Human in the Loop, Evidence on File: An AI Sign-Off Checklist for UK MedComms Teams

The tools arrived quietly. A drafting assistant here, a reference-checking plug-in there, a slide template that assembles itself from a structured content file. By 2026, few UK medical communications agencies would claim that no generative system touches their output at any stage. Far fewer can produce, on request, a coherent account of who checked what, against which sources, and on what date.

That gap matters. The regulatory framework has not been rewritten around artificial intelligence, and it does not need to be. The ABPI Code of Practice still places accountability on named certifying signatories. The MHRA's expectations of advertising and promotional material remain unchanged in substance. What has changed is the number of invisible hands between a blank page and a final deliverable — and the ease with which a review step can be assumed rather than performed.

The principle that settles most arguments

The machine never signs. Whatever a tool produces, a named human being is accountable for its accuracy, balance and compliance, and that person must be identifiable from the project record alone. This is not a philosophical position; it is the practical test a client auditor or a self-regulatory case officer will apply when something goes wrong.

The rest of this article is a checklist built around that principle. It is designed to be lifted into a standard operating procedure, a project initiation document or a quality management system without much adaptation.

The checklist

1. Named ownership at each stage

Every deliverable should carry a simple three-column record: stage, activity, named owner. Typically this means a writer for first draft, an editor for source verification, a medical reviewer for scientific accuracy, and a nominated signatory for final approval. Where a tool contributes to a stage, the human owner of that stage remains unchanged. A useful discipline is to write the owner's name before the tool is used, not afterwards.

2. Prompt and version records

Where generative tools have contributed substantive text, retain the tool name, version, date of use and the prompt or instruction set. This is not bureaucracy for its own sake. If a claim is later challenged, the difference between "the writer drafted this" and "the writer directed a tool, then verified the output against the approved reference" is material to how the error is understood and remediated.

3. Source verification against approved references

Automated citation tools are useful for locating a paper; they are not evidence that the paper says what the draft claims. Reviewers should confirm that every factual statement traces to a source on the client-approved reference list, that the specific passage supports the specific claim, and that the citation is formatted to the required journal or client style. Record the initials of the verifier and the date. Where a reference could not be verified, note the escalation rather than deleting the trail.

4. Balance, claims and comparisons

AI-drafted copy has a tendency towards confident phrasing. Reviewers should interrogate superlatives, unqualified comparative claims, implied efficacy and any statement that would require substantiation under the Code. Check that fair balance has survived the editing process intact — particularly where a summary section has been regenerated or shortened. Comparative claims against named competitors deserve a second, independent review.

5. Data, confidentiality and tool approval

Confirm that the tool in use is on the agency's approved list and that no client-confidential material, unpublished data, patient-identifiable information or pre-approval content has been entered into a public system. UK GDPR obligations apply to anything typed into a prompt. A short project-level declaration — tool approved, data classification observed, named approver — takes minutes and closes an obvious exposure.

6. Visual and slide production

Automated slide assembly and image generation introduce their own risks: mislabelled axes, distorted graphs, synthetic imagery that misrepresents a patient population, or accessibility failures. Reviewers should confirm chart data against the source table, check alt text and colour contrast, and verify that no generated image implies a real person or a clinical outcome. Any figure that appears in a promotional deck should be traceable to an approved data source.

7. The sign-off record itself

A single page per deliverable is usually enough: deliverable title and version, tools used, stages and owners, verification dates, open issues, and the final signature with date and role. Store it with the project files, not in a separate folder that nobody opens. The record should be intelligible to someone who joins the project six months later and has no memory of the conversation in which decisions were made.

What auditors and clients actually ask

In practice, three questions recur. First, can you show me the review trail for this document? Second, who verified the references, and when? Third, how do you know the tool did not introduce content that nobody checked? A team that can answer those three questions in under five minutes is in good shape, regardless of how sophisticated its tooling is.

Common failure modes

• Review by osmosis. A document passes through four pairs of hands and nobody can say which of them read the references.
• Tool-by-tool approvals. Individual freelancers use different assistants under no common policy, so no single record exists.
• Silent regeneration. A section is re-run through a tool after review, invalidating the earlier check without anyone noticing.
• The unwritten escalation. A reviewer spots a doubtful claim, raises it verbally, and the trail records nothing.

A proportionate response

None of this requires a large compliance function. It requires three habits: name the human, record the source, file the sign-off. Teams that adopt them tend to find that the process sharpens the work as well as protecting it — a draft whose every claim has been traced is usually a better draft.

For UK medical communications, the competitive advantage in 2026 lies less in who has the cleverest tools and more in who can demonstrate, credibly and quickly, that a competent person stood behind the output. That is a reassuringly old-fashioned answer to a very modern question.