The Machine in the Margin: A Provenance File for AI-Assisted UK Proposals

AI drafting tools have moved from novelty to ordinary desk equipment in UK advisory firms. By late 2026, the more interesting question is not whether a proposal was touched by a machine, but whether the firm can say precisely where, how and under whose name. Clients have noticed. Procurement teams, general counsel and sponsor-level reviewers are asking sharper questions about provenance, and they are asking them earlier in the process.

That shift deserves a considered response. The answer is not a theatrical pledge that no AI is ever used, nor a vague assurance that all output is reviewed. It is a short, repeatable record — a provenance file — that sits beside the proposal and the project file. It should be unglamorous, quick to complete and robust enough to survive a challenge six months later.

The question behind the question

When a UK client asks how a proposal was produced, they are usually testing three things at once. First, whether the firm understands the sensitivity of the material: client names, commercial figures, regulatory assertions, personal data and unpublished intentions. Second, whether a named person has taken responsibility for the substance, rather than merely pressed send. Third, whether the firm can reconstruct its sources if a statement is questioned.

Those are familiar professional concerns. They apply to work drafted by a junior colleague, a specialist subcontractor or an external copywriter. AI does not create a new category of trust; it changes the speed and scale at which the question can arise. A document that once took a week to draft can now be assembled in an afternoon, with several iterations and sources blended together. The faster the draft, the more important the trail.

Four fields, honestly completed

A practical provenance file need not be a long policy document. For most UK advisory proposals, four fields will do more work than a page of general assurances.

What was machine-drafted. Describe the component, not the whole document. A first-pass structure, a rewritten executive summary, a comparison table or a suggested risk list can each be recorded separately. There is no need to disclose prompts or working conversations, particularly where those contain client confidences. The useful fact is the boundary: which passages began as machine output and which began with a person.

Who checked it. Name the individual. A role such as associate director or bid lead is not enough. The record should identify the person who verified facts, figures, commitments and tone, and the date on which that check took place. Where a second reviewer is required — for pricing, legal statements or regulated claims — record that too.

Which sources were used. This is the field most likely to be skipped and most likely to matter. A short source register can list the client brief, the published accounts, the framework terms, the named specialist's CV, the internal case study and the policy document relied upon. It need not reproduce the sources. It must make them findable.

What was excluded or escalated. Some material should never leave the firm without a human decision. Client names in unapproved case studies, forward-looking commitments, fee assumptions, regulatory interpretations and personal data all belong on an escalation list. A provenance file should show that the list was considered, not merely that it exists.

Two layers, one discipline

The error to avoid is treating disclosure as a single client-facing statement. In practice, two layers work better. The internal audit record is the fuller version: version names, reviewer initials, source references and a short note of any material change. It stays in the project file and supports quality review, handover and any later query.

The client-facing note is shorter. It might appear as an appendix line in the proposal or as a paragraph in the covering email: the firm uses AI-assisted drafting for specified elements; a named reviewer has checked all facts and commitments; the sources relied upon are available on request. That is usually enough. It signals control without turning the proposal into a technical manual.

This distinction matters in UK public sector and regulated work, where a client may have its own AI policy and its own disclosure expectations. The firm's record should be able to answer those expectations without disclosing confidential working methods. The aim is not transparency for its own sake. It is a credible account of professional judgement.

The commercial signal

There is a commercial argument here that goes beyond risk management. In a crowded UK advisory market, many firms now claim to use AI responsibly. Fewer can show a consistent trail. A client reviewing two comparable proposals may not reward the firm that says the most about AI. They may reward the firm that answers the practical question quickly: who stands behind this document, and how would we know?

A clear provenance file also reduces friction inside the firm. It gives reviewers a checklist, shortens the interrogation of late-stage drafts and makes handover to a delivery team less dependent on memory. It supports the same instinct that underpins good project governance: decisions should leave evidence, not just outcomes.

None of this requires a new bureaucracy. It requires a habit. Add a provenance section to the proposal template. Use file names that distinguish machine-assisted drafts from reviewed versions. Record the reviewer's name at the point of sign-off. Keep the source register short but real. Treat the trail as part of the work, not an admission about it.

The boring test

The best test of an AI trail is whether it is boring. If completing it takes ten minutes and reading it takes two, it will survive busy weeks. If it requires a separate committee, it will be skipped and replaced by assurances that nobody can verify.

The machine will remain in the margin of UK advisory work. The professional question is whether the firm's name in the centre of the page is supported by a record that can explain the margin. A modest provenance file does exactly that. It turns an awkward question into a demonstration of control — and control, as clients have always known, is rather more valuable than novelty.