AI-assisted valuation should be tested on the quality of the final reasoning, not just the speed or fluency of a draft. Separate evidence extraction from the judgement about what that evidence means.

A valuation practice can have an accurate summary and still have a weak decision process. The summary might preserve the transactions, rents and dates, while leaving unclear why one comparable deserves more weight than another. The useful question is therefore not simply whether AI can write a valuation report. It is which parts of the work it should perform, and where the valuer must form and explain a view.

Separate the evidence from the inference

In a comparable review, distinguish source facts, interpretations and the final judgement. An AI tool might organise transaction evidence, identify missing fields and draft questions for review. The decision about relevance requires a separate step: differences in location, condition, lease terms, timing and transaction circumstances need to remain visible.

For each material statement, ask whether the reviewer can reach the underlying source. A linked source is useful only if it supports the specific statement. A transaction date should not silently become an assumed valuation date; a reported rent should not silently become an effective rent.

Design the expert's contribution

One design worth testing is to ask the valuer for an initial view on the most relevant evidence before showing an AI recommendation. Another is to display the AI's suggested weighting alongside the source details, with an explicit reason required when accepting or changing it. Neither arrangement should be assumed superior in every task.

The record should make disagreement intelligible. A reviewer who rejects an AI suggestion needs a place to record what changed their mind. A reviewer who accepts it should be able to show the evidence supporting that acceptance. A final approval box alone does not explain this reasoning.

Test a bounded task

Start with one part of the workflow, such as reviewing comparable evidence. Use authorised past cases with the information that was available at the time. Keep later knowledge out of the material presented to participants, and reserve a separate set of cases for the second test.

  • Evidence fidelity: are material facts, qualifications and sources preserved?
  • Reasoning: can another valuer follow the selection and weighting of evidence?
  • Exceptions: are missing information and unusual cases recognised?
  • Expert effort: how much time goes into checking, correction and rework?

Agree scoring criteria before reviewing results. Where professional judgement permits more than one defensible conclusion, assess the reasoning and evidence rather than treating agreement with a historic figure as the only measure of success.

Professional judgement remains central

RICS' professional standard on responsible AI use addresses maintaining professional judgement, data governance and output reliability. That provides relevant professional context. The workflow suggestions here are Axrea's practical design proposals, to be tested in the firm's own setting.

Our approach is test, redesign, test again. The outcome should be a clearer basis for deciding what to retain, change or stop, with evidence about both quality and expert effort.