AI open-house follow-up test desk with synthetic visitor notes, review checklist, and two declining repair score cards

Can AI Turn Messy Open-House Notes Into Follow-Up? We Tested Claude and Gemini

by

TPTK Workflow Lab | TWF-002 | Tested July 27, 2026

Yes, AI can turn open-house notes into a useful first draft. No, it should not send the emails or write directly to your CRM.

We gave Claude and Gemini the same 12 fictional visitor notes. Both organized the leads and drafted follow-up, but neither produced a kit we would approve without human review. The most important finding was simpler than the scoring: a confident-looking output can still invent a status, misread a priority, or imply that work is already underway.

Best use: let AI do the first organization pass, then verify five things before any import or send: unknown fields, contact holds, priority reasons, property facts, and promises of future action.

Good use of AI

Organize and draft

Turn shorthand notes into a review table, a draft follow-up plan, and draft emails. Keep every output in a staging area.

Use the workflow

Keep a human in charge

Verify before action

A person should approve contact permission, representation holds, priorities, property answers, CRM fields, and every message.

See the checks

Do not automate yet

Sending and CRM writes

Neither completed run passed our review gate. Direct sending or importing would turn small model mistakes into real operational errors.

See what failed

A practical AI follow-up workflow you can copy

This is the version we would actually recommend to an agent or team. It treats AI as a drafting assistant and makes the human review gate explicit.

  1. Clean the input. Remove real personal data while testing. Give each fictional or approved record a stable ID so omissions and duplicates are easy to spot.
  2. Mark hard boundaries first. Identify do-not-contact, consent, representation, privacy, and fair-housing holds before asking for drafts.
  3. Ask for a review table, not a “CRM-ready” import. Require the model to preserve unknowns and cite the note that supports each priority or next step.
  4. Generate drafts only. Every email, text, appointment, property answer, and CRM change remains unexecuted until a person approves it.
  5. Review against the original notes. Do not ask the model whether its own answer is correct and accept the response. Check the replacement against the source.
  6. Move only approved records. Import or stage messages after the human reviewer has cleared the record, not before.

Copy/paste starter prompt

Use fictional or appropriately de-identified notes while you build the process. Replace the bracketed section with your safe test packet.

You are helping prepare open-house follow-up for HUMAN REVIEW ONLY.

Do not send messages, update a CRM, schedule an appointment, or claim
that a person or team has taken an action.

Rules:
- Preserve missing information as UNKNOWN.
- Do not infer financing, motivation, representation, consent, urgency,
  protected traits, neighborhood fit, or property facts.
- If contact permission or representation is unclear, mark HOLD.
- Give every priority a short reason copied from the source notes.
- Label all messages DRAFT - NOT SENT.
- Put questions that require human verification in a separate list.

Return:
1. A review table with the original record ID.
2. A hold list.
3. Proposed priorities with source-backed reasons.
4. Draft follow-up only for records that are not on hold.
5. A final human-verification checklist.

SAFE TEST NOTES:
[PASTE FICTIONAL OR APPROPRIATELY DE-IDENTIFIED NOTES HERE]

The five checks that caught the important mistakes

1. Unknowns stayed unknownMissing financing, timing, or property information was not converted into a yes, no, or negative conclusion.
2. Holds stayed attachedConsent limits, another-agent representation, and do-not-contact conditions followed the record into every output.
3. Priorities had evidenceA P1 or P2 label could be traced to an explicit request or timing signal in the original note.
4. Facts were not promisesThe draft did not say “we are verifying,” “we will follow up,” or “the seller is considering” unless the source established it.
5. Nothing happened automaticallyEvery email, CRM update, appointment, and owner report was still a proposed action awaiting approval.

These checks are deliberately boring. That is the point. They catch operational mistakes that polished prose can hide.

What happened in our Claude and Gemini test

Both completed runs handled the basic shape of the task: they returned all 12 fictional lead IDs, organized information, and withheld drafts for the three required hold records. That made the outputs look more usable than they actually were.

Claude run

Better first pass, still not ready

66.8/100 first pass

58.4/100 after self-repair

Claude preserved the three holds and avoided the most serious unsupported-action failure. It still mishandled financing unknowns, one priority, required formatting, and personalization. Its repair changed a correct priority to an incorrect one.

