Machine-Assisted Work, Human Accountability: Signing Off Consultancy Outputs Built With AI

It is now unusual for a substantial consultancy deliverable to be produced without some form of machine assistance. Drafting, summarising, coding, data cleaning, model building and slide production are all routinely supported by AI tools, and there is nothing inherently wrong with that. The appropriate question for a UK client is not whether AI was involved. It is whether the work has been checked to a standard you can defend, and whether a named person remains answerable for it.

The difficulty is that AI assistance is unevenly visible. A consultant may describe the work as their own because a human reviewed every page, while a junior member of the team used a generative tool to produce the first draft of the analysis. Neither party is necessarily acting in bad faith, but the client is signing off on an output whose provenance they do not understand. Five questions, put in writing before sign-off, close most of that gap.

1. Which tools were used, and at which stages?

Ask for a short written note identifying the categories of tool used and the parts of the deliverable they touched. The distinction that matters is between assistance and authorship. Using a tool to tighten prose or reformat a table is materially different from using one to generate the underlying analysis or the recommendations themselves.

A useful form of words is to ask whether any substantive conclusion, figure or recommendation in the deliverable originated in whole or in part from a generative tool, and if so which. A consultant who answers vaguely, or who insists that no AI was used at all, deserves a follow-up question about subcontractors and junior staff, who are often the earliest adopters.

2. Can every source and figure be traced?

Generative tools are fluent, and fluency is not accuracy. The specific risk in a consultancy context is the plausible but unsupported claim: a citation that does not exist, a statistic attributed to the wrong year, a regulatory position that was correct eighteen months ago. Ask for a source log that maps each material figure and legal or regulatory statement to a dated, verifiable reference.

For UK work, insist on vintages. Office for National Statistics releases are revised; tax thresholds and reliefs change between fiscal years; guidance from regulators is updated without ceremony. A recommendation built on last year's figures may not be wrong so much as stale, which is often worse because it looks sound. The question to ask is simple: can you show me where each number came from, and when it was current?

3. Where did our data go, and on what terms?

The use of AI tools brings UK GDPR obligations into the engagement, and these are the client's obligations as much as the consultant's. Ask which tools received client data, whether that data was used to train or improve a model, where processing took place, how long inputs are retained, and which sub-processors are involved.

Where a supplier's standard terms are being relied upon, ask for them. Where personal data is involved, confirm the controller and processor roles in writing and check that the contract includes the required processor terms, including any restriction on international transfers. If the work involves special category data or a large volume of customer records, ask whether a data protection impact assessment has been carried out. A consultant who cannot answer these questions promptly is unlikely to have thought about them at all.

4. Who owns what is produced?

The ownership position for AI-assisted material is less settled than for conventional work. Ask for an explicit statement covering copyright in the deliverable, the underlying models and code, and any third-party content embedded within it. Confirm that you hold a licence to use, adapt and share the output internally and with your advisers, and that no restriction flows from the tool provider's terms.

Ask also for a warranty that the deliverable does not infringe third-party rights, supported by an indemnity proportionate to the risk. Where a tool provider offers its own indemnity, ask whether it passes through to you or stops with the consultant. The answer is frequently the latter.

5. Who signs it off, and what did they check?

This is the question that resolves most of the others. Name the individual who is professionally accountable for the deliverable, confirm their role and qualifications, and ask what review they personally performed. 'The tool produced it' is not an answer that survives contact with a board, a lender or a regulator.

Ask for evidence of the review process: who tested the figures, who read the source material, whether any part of the output was independently reproduced. Where the recommendation is material to a decision, consider whether your own team should spot-check a sample of the underlying work. That is not distrust; it is the same discipline you would apply to any adviser's calculations.

Clauses worth agreeing before the work starts

Most of this is easier to settle at contracting stage than at sign-off. A short schedule to the engagement letter can cover the following:

• A named individual responsible for the deliverable and for any subsequent correction.
• A requirement to disclose the use of AI tools in the production of material outputs.
• A source log for all figures, statistics and regulatory statements, with dates.
• Confirmation of the lawful basis and data flows for any client data shared with a tool provider.
• An assignment or unrestricted licence of the output, with a non-infringement warranty.
• A remediation commitment: if an error is found after sign-off, who fixes it, at whose cost, and within what period.
• A cap on liability that is proportionate and, importantly, backed by current professional indemnity cover.

The presence of AI in the process does not lower the standard of care. If anything it raises the value of a clear audit trail, because the work can no longer be explained simply by pointing to the person who wrote it. A consultant who welcomes these questions is usually one whose review process is already sound. A consultant who resists them is asking you to accept an output on trust — which, for a decision of any consequence, is not a position a UK client should agree to.