Several buyers can view the same property and reach opposing conclusions. One likes the open-plan kitchen; another says it provides too little separation. A summary saying buyers liked the layout loses the second view. A summary saying buyers disliked it loses the first. The useful report preserves both statements and gives the agent enough evidence to judge what response, if any, is appropriate.
The proposed workflow below groups feedback without manufacturing consensus. It does not predict sale probability, determine an appropriate price change or claim that a viewing comment represents the entire market. AI prepares a report from attributable records, while the agent decides how to explain the feedback and what action to recommend.
Establish which feedback belongs in the report
Start with a stable property and listing identifier, a defined reporting window and the viewing records to include. Link each response to its viewing and respondent reference. Use a privacy-appropriate reference in a seller-facing report; keep any contact details in the authorised underlying record rather than copying them into every summary.
Separate a buyer's words from an agent's note about the conversation. If the note paraphrases rather than quotes, label it accordingly. A message saying the buyer seemed hesitant is the agent's observation, not a buyer statement that the price is too high. Keep the evidence type with the text.
Record nonresponses explicitly. A buyer who has not replied is not automatically uninterested. A cancelled viewing has no completed-viewing feedback unless the person gave a relevant reason, which should remain labelled as cancellation context. Scope decisions prevent the model from interpreting every event as a comparable opinion.
Group themes without choosing the dominant story
Use proposed themes such as layout, condition, location, stated price concern and requested next step. Allow a response to belong to more than one theme while retaining its source. Do not force a comment into the nearest available category when it raises a different issue.
Within each theme, preserve favourable, unfavourable and ambiguous statements. A short report can say the layout received opposing comments rather than hiding the minority view. Avoid an unsupported confidence score or a universal sentiment scale. The agent should understand what was actually said and how the grouping was performed.
Source: OpenAI Structured Outputs documents schema-constrained responses. A proposed schema can require each theme to include supporting response references, opposing evidence and unresolved wording. The schema controls structure, not the faithfulness of the summary. The reviewer must compare the paraphrase with the original meaning.
Keep counts reproducible and modest
Calculate counts outside the model using the accepted response identities. Two messages expanding one buyer's opinion should not become two buyers. If several people attend together and provide one joint response, state how the agency counts it. Do not imply independent opinions where the source is a shared statement.
Report the denominator with any count: three attributable responses out of five completed viewings, for example, is more useful than most buyers. Small samples should remain descriptive. They do not establish market demand or justify a price conclusion on their own. If the report window changes, preserve the earlier report version so its counts can be reproduced.
Source: OpenAI file search describes retrieval and citations from uploaded files. It can help locate relevant viewing notes, but retrieval alone does not guarantee complete coverage. The reporting process should start from an accepted viewing/response list and reconcile missing records rather than assume search results contain every response.
Reusable disagreement-preserving feedback report
Use this proposed template for one property and reporting window. Fill the counts from the reconciled list rather than asking the model to infer them from prose.
Property: [Listing and property IDs; approved record version]
Scope: [Reporting dates; completed viewings; attributable responses; missing replies; counting rule]
| Theme | Supporting comments | Opposing or qualifying comments | Evidence references |
|---|---|---|---|
| [Theme] | [Faithful summary; response count] | [Faithful conflicting summary; count, or none received] | [Viewing and response IDs] |
| [Theme] | [State observations without overclaiming] | [Ambiguous wording and question to resolve] | [Original message or note reference] |
Unresolved evidence: [Missing response, unclear wording, disputed viewing match or unattributed note; owner and next action]
Agent interpretation: [Explicitly labelled interpretation, evidence considered and limits of the sample]
Agent recommendation: [Proposed next step and rationale, or no action recommended; kept separate from buyer statements]
Review decision: [Agent accepts, revises or holds this report version; date and reason]
Before release, check that every attributed statement points to the right response, opposing views remain present, counts match the accepted list and missing replies are not described as opinions. The completion check is a report the agent can trace back to source messages without guessing how the summary was formed.
Work through opposing comments and duplicate messages
Consider a hypothetical property with four completed viewings and three distinct responses. Buyer A likes the open-plan layout. Buyer B says the same layout does not suit their need for a separate kitchen. Buyer C reports that the bedroom size is suitable but asks about a feature not confirmed in the record. The fourth viewing has no reply.
The proposed report records two opposing layout comments, one bedroom-size comment and one unresolved feature question. It states three responses from four completed viewings. It does not label the layout generally popular or unpopular, and it does not treat the nonresponse as rejection. The agent may recommend checking the feature question before the next viewing, clearly as their recommendation.
Suppose Buyer B later sends a second message explaining the separate-kitchen preference. That adds evidence to the same response; it does not raise the response count to four. If two notes use the same name but different viewing IDs, staff verify whether they refer to one person or two. Similar names do not justify merging feedback automatically.
Now imagine an agent note says buyers found the property expensive while none of the retained responses contains that statement. The report flags the unsupported generalisation and asks the agent for the underlying evidence. It can retain a labelled agent interpretation, if the agent chooses, but must not rewrite it as a shared buyer view.
Review recommendations separately from the evidence summary
The agent should approve both the faithful summary and any seller-facing recommendation. A correct feedback summary may still be followed by an unjustified instruction to lower the price or renovate. Keep those decisions in a separate section and route them through the agency's professional process.
Source: n8n human review describes approval or denial of selected agent tool actions. It can provide a review boundary before a report is sent. It does not judge whether an agent's pricing recommendation is justified or whether the feedback sample represents prospective buyers generally.
Evaluate the proposed summariser against agent-reviewed examples that include mixed opinions, sarcasm, unclear pronouns, joint responses and missing replies. Check omitted disagreement, invented sentiment, wrong property references and duplicate counts. Review the effort required to trace evidence, not merely whether the prose reads smoothly. Any claimed improvement should follow an actual comparison with the team's current reports.
Questions about viewing feedback summaries
Should we report only themes mentioned by most buyers?
No universal majority rule is appropriate here. A less common comment can be important, and opposing views may concern the same feature. Define the reporting policy with the agent, retain the evidence and state counts rather than suppressing minority comments automatically. The report's purpose is faithful communication, not consensus manufacture.
Can a model infer why a buyer did not respond?
It should not. Record the missing reply and any known follow-up state. Silence can have many explanations that are not present in the evidence. If the agent obtains a reason later, add it as a new attributable record instead of treating an earlier model guess as established history.
Does a repeated price concern mean the asking price should change?
That is an agent decision requiring broader professional judgement. The report can state the attributable concern and its limited sample. It should not turn descriptive feedback counts into an automatic pricing conclusion. Separate the agent's recommendation and rationale from the buyer evidence so the seller can understand both.
If your business needs help defining this process, explore AI agents for estate agencies, the wider AI automation services, and our custom-agent workflow guide. The agents and automation comparison and custom AI agent glossary explain the terms. To discuss your records and approval rules, get in touch.