Skip to main content
Curvestone AI
How to

How to validate an AI checking model before you rely on it

Updated
  1. Why a vendor's own test is not enough

    Vendors often share accuracy figures, test reports or an outside review of their AI. Treat these as the vendor's evidence about the vendor's test set. They do not show how the check performs on your case types, your checklist and your document quality.

    The FCA's AI paper of 19 October 2026 is direct on this point. Firms "can outsource AI services, but remain responsible for them". It also says firms "should be able to show that the controls they choose work in practice". A report you did not design is hard to use as that proof.

    The paper also warns against the opposite error: leaning on human review to cover for a model nobody tested. It lists "excessive reliance on human-in-the-loop controls" as an area for improvement and notes that human review "may mask weaknesses in the model". Validation is how you find those weaknesses before your reviewers start to trust the output.

  2. Six steps to validate an AI checking model

    1. Choose the files. Take a set of completed files your team has already closed. Include the case types you see most, the ones that fail most, and a few unusual ones. Do not let the vendor pick them.
    2. Build the answer key blind. Ask experienced reviewers to check each file against your own checklist, without access to the AI's result. The FCA describes a firm doing exactly this for affordability decisions, so that it could "compare independent human and AI judgements". Where two reviewers disagree, settle it and record why. Disagreement between people tells you how hard the question is.
    3. Run the AI check on the same files. Use the configuration you plan to use live: the same checklist, the same document formats, the same route in.
    4. Sort every disagreement by error type. Count the issues the AI missed that your reviewers found, and the issues the AI flagged that were not issues. A missed issue on a suitability or affordability point matters more than a false flag on a formatting point, so record which checklist question each error sits under.
    5. Decide where people stay. Use the error pattern to set the line. Questions where the AI and your reviewers agree closely can move to review by exception. Questions where it misses things stay with a person, or need a change to the checklist or the configuration before you test again.
    6. Sign off and keep the record. Write down the files used, the answer key, the results by question, the decisions you took and who approved them. The FCA points to "structured challenge, testing and sign-off before deployment", with control functions involved.

    Smaller firms can scale this down. The FCA notes that smaller firms used "carefully filtered high-quality datasets that had been reviewed by people" to give known good outputs for testing. A few dozen well-chosen files that your best reviewer has checked show more than a large sample nobody has read.

  3. What independence means in practice

    Independence is about who builds the answer key and who reads the results. The people who chose the tool, or who are measured on its rollout, should not be the only people who judge whether it works. In a network or a lender, that often means compliance or quality assurance runs the test and risk or a governance committee reviews it.

    PRA SS1/23, the Bank of England's model risk supervisory statement, sets out "Principle 4 – Independent model validation". Its principles apply to models "developed in-house or externally (including vendor models)". Its formal scope is banks, building societies and PRA-designated investment firms with internal model approval for regulatory capital, so most brokers and networks are outside it. The idea still carries over: the test should not be run only by the people who want the answer to be yes.

  4. Validate again when something changes

    A validation result describes one model, one checklist and one set of files. Each of those can change. The FCA notes that some firms used contracts to obtain "notice of, and further information about, significant model changes" so they could test updates before go-live. It also lists "periodic performance reviews" among stronger vendor oversight practices for higher-impact AI.

    Three triggers are worth writing into your procedure:

    • The vendor tells you the underlying model or its configuration has changed.
    • You change your own checklist or add a new product type.
    • Live monitoring shows a rise in overrides on one checklist question.

    After go-live, keep a sample of files under blind human review. The FCA describes firms "assessing their model's performance against human-led processes" to see whether they need more checks. That is step 2 again, at a smaller scale.

    Curvestone checks every case against a firm's own compliance criteria and flags gaps with evidence references, which gives your reviewers a reason to test against for each result. For the wider governance picture, see model risk management for AI and what the FCA's AI paper means for lenders.

Questions

Frequently asked questions

Can we rely on a vendor's third-party validation report?
You can use it as an input to your due diligence. It does not replace your own test. The FCA says firms remain responsible for AI they outsource and should be able to show their controls work in practice. A report on the vendor's test set does not show how the check performs on your files and checklist.
How many files do we need to validate an AI check?
There is no FCA number. Choose enough files to cover your main case types and the questions that fail most often, and have experienced reviewers check them blind. The FCA notes that smaller firms tested with smaller, human-reviewed datasets. A well-chosen set your best reviewers have checked is worth more than a large unreviewed one.
Who should sign off the validation?
Someone who did not choose the tool and is not measured on its rollout. In practice that is often compliance or quality assurance running the test, with risk or a governance committee approving the result. The FCA points to structured challenge, testing and sign-