AI for Compliance Evidence in Pharma: Assembling the Pack, Not Signing It
On a pharmaceutical site the deliverable is rarely a decision on its own. It is a pack. A batch is not released because someone believes it is fine, it is released because a file exists showing every entry made, every result within limits, every deviation assessed, and a qualified person who put their name to it.
Most of the hours go into building those packs: retrieval across systems that do not talk to each other, cross-referencing, reconciliation, chasing signatures. The judgement, when it arrives, takes minutes and depends entirely on the pack being complete and true. That is why agents fit here. They are good at assembly across messy sources. They are not the signatory.
Batch record review and what review by exception really demands
Review by exception is defensible only when the checks are defined in advance, qualified and evidenced. You still have to show what passed, not only what was flagged.
An agent works that check set as the record closes: every required entry present and signed in the right order, in-process results against master limits, yields and reconciliations recalculated rather than trusted, equipment status and cleaning records covering the run window, every deviation named in the record matched to an event in the deviation system, and paper entries reconciled against the data the equipment recorded.
It returns two things. An exception list naming the page, the field, the expected condition and the observed one, ranked by criticality. And a positive assertion record of which checks ran, against which version of the master record, and passed. Without the second you have automation with no evidence it happened.
Deviation and CAPA files: assembly before investigation
Investigators lose the first days of a deviation to gathering. An agent compresses that into the first hour, assembling process and environmental data around the window, prior events on the same line or equipment, maintenance and calibration history, batch genealogy so the impact scope is a list rather than a guess, training records for the people who signed the affected steps, and the procedures as they were effective on the day rather than as they read now. Investigations get rewritten when someone notices the quoted procedure changed after the event.
It drafts the chronology, each line citing its source, and proposes the scope of the impact assessment. Root cause is not assembly work, and neither is CAPA design.
CAPA effectiveness checks fall due long after attention has moved on and are often closed with an assertion rather than data. An agent holds the definition, watches the due date and drafts the report from recurrence data a reviewer can verify.
Change control packs and closure evidence
The expensive question in change control is what the change touches. Not the regulatory strategy, which is a human call, but the mechanical sweep: which procedures quote the parameter, which forms and logbooks reflect it, which qualification documents describe the old state, which risk assessments assumed the old one, which curricula need reassignment, which quality agreements oblige you to notify a partner. An agent searching the controlled set returns a candidate action list citing the clause in each document at its effective version, so reviewers argue with a list rather than building one.
Closure is the weaker half at most sites, because changes are closed against a checklist of actions rather than the evidence each produced. An agent will not assemble a closure pack until the artefact exists: the revised procedure at its new effective version, completed training records, the executed requalification protocol with its attachments.
Supplier and contract manufacturer qualification evidence
Qualification files rot quietly. Audits fall due, certificates expire, a quality agreement stops describing what the contract manufacturer actually does, a supplier change notification arrives by email and is never assessed, and none of it surfaces until an inspector asks how you exercise oversight.
An agent maintains a live dossier per supplier and contract acceptor: approval scope, audit history and next due date, agreement version against current activities, certifications with expiry, change notifications and their assessment status, and performance data drawn from your own systems rather than their claims. From that it drafts the periodic re-evaluation and flags the forgotten gaps: person-in-plant reports not filed, batch records received but never reviewed, an activity being performed that the agreement does not cover.
Training records and SOP currency
Training completeness is genuinely binary, which is why it is checked so often and fails so often. The failures are structural. The population is never one clean list: employees, contractors, agency staff, engineers in for a shutdown, people who changed role mid-quarter. An agent reconciles that real population against curricula and completions and finds the recurring gaps: someone performing a task their curriculum does not cover, a leaver still holding assignments, a contractor with no curriculum, and above all a revised procedure whose read-and-understand was assigned but never completed before its effective date. That last one turns a document control finding into a data integrity conversation.
Procedure currency runs on the same machinery: documents past their review date, references to superseded documents, forms out of step with the procedure that calls them. The more valuable piece is the periodic review record itself, too often a signature stating no change is required. An agent drafts that review with its evidence attached, meaning the deviations, changes, complaints and CAPAs that referenced the procedure since the last review.
