Are AI Audit Findings Confidential? What Valuers Need

A draft valuation report can contain far more than a proposed market value. It may include borrower information, lease terms, tenant covenant commentary, floor areas, comparable transactions and lender-specific instructions. So, are AI audit findings confidential? They should be, but that depends on how the provider processes the report, stores data and controls access to the output.

The point is easily missed because an audit finding can look like a short, separate note: a yield is inconsistent with the valuation calculation, a date differs between the executive summary and body of the report, or an EWS1 disclosure is absent. Yet that finding is derived from, and may repeat, confidential report data. It needs the same care as the report itself.

For a valuer, the right question is not simply whether a tool uses encryption. It is whether the full journey of the report and its findings is controlled, documented and proportionate to the instruction.

Why AI audit findings are confidential

An AI review tool reads the report as one document. To flag that a £5,250,000 figure in the executive summary does not match the £5,200,000 figure in the valuation conclusion, it must process both figures and understand their context. The finding may also identify the property, instruction or report section in which the discrepancy appears.

The same applies to a finding on comparable evidence. A tool might identify that the report refers to six comparable transactions in the narrative but includes only five in the schedule. That output can reveal transactional evidence, the approach adopted by the valuer and information supplied for a particular lending decision.

Confidentiality therefore applies to three things, not just the uploaded file: the report itself, the extracted data used during processing, and the audit findings presented to the user. Treating the findings as less sensitive because they are shorter is a mistake.

What confidential processing should look like

A provider should be able to explain the process in plain terms. Technical language is useful where it answers a real question, but it should not obscure the basics: who can access the data, where it goes, how long it remains available and whether it is used for any other purpose.

Encryption is necessary, but not the whole answer

Reports and findings should be encrypted when sent to the platform and while held by its systems. This reduces the risk of interception and unauthorised access. It does not, by itself, answer whether the provider retains reports, whether staff can inspect them, or whether another service processes their content.

Ask how access is restricted internally. A sensible arrangement limits access to authorised personnel with a genuine operational need, rather than making customer documents available across a wider team. Account controls matter too. A report prepared for one instruction should not be visible to another valuer or another office simply because they use the same system.

No model training on client reports

This is often the clearest dividing line. Some public AI services may use submitted content to improve their models, subject to their terms and settings. That is unlikely to be appropriate for a confidential valuation report.

A specialist audit provider should state clearly whether reports, extracted information and findings are used to train any model. For confidential instructions, the expected answer is no. The report should be processed solely to produce the requested review, not retained as material for a broader dataset.

This point is worth checking in writing. “We take privacy seriously” is not the same as a clear commitment that customer report data is not used for training.

Storage and retention need a clear purpose

Some platforms need to retain a document briefly so that a user can revisit an audit result, download it or resolve a technical issue. Others may offer configurable retention periods. Neither approach is inherently wrong, provided the retention is transparent and supports the user’s needs.

The question is whether there is a defined deletion process. If a report is retained, how long is it retained? Can the customer request deletion? Are backups subject to a separate retention period? Does deletion cover both the original report and the audit output?

A provider should be able to answer these questions without sending you through several layers of generic policy wording. If the answer is unclear, assume the arrangement needs further scrutiny before live reports are uploaded.

Questions to ask before uploading a report

The practical test is straightforward. Before introducing any AI review process, ask the provider to set out how data moves through the service. The most useful questions are these:

• Is the report data encrypted in transit and at rest?
• Are uploaded reports, extracted data and audit findings used to train AI models?
• Who can access our reports and findings, and under what circumstances?
• Where is data processed and stored, including by any subcontracted service providers?
• How long are reports and findings retained, and how are they deleted?
• What contractual terms govern the handling of confidential client information?

These are not merely procurement questions. They go to whether the tool can sit within the way your firm already handles valuation instructions. A small practice may need confidence that each registered valuer can control access to their own work. A lender-side review team may need account permissions, an audit trail and a clear separation between panel firms.

It also helps to ask what the provider means by an “audit trail”. A useful trail records when a report was processed, the version reviewed and the findings generated. It should not create unnecessary duplication of confidential report content or expose it to people who do not need to see it.

Confidentiality and professional judgement are separate issues

Data confidentiality does not change the valuer’s responsibility for the report. An AI audit is a second pair of eyes. It can flag that the adopted rent of £185 per sq m does not reconcile with the stated annual income and floor area, or that a report says an unexpired term of 72 years in one section and 71 years elsewhere.

The surveyor decides whether the finding is material, whether there is an explanation, and whether the report needs amendment. Sometimes the tool will flag a point that is correct in context. A comparable may have been excluded for a stated reason, or a quoted figure may relate to a different tenancy assumption. That review remains a professional task.

Keeping the valuer in control is also good confidentiality practice. Findings should be presented to the person responsible for the report, with enough context to assess them, rather than automatically circulated or treated as a decision in their own right.

A practical approach for valuation teams

Start with a limited internal trial using reports that your firm is comfortable using for testing, subject to your own information-handling procedures. Review the provider’s data terms alongside the workflow. Decide who can upload reports, who can view results and whether findings should be saved to the job file.

Then test the output against realistic issues. For example, assess whether the system flags a missing EWS1 disclosure where it is required by the instruction, a market rent that differs between the valuation rationale and summary table, or a yield that does not reconcile with capital value and income. The aim is not to prove that software can replace review. It is to establish whether it helps your existing review process catch what a manual read-through might miss under time pressure.

WriteUp is designed around this use case for Red Book reports. It processes reports privately, with encrypted handling, no storage for model training and the surveyor retaining control of every finding. The important point is not the label on the technology. It is whether the processing arrangements and the review output stand up to the same scrutiny you would apply to the report itself.

A confidential AI audit should leave you with clearer questions before the report leaves your desk, not uncertainty about where your client’s information has gone.