Article

Your AISOP: What Does Your SOP for the AI Look Like?

· Matthew DeWolfe

What an AISOP is

An AISOP is a standard operating procedure for the AI. It says what the AI may do, what a person must decide, and how you prove it afterward.

We write SOPs for everything people do in a regulated business. AI is now doing work in those same businesses, often with no equivalent. So the useful question is no longer whether to use AI. It is this:

What does your AISOP look like?

Where things go wrong

Compliance rarely fails inside one system. It fails in the gaps between systems, and AI can widen those gaps.

The seams. Ask a quality manager how many software systems their company runs. One for documents, one for training, one for audits, one for CAPAs, one for the lab, plus a few that came with the last acquisition. Each works on its own. But a lapsed training record never reaches the inspection schedule, and a failed QC result never triggers the CAPA.

The quiet disagreements. Then the systems start to contradict each other. The SOP is on revision 4 in document control and revision 3 in the training system. A client is “approved” in one place and “on hold” in another. None of it looks serious alone. Together, it is what an assessor finds in an afternoon.

The AI on top. Every vendor is adding AI to its own product, and each AI sees only its own slice of the data. So which AI, reading which version of the truth, is producing your work? And who checks it before it goes out the door?

This is no longer theoretical. In April 2026, FDA issued what DLA Piper describes as its first warning letter citing a drug manufacturer for improper reliance on AI. The firm used AI agents to generate required records. Staff said they were unaware of certain requirements because the AI never flagged them. FDA's position, as summarized there, is that AI outputs must be reviewed and approved by an authorized quality unit representative.

Trust needs checking too. In October 2026, an outside researcher found that eight intermediate keys at a certificate authority were trivially weak. One had a value of just 4. Testing, inspection and certification works on the same kind of trust: people rely on the certificate without looking underneath. So who checks the checker?

The rules are moving the same way. ISO/IEC 17020:2026, published on 27 March 2026, adds a section on data and information control. It also extends facility and equipment controls to AI and remote tools. The TIC Council says accountability for compliance decisions “remains non-delegable.” “We have a system for that” will not be enough. You will need to show how the system, and the AI inside it, is controlled.

The three layers of return

A good AISOP protects three kinds of return, and most AI business cases only count the first.

LayerWhat it meansWhen you see itWhat the AISOP must settle
TimeHours saved on review and paperworkEarlyWhat the AI checks, and what a person still checks
ProtectionProblems caught and stopped before they reach a client or regulatorUsually only when it is goneWhich gates cannot be skipped, and how you prove it
CapabilityMore people doing expert-level work, consistently, with a system behind themOver yearsWhose knowledge the AI carries, and who keeps it current

Time is the easy one. Enveda published a study of AI-assisted review on 1,061 pages of executed batch records. Across its runs, the AI raised all 11 actionable issues the human reviewers found, and one run added seven more. Review time fell about 89%, from 25.9 hours to 166 minutes.

The guardrails matter as much as the numbers. It was one exploratory campaign. The tool was advisory only and not validated as a GMP system. Each run also made a misread, and humans decided everything. That is what a sensible AISOP looks like in practice.

Protection is the one you only notice when it is gone. Nobody sees the recall that did not happen. I believe the true return on strong quality takes years to show, so any number quoted today is a hypothesis.

Capability changes an industry. Many experienced reviewers are nearing retirement, and their know-how can leave with them. AI will not replace their judgment. It can carry their checklists forward and give the next generation a system to stand on. But automation is a snapshot of human know-how, and it drifts unless someone owns it.

What a good AISOP contains

A useful AISOP fits on one page and answers ten questions. If a section is blank, that is your gap.

SectionThe question it answers
Purpose and scopeWhich work does this AI touch, and which work does it never touch?
InstructionsWhat is the AI told to do, and did an expert approve those instructions?
Human gatesWhich outputs must a qualified person review and sign before they go anywhere?
BoundariesWhat can the AI read, what can it change, and what must it never do?
OwnerWhich named person is accountable for this AI?
Version and change controlWhat changed, when, and who approved it?
Review scheduleWhen do you check that it still performs as expected?
Evidence it worksWhich tests show it does the job, and what happens when it fails one?
TrainingWho is trained to use it and to challenge its output?
Records and logsCan you show what the AI produced, which records it drew from, and who made the final call?

An instruction file is a good start, but it is not enough alone. It needs gates, tests and an owner. Competence without checks breeds complacency, and the same is true of AI.

The question to take back to your team

Compliance is not a set of modules. It is a closed loop: one record, one check, one decision, one proof.

So read your AISOP three times: once for time, once for protection, once for capability. Which one would it fail?

And if you do not have one yet, start with the three lines that matter most. What the AI checks. What a human decides. How you prove it.

Published by AI Lab Services (AILS), applied AI R&D for regulated industries. Questions and comments: matthew.dewolfe@brslab.com

Sources