The annual product quality review
The product quality review is close to pure assembly, and it is usually a scramble in one month against a deadline. Batches manufactured, rejected, reworked or returned. In-process and release results with trends. Stability data. Deviations, changes and complaints. Qualification status of processes, equipment and utilities. Contract acceptor performance. Commitments made to authorities and whether they were met.
An agent compiles that continuously, keeps tabulations and trends current through the year and drafts the descriptive sections. The interesting output is the cross-reference a compiler misses: a trend a known deviation explains, or a parameter that never approaches its limit and may not be controlling anything.
Conclusions on process capability and specification adequacy stay with the reviewer, but turning the review from a retrospective compilation into a monitoring activity changes what it is worth.
Validation and computer system validation documentation
Validation packs age badly, and the gap between the documentation and the running system is a common finding. An agent maintains traceability from requirements through specification, test and defect to release, and reports the two failure shapes: requirements with no test evidence, and tests with no executed evidence, plus protocols missing attachments, signatures or dates. For periodic review of a validated system it assembles the pack from change records, incidents, access reviews and configuration state.
The recursion is worth stating plainly. An agent used in a GxP process is itself a system requiring qualification, and a non-deterministic one, so it is qualified on measured performance against real cases rather than a scripted pass. Prompts, retrieval configuration and model versions are versioned items under change control. Scope the intended use narrowly: an agent that drafts for a named reviewer is a far smaller exercise than one that decides.
Inspection readiness and the request in the room
Inspection is where evidence assembly becomes a timed exercise. A request arrives as a chain: the record for a given lot, the deviation it references, training records for the operators who signed the affected steps, the procedure as it was effective that week.
An agent in the back room resolves that into specific documents, retrieves both current and as-effective-on-the-day versions, checks each for completeness, notes anything requiring redaction, and produces a numbered, indexed pack listing whatever was not found and why. Between inspections the same checks run as a standing readiness posture, so gaps are found by you rather than in front of someone.
Speed pressure in the back room is exactly when wrong documents reach the front room, so the agent's job is to get a complete pack to a qualified reviewer faster, never to remove that review.
Audit trails, data integrity and who signs
An agent operating in a regulated process is an account, not a tool. It needs provisioned, reviewed access, and every action needs an attributable, time-stamped, non-alterable entry recording what it read, what it proposed and what the human did with the proposal. A reviewer must be able to reconstruct the evidence behind an output, which makes citation to source a design requirement.
It does not write into the record of truth unasked. Output lands in a review layer or as a draft and becomes record when a person accepts it. Signature stays human, and the agent must never be the reason a record looks contemporaneous when the activity was not.
Where AI fails here
Failures in the evidence
If the executed record is wrong, the agent assembles the wrong thing neatly, and neat is more persuasive than messy. Assembly quality is not evidence quality.
Handwritten and scanned entries carry real extraction error rates, and a misread value that falls within limits is more dangerous than an unreadable one that gets flagged. Tune for escalation over guessing. Retrieving a document as it was effective on a past date is hard where the system holds that history poorly, and citing today's procedure for a batch made under the previous version is a quiet, plausible error.
Failures in the deployment
Root cause is not assembly. A tidy, well-cited chronology can make a wrong cause look supported, and the reviewer complacency that sets in once output is usually good is the failure mode to watch.
Some evidence is not reachable at all: older systems and paper archives mean parts of the pack stay manual. The validation and lifecycle cost is real and ongoing, and for a low-volume process it will exceed the saving. Accountability does not move: the qualified person, the quality head and named signatories remain answerable.
A sensible first deployment
Start where ground truth exists and the agent only reads: training reconciliation, procedure currency, supplier qualification status. Run it in parallel with the current process and keep the comparison, which is both qualification evidence and your answer on whether it is good enough.
If you want a plain view of which of these an agent could carry inside your quality system, describe your current process to Adopt AI Partners and we will tell you where it holds and where it does not.