Subfractal
Set the rules

If an AI Drafts It and You Sign It: Ten Things an Assessor Is Likely to Ask

We could find nothing in NATA's criteria that prohibits AI; the exposure sits under requirements that already exist. Here are ten questions to be able to answer before a tool goes near a report.

For: People authorised to release results at NATA-accredited laboratories and inspection bodies, and the quality managers asked to approve an AI tool

You are authorised to release results. Someone proposes a tool that will draft part of a report, gather the evidence you review, or flag which results need a second look. You will still sign.

What will an assessor ask about AI?

Nobody outside NATA can tell you for certain. This list is our inference from public documents, not knowledge of how any assessment will go: NATA's application documents (the inspection one now names AI), NATA's commentary on AI, and the bulletins the UK's accreditation body, UKAS, has issued. The background is in What NATA has said about AI.

On 13 October 2026, NATA published a short article, "Can you Trust an AI-Supported Result?". It offers six questions: purpose, data, validation, limitations, oversight and accountability. NATA says:

"They should not be treated as a formal NATA checklist, but as useful questions to support responsible decision-making."

The same is true of ours, which go further into the paperwork.

The ten at a glance

# Question What to have on paper
1 What does it do, and not do? One paragraph of scope and limits
2 How do you know it works? A validation record with criteria set beforehand
3 What happens when it changes? Change control, and a fixed version
4 Who released this result? A named, authorised person and a record
5 Can that person explain it? Their reasoning, without the tool
6 What does the record show? Tool output, its evidence, the decision
7 Where does the data go? A data-flow answer you can say in one breath
8 What happens when it is wrong, or off? A risk entry and a fallback
9 What did you ask the supplier? Written answers to the questions below
10 Are your people still competent without it? A periodic unaided check

The ten, in detail

1. What does it do, and what does it not do?

One written paragraph. "It lays the batch's control results beside the last thirty." "It does not accept or reject a result." The second sentence matters more. A tool with no stated limit will be assumed to have none.

2. How do you know it works?

A validation record. Before use, the tool was run against cases where you already knew the right answer, with acceptance criteria written down beforehand. NATA's inspection criteria say validation "must demonstrate the suitability of the system for its intended purpose". The UK bulletin asks facilities to validate AI systems "according to a documented procedure", as extensively as the application needs.

NATA notes that commercial off-the-shelf software "may be considered pre-validated when used within its specified application range". Whether a general-purpose AI model has a specified application range is worth asking your assessor rather than assuming.

3. What happens when it changes?

NATA's inspection criteria are direct. Software changes, configuration amendments and system updates go through change control, "including authorisation, documentation and validation, before implementation". The UK bulletin says the same of AI: work out the impact of a change, and validate again if it affects the original validation.

This is where cloud services can be hard to pin down. If the supplier updates the model without notice, you cannot authorise the change before it happens, because you do not know it has happened.

4. Who released this result?

A named person, authorised by the facility, with a record that they did it. NATA's application documents for laboratories and inspection bodies both require results to be authorised before release. The joint UKAS and DAkkS bulletin goes further: final decisions on conformity "must not be delegated to an AI system under current accreditation frameworks". A draft that can reach a customer without passing that person is the obvious gap.

5. Can that person explain it?

An assessor may well ask them. NATA's guidance to its technical assessors tells them to probe "through discussions, with facility staff", and to establish whether staff "understand the principles and limitations of the methods and procedures adopted". Its remote assessment policy lists interviews with staff authorised to release results among activities that can be done remotely.

"The system flagged nothing" is not an explanation. The person signing should be able to say what they looked at and why they were satisfied, as they would if no tool existed.

6. What does the record show?

For any result the tool touched: what it put forward, on what evidence, and what the person decided, including any override. NATA asks that records be kept in a way that "prevents amendment or loss of the original". A report that looks identical whether a person or a tool drafted it leaves a gap.

7. Where does the data go?

NATA's Member Advisory Forum, discussing AI in February 2024, said "confidentiality and software verification or validation, as appropriate, should be prioritised". NATA's guidance to its own assessors discourages tools like ChatGPT for assessments, partly on confidentiality grounds. The joint UK and German bulletin gives "processing data locally to ensure confidentiality" as one example of how a body might comply.

The simplest answer is that the tool runs on your own hardware and nothing leaves the building. It is not the only acceptable answer, just the one that needs the least explaining.

8. What happens when it is wrong, or off?

An entry in your risk register. What if it misses something? What if it flags so much that people stop reading? What do you do on the day it is down? If the honest answer is "we could not report", you have built a dependency, and should say so.

9. What did you ask the supplier?

NATA's inspection criteria say bodies "remain responsible for ensuring that externally provided products and services affecting inspection activities are suitable and fit for purpose". The UK bulletin says an off-site or externally provided AI system should meet the same requirements as one you run yourself. Ask any supplier what you would ask a calibration provider, in writing:

  1. Does the tool decide anything, or only put things forward for a person to decide?
  2. For every draft or flag, can it show the evidence it used?
  3. Where is our data processed and stored? Can it run on our own hardware? Is our data used to train anything?
  4. Can we fix the version? How are we told before it changes, and can we hold an update until we have validated it?
  5. What validation evidence will you hand over, and what must we validate ourselves?
  6. How do you control your own changes?
  7. If you stop trading, do we keep our records, and can we switch the tool off without losing the ability to report?
  8. What insurance do you carry?

A supplier uncomfortable with those questions has answered them.

10. Are your people still competent without it?

NATA's February 2025 commentary warned that "over-reliance on AI could result in a gradual reduced role for human judgment". The UK bulletin expects staff to be able to "recognise and evaluate the significance of any unexpected outputs". A reasonable answer is a periodic check: a sample of results reviewed the old way and compared.

What is not on the list?

The UK bulletin says it does not "establish new or additional requirements for accreditation", only contextualises existing ones. Every question above could be asked of a spreadsheet with macros, and you have probably answered most for your reporting software already.

The uncomfortable part is that questions 3 and 7 are often the hardest to get straight answers to.

Two caveats

NATA's technical assessors are volunteer industry experts, and NATA's own guidance says they "may differ in their approach". Without specific AI guidance, two assessors may weigh the same tool differently. Clear paperwork narrows that gap. It does not close it.

And we could find no NATA guidance on whether it wants to be told when a facility adopts a tool like this. UKAS does. A careful facility might ask NATA in writing; we explain why in What NATA has said about AI. Inspection bodies should also read what clause 6.2.9 asks of a body that uses AI.

What can you do this month?

It is worth doing before anyone offers you a tool.

  1. Pick one decision you sign: releasing a report, accepting a batch, closing a nonconformance.
  2. Write down what you look at before you decide. All of it, including what is not in the procedure.
  3. Put the ten questions beside it and answer each as if a tool were gathering that evidence for you.
  4. Mark the ones you could not answer. Those are your conditions for any supplier, and the start of your written questions to them.

One question

A question for people who sign: what would have to be true for you to trust a draft you did not write?

Write to [email protected]. We will share what we hear, without names.

Sources

  1. General Accreditation Criteria: ISO/IEC 17020 Standard Application DocumentNATA, Issued September 2026, effective October 2026
  2. General Accreditation Criteria: ISO/IEC 17025 Standard Application DocumentNATA, October 2024
  3. Technical Assessor Information and Guidance DocumentNATA, May 2026
  4. General Accreditation Criteria: Remote assessment policyNATA, December 2022
  5. Member Advisory Forum minutesNATA, 20 February 2024
  6. Part 2 – NATA: Assessing the adoption of AINATA, 3 February 2025
  7. Can you Trust an AI-Supported Result?NATA, 13 October 2026
  8. Technical Bulletin: The use of Artificial Intelligence (AI) technologies in accredited conformity assessmentUKAS, 20 June 2025
  9. Technical Bulletin: The use of Artificial Intelligence (AI) technologies in accredited conformity assessment (joint UKAS and DAkkS bulletin)UKAS and DAkkS, March 2026

Facts last checked:

Subfractal is not affiliated with or endorsed by NATA, and nothing here has been reviewed by it. This is general information, not accreditation or legal advice.

Who wrote this

Dom O'Carroll is a co-founder of Subfractal.