Aprentiz home Apply for a founding research place
Menu

Questions to ask any AI supplier — including us.

Built from what the FCA and the SRA now expect of firms using AI. Our own answers are set apart and held to the same standard.

Printed, it leaves out our answers, so you can use it with any supplier. It isn’t legal or regulatory advice, and every reason we give is quoted word for word, with its source.

Testing before you rely on it.

Questions 1 to 3 of 14.

  1. Question 1: How was it tested before we use it, and can we keep the results?

    Why it matters

    Structured challenge, testing and sign-off before deployment. FCA, 19 October 2026, §3.1.

    Overreliance on untested vendor assurances. FCA, 19 October 2026, §4.2.

    A good answer
    Testing on your own work, before go-live, with results you keep.

    Aprentiz’s answer

    The founding programme rehearses your workflow on example clients before anything touches a real client, and you keep the results. We also try to break it: pressure, repeated asking, a different model.

  2. Question 2: What does it do when the evidence isn’t there?

    Why it matters

    …where the AI tool should abstain, redirect or escalate to a person FCA, 19 October 2026, §2.1.

    A good answer
    It says so, and stops, rather than filling the gap.

    Aprentiz’s answer

    It stops and names what’s missing. It won’t give a conclusion without evidence, and a gap is never shown as covered.

  3. Question 3: Can we check every citation against the source?

    Why it matters

    Reliance on an output of AI would not be a suitable defence in this scenario. SRA, 17 August 2026, AI hallucinations: relevant considerations.

    A good answer
    Every citation links to the exact source text.

    Aprentiz’s answer

    Yes. Each cited sentence is checked against the exact source text, and the original is one click away. A citation that doesn’t match is rejected.

Accountability and oversight.

Questions 4 to 6 of 14.

  1. Question 4: Who in our firm is accountable, and how does the tool help them show reasonable steps?

    Why it matters

    Assigning clear SMF ownership for AI risks. FCA, 19 October 2026, §3.3.

    …must take all reasonable steps to ensure compliance with relevant regulatory arrangements, including those relating to supervision SRA, 17 August 2026, AI hallucinations: relevant considerations, citing paragraph 9.1.

    A good answer
    A named person decides, and the record shows what they saw.

    Aprentiz’s answer

    A named person in your firm makes the call and signs it off. The record shows who approved it, what they saw and why.

  2. Question 5: Is the human check real oversight, or box-ticking?

    Why it matters

    There was a risk of human review of AI outputs being a ‘box-ticking’ exercise, rather than meaningful oversight. FCA, 19 October 2026, §3.4.

    A good answer
    People decide where it matters, with the evidence in front of them.

    Aprentiz’s answer

    It stops for a person only where a decision matters, with the evidence in front of them. Their reasons are recorded, not just a tick.

  3. Question 6: What will we have on file for the board, the regulator and our insurer?

    Why it matters

    Some firms’ decisions on AI tools were poorly documented with limited audit trails or supporting rationale. FCA, 19 October 2026, §3.3.

    In the event of a complaint, the file will determine the outcome. Howden, 1 June 2026, Adviser claims trends.

    A good answer
    A complete record for each decision, not a log of prompts.

    Aprentiz’s answer

    A sealed record for each decision: the rule text and its version, the evidence, the gaps, the decision, and who approved it and why.

Model changes and exit.

Questions 7 and 8 of 14.

  1. Question 7: Will we know before the model changes, and what happens to decisions we’ve already signed?

    Why it matters

    Maintaining visibility of vendor model changes. FCA, 19 October 2026, §4.2.

    A good answer
    Changes are visible in advance, and signed decisions don’t silently change.

    Aprentiz’s answer

    A signed-off decision isn’t reopened without new evidence, whichever model runs. A different model’s conflicting answer is set aside for a person to review.

  2. Question 8: If the model is retired, or the supplier is, what’s our exit plan?

    Why it matters

    When the model was retired, the firm had to redesign the product. FCA, 19 October 2026, §4.2, case study T.

    Single-platform dependence without exit options FCA, 19 October 2026, §4.2.

    A good answer
    Your method and records survive a change of model or supplier.

    Aprentiz’s answer

    Aprentiz isn’t tied to one model. Your method, your data and your records are yours to keep.

Your data and confidentiality.

Questions 9 to 11 of 14.

  1. Question 9: Where is our data stored and processed, and by whom?

    Why it matters

    …legal professional privilege may be permanently waived and unable to be recovered SRA, 17 August 2026, Client confidentiality: relevant considerations.

    A good answer
    A named place, named providers, and nothing sent to outside AI services.

    Aprentiz’s answer

    In a private environment for your firm alone, on AWS in London (eu-west-2), which we run during the founding programme. No outside AI service sees your work: the models run privately, with no fallback to one.

  2. Question 10: Will our data be used to train any model?

    Why it matters

    …is not used to train AI models except where explicitly authorised and appropriate to do so SRA, 17 August 2026, Client confidentiality: relevant considerations.

    A good answer
    Never without your explicit, written authorisation.

    Aprentiz’s answer

    No. We never train or fine-tune any model on your firm’s material. During the founding programme we work on example clients, not your client files.

  3. Question 11: Does any of our data go to an outside AI service?

    Why it matters

    …legal professional privilege may be permanently waived and unable to be recovered SRA, 17 August 2026, Client confidentiality: relevant considerations.

    A good answer
    No, and the system refuses to run if it could.

    Aprentiz’s answer

    No. The models run inside the environment, with no fallback to an outside service. At every start-up it checks that it can’t reach outside addresses, and won’t start if it can. Checked on our R&D cloud build on 9 October 2026.

Security and the supplier itself.

Questions 12 to 14 of 14.

  1. Question 12: What independent security assurance do you have?

    A good answer
    Named certifications and independent tests, with dates.

    Aprentiz’s answer

    Built in: named sign-in with passkeys, the start-up check on outside connections, and signed, sealed records. Your IT provider can ask us for the detail.

  2. Question 13: Can we see what the system does today?

    A good answer
    A plain-English list of what is built and working.

    Aprentiz’s answer

    Yes. Our trust centre lists what is built and working.

  3. Question 14: What does it cost, and what are we committing to?

    A good answer
    Terms in writing before you start, and no lock-in.

    Aprentiz’s answer

    A founding research place is research and development, not a sale: we choose six firms by application, you bring a named sponsor, one workflow and your time, and there is no fee. If you want to continue after the 90 days, we agree terms together before any fee is discussed.