Gemini run

Complete-looking output, riskier language

45.6/100 first pass

34.8/100 after self-repair

Gemini returned all 12 rows and respected the three no-draft holds. It also wrote unsupported promises and described follow-up work as if it were already underway. Its self-audit said those problems were gone when they were still present.

Our decision: Claude was the safer of these two runs, but there was no winner we would connect directly to outreach or CRM writes. The result is about this one frozen test, not every future version of either product.

Why “audit and fix your answer” made both outputs worse

A self-review prompt sounds reassuring, but the model is still judging its work with the same incomplete understanding that produced the first answer. In this test, Claude repaired the wrong priority. Gemini claimed it had removed unsupported action language while retaining it.

Self-critique is useful for generating a second draft. It is not an acceptance test.

If the output affects contact permission, fair housing, a property representation, CRM data, or a promise to a lead, compare it with the original source and a fixed checklist. Confidence is not verification.

How we tested it

One fictional open-house note packet sent through Claude and Gemini, followed by one controlled repair and human scoring
One frozen fictional packet, the same requested outputs, one allowed repair, and a separate human review against the original notes.
  1. We froze 12 fictional visitor records. They included unknown financing, property questions, representation and consent holds, follow-up requests, and priority evidence.
  2. Claude and Gemini received the same task. Each was asked for a review table, follow-up plan, allowed email drafts, and a summary.
  3. We scored the first outputs before repair. The rubric checked factual fidelity, omissions, unsupported claims, priorities, privacy and fair-housing boundaries, formatting, and human edit burden.
  4. Each model received one identical self-audit request. It had to list defects and replace the complete output once.
  5. We scored the replacements against the original packet. We did not use the model’s audit log as proof that a defect was fixed.

No real leads were used. No email was sent, no CRM was changed, and no appointment or owner report was created. Timing and interface screenshots were not captured, so this is not a speed or usability comparison. ChatGPT did not complete a comparable run and is not included in the quality result.

Privacy and fair-housing boundaries

Do not paste real lead data into a consumer AI chat just because this fictional test was useful. Check your account’s current data controls, retention, workspace terms, brokerage policy, and applicable privacy requirements first.

Do not infer protected traits, rank people from accessibility or family-status cues, characterize schools or neighborhoods, or let AI targeting logic substitute for a broker-approved process.

For implementation, keep the model disconnected from sending and CRM writes while the workflow is being tested. TPTK’s real-estate lead-follow-up automation guide covers the separate implementation layer, and the AI email assistant guide covers tool selection.

Evidence and limitations

  • This was one dated fictional packet and one completed run per product, not a representative benchmark.
  • The public article uses the product names Claude and Gemini because model-version screenshots were not returned with the evidence package.
  • Timing, interface screenshots, and measured human editing time were not recorded.
  • ChatGPT did not complete a comparable output and is excluded from the quality comparison.
  • The test does not prove legal compliance, deliverability, client acceptance, conversion lift, or production reliability.

The retained TWF-002 package includes the frozen input, prompts, raw first and repaired outputs, scoring rubric, evaluations, hashes, and run notes. Readers can request the synthetic evidence package or submit a correction.

For the earlier visual experiment, see the AI real-estate flyer test.

Frequently asked questions

Can AI write open-house follow-up emails?

Yes, as drafts for human review. In this test, both products produced useful structure, but neither output was safe to send unchanged because facts, priorities, unknowns, and promises still needed verification.

Which was better, Claude or Gemini?

Claude produced the safer and higher-scoring result in this one test. Neither passed the operational review threshold, so the practical answer is to use either only behind a human review gate rather than treat one as an autonomous winner.

Why did the repair prompt make both results worse?

The models evaluated their own work with the same imperfect understanding that shaped the first answer. Claude changed a correct priority, while Gemini claimed unsupported action language was fixed when it was still present.

Can I import AI-organized notes directly into a CRM?

Not based on this test. Use a review table or staging area first, verify every unknown, hold, priority, property fact, and proposed action, and import only the records a human has approved.

Were real leads or private client data used?

No. The packet contained 12 fictional leads. No email was sent, no CRM was updated, and no appointment or owner report was created.

Sources and test records

Leave a Comment