. Ortiz, approved with changes, 2026-09-11
Where the released version is Stage 7 Transaction file, "remarks-final"

Point to documents; do not copy them. The record names the source and where it is filed. It does not paste confidential content into itself.

Keep it with the transaction file. Filing the record where the broker keeps the transaction's other records means it follows the same retention and access rules. In Texas, for example, the CMA it describes must be kept for at least four years; keeping the AI record alongside it is best practice, not a stated Texas requirement.

Record the rejections too. A rejected draft, a fair housing flag rewritten, or a task moved to Route D shows supervision in action. Law / rule (South Carolina): S.C. Code §40-57-820 makes a South Carolina licensee responsible for AI-assisted work product; a contemporaneous record is how a licensee shows what they did to meet that responsibility.

Remember this

  • Write the record as each stage finishes, not afterwards from memory.
  • Record the review, not just the tool: sources, counts, fair housing result, reviewer and date.
  • Point to documents; never paste confidential client material into the record.
  • Keep the record with the transaction file so it follows the broker's retention and access rules.
  • California's DRE recommends keeping records of how AI outputs were reviewed; Texas requires brokers to keep CMAs and substantive communications for at least four years.

Real estate example

Copperline Properties is a fictional California brokerage whose written AI policy, set by responsible broker Irene Castellanos, requires a transaction AI use record for every AI-assisted client or public deliverable. Buyer's agent Priya Raman is preparing an offer summary for her fictional clients, Sam and Lena Duarte, comparing two counteroffers on a fictional townhouse. She opens the record first and writes the triage line: client-facing, confidential terms, fully verifiable against the two counteroffers, no fair housing exposure, sign-off by Priya under Irene, Route B. At Stage 2 she lists the two counteroffer documents by filename. At Stage 3 she notes that she removed the Duartes' loan details, which the summary does not need. At Stage 4 she records the prompt name, "offer comparison v2". At Stage 5 her grid has 10 statements: 8 supported, 1 changed (the draft gave the second counteroffer's option period as 10 days; the document says 7) and 1 beyond the licensee (a sentence calling one clause "legally safer", which she deletes and replaces with a note suggesting the clients ask an attorney about that clause). At Stage 6 she records her own approval with the date. At Stage 7 she notes where the sent version is filed. When Irene samples files at month-end, she reads Priya's record in under three minutes and sees exactly what the tool did and what Priya changed. She notes one improvement for the team: the record should also say which version of the summary the Duartes received, because Priya revised it once after a phone call.

Responsible-use warning

Do not reconstruct the record afterwards. An entry written from memory on Friday is not evidence of what happened on Tuesday, and it rarely captures the prompt version, the sources or the draft that was actually released. Do not paste client documents into the record, and do not keep it in a personal app the broker cannot reach. A record that shows only the tool, and not the review, proves nothing a broker or regulator needs to know.

Human-review requirement

The licensee who approved the deliverable completes and signs the review fields, and the supervising broker, or a delegated supervisor, samples records under the brokerage's written policy. The broker checks that entries were made during the work, that they point to documents rather than copy them, that verification counts and fair housing results are present, that the reviewer is a licensee, and that the record sits with the transaction file. Gaps are raised with the licensee within the brokerage's response timeframe and fixed in the process, not in the old record.

Practical exercise

Use a fictional transaction only.

  1. Create a record template with the nine fields in this lesson's table, and a column showing the workflow stage at which each field is written.
  2. Choose a fictional task, such as listing remarks or an offer summary, and write the Stage 1 triage line, naming the licensee and supervising broker.
  3. List two or three fictional source documents by name, date and filing location.
  4. Record one item of data you removed and why the task did not need it.
  5. Write a prompt name and version, and the verification counts from a fictional grid. Check that the counts add up to the total number of statements.
  6. Record one fair housing flag and how it was resolved, or state why the task had no fair housing exposure.
  7. Record the reviewing licensee, the decision and the date.
  8. Review your record: remove anything that copies confidential content rather than pointing to it.
  9. Swap records with a colleague and answer, from their record alone, who wrote the output, what it was based on, what was changed, and who approved it.
  10. Note any question you could not answer and add the missing field to your template.
  11. Write one sentence stating where your brokerage keeps transaction records and how long, and check it against your own state's rules and your broker's policy.

Knowledge check

  1. What does the California DRE bulletin recommend about records of AI use?

    Answer guidance: Among its best practices, DRE recommends keeping records of how AI outputs were reviewed, alongside written broker policies, training and human review before publication. These are recommendations for California licensees.

  2. Which Texas records must a broker keep for at least four years, and does the rule mention AI?

    Answer guidance: Records including disclosures, commission agreements, substantive communications, offers and contracts, and appraisals, BPOs and CMAs. The rule does not mention AI; keeping the AI record with them is best practice.

  3. Why should the record point to documents instead of copying them?

    Answer guidance: Copying confidential client material into the record creates another copy in another place. Naming the source and its filing location gives the broker what they need without spreading the data.

  4. Why record rejected drafts and caught fair housing flags?

    Answer guidance: They show that review actually happened and that supervision works. A record of successes alone cannot show what the licensee checked or changed before release.

Sources and further reading