Surveyor Guide to AI Governance for Valuations

A draft valuation can be technically sound and still contain a problem that matters to a lender: £5,000,000 in the executive summary, £5,050,000 in the conclusion, or an EWS1 disclosure mentioned in one section but absent from another. A surveyor guide to AI governance is not about handing those decisions to software. It is about setting rules for where technology can help review a report, what it can access and how the valuer remains in control before the report leaves the desk.

For valuation firms, AI governance should be practical. It should reflect the work being done: Red Book reports, lender instructions, comparable evidence and short review windows. A two-page policy that people follow is more useful than a lengthy document written for a generic technology risk register.

What AI governance means in a valuation practice

AI governance is the set of controls around the use of AI tools. It answers straightforward questions. Which tools may staff use? What data may be entered? What work can the tool assist with? Who checks its output? What record is kept?

The starting point is that the registered valuer remains responsible for the valuation opinion and the final report. AI can read, compare, calculate and flag. It does not inspect the comparable, assess its relevance, judge the quality of market evidence or decide whether an assumption is reasonable.

That distinction matters because a useful review tool can find contradictions that are difficult to spot when a report has been revised several times. It may identify that a floor area used in the valuation calculation differs from the area stated in the property description, or that a market rent of £32 per sq ft does not reconcile with the adopted value and yield. The finding is a prompt for review, not an instruction to amend the report.

Start with permitted uses, not broad statements

A policy saying that staff may use AI responsibly gives little direction when a report is due that afternoon. Define permitted uses by task instead.

For example, an approved review tool may be used to compare figures and statements across a draft report, test calculations, identify missing lender-required disclosures and check stated assumptions against client criteria. A separate approved tool might assist with meeting notes or internal administration, provided no confidential instruction material is entered.

The policy should also identify uses that need particular care or are not permitted. A public chatbot should not become a place to paste a full draft valuation report, borrower information or confidential comparable details simply because it is convenient. Nor should generative text be inserted into a signed report without the valuer checking each statement against the evidence and instruction.

The point is not to prohibit useful tools. It is to prevent staff making their own judgement about data handling and acceptable use under time pressure.

Match the control to the task

Not every AI use carries the same exposure. A spelling suggestion in an internal note is different from an automated review of a secured lending valuation. Governance should reflect that difference.

A report review tool needs controls over confidentiality, access, auditability and the way findings are presented. A tool that drafts prose needs additional controls because plausible wording can conceal an unsupported statement. Where a tool performs calculations, the firm should understand what it is calculating, what inputs it relies on and how the result is checked.

This is where a purpose-built valuation review platform differs from general-purpose AI. It can be configured to look for relevant issues: unreconciled market values, inconsistent yields, lease dates, tenant covenant statements, comparable evidence and instruction-specific requirements. It should still be treated as a second pair of eyes, not a replacement for the reviewer's own reading of the report.

Keep report data under control

Data security is usually the first concern, and rightly so. Draft reports contain information that is commercially sensitive and often personal. Before approving a provider, establish where data is processed, how it is encrypted, who can access it, how long it is retained and whether it is used to train any model.

The answers should be clear enough for a director or compliance lead to understand without interpreting technical language. If the provider cannot explain them plainly, that is a warning sign.

For a valuation practice, sensible minimum controls include named user access, strong authentication, defined retention periods and a clear process for removing data when it is no longer needed. Staff should use firm-approved accounts rather than personal logins. Access should be removed promptly when someone changes role or leaves.

It also helps to distinguish between processing and storage. A tool may need to process a report to carry out an audit, but that does not mean it needs to retain the document indefinitely. WriteUp, for example, uses private encrypted processing and does not store reports or use them for model training. Those details are the kind that should be confirmed and recorded for any supplier, rather than assumed.

Put human review at the centre of the workflow

The most effective governance control is a defined human decision point. A report should not move from an AI audit to issue without the valuer or designated reviewer considering the findings.

That does not mean every flag will result in a change. A comparable may appear to sit outside a stated parameter for a sound reason. A yield may differ between sections because one refers to an initial yield and another to an equivalent yield. The reviewer needs to decide whether the explanation is clear, correct and appropriate to the instruction.

A workable process is simple. Run the audit on the draft once the principal inputs and narrative are in place. Review each relevant finding against the source evidence. Correct genuine discrepancies, document why any material finding has not led to a change, then complete the usual professional review before issue.

This approach helps with confirmation bias. Once a valuer has worked through a report, it is easy to read what one expects to see. A system that reads the report as one document may flag that the adopted market rent in the valuation rationale is £28 per sq ft while the schedule and calculation use £30 per sq ft. The valuer then establishes which figure is right and whether the narrative needs correction.

Record enough to show control, not bureaucracy

Governance should leave an evidence trail, particularly for higher-risk work or lender-facing reports. The record does not need to be elaborate. It may show the approved tool used, the date of review, the person who considered the findings and any material actions taken.

This gives a firm a consistent answer when a director asks how AI-supported reviews are controlled. It also makes internal sampling more useful. If a recurring issue appears, such as incomplete marketing-period commentary or inconsistent treatment of unexpired terms, the firm can address the source of the issue through templates, training or review procedures.

Keep the tool's role accurately described. Do not record that an AI system approved a valuation or confirmed compliance. It reviewed the draft against defined checks and raised findings for a surveyor to assess. That wording reflects the actual process and preserves professional accountability.

Test the tool before relying on it

Supplier demonstrations are useful, but a firm should test an AI review tool against its own work. Use a small set of historic or test reports with known issues, appropriately handled for confidentiality. Include mistakes that occur in real drafting: a transposed date, a capital value mismatch, a missing EWS1 reference, an inconsistent net internal area and a calculation that does not reconcile.

Assess whether the tool identifies the issue, whether it produces irrelevant flags, and whether the finding is clear enough for a reviewer to act on. Also test reports from different sectors and formats. A process that works well for a straightforward residential report may need adjustment for a multi-let commercial asset with several lease events and more extensive comparable analysis.

Testing should not be a one-off exercise. Templates change, lender instructions change and software products change. Review performance after material updates and periodically sample issued reports to see whether the tool remains useful in practice.

Give staff a clear route for questions

Most governance failures are operational, not theoretical. Someone is unsure whether a new tool is approved, cannot access the firm account or receives a finding that does not make sense. If there is no clear route for resolving that question, people will work around the process.

Assign a named owner for approved AI use, even in a small firm. That person does not need to be a technology specialist. They need enough authority to maintain the approved-tool list, collect supplier information, arrange brief training and escalate concerns.

Training should use real report scenarios. Show staff how to interpret a flag for a £50,000 discrepancy in market value, how to check a missing disclosure against the instruction, and when a finding is simply a false positive. The aim is judgement, not button-pressing.

Good AI governance gives valuation professionals a controlled way to use a helpful review layer without diluting the professional thinking that sits behind every opinion. The best test is simple: can the firm explain, clearly and honestly, what the tool checked, how report data was handled and who made the final decision? If it can, the process is likely serving the report rather than distracting from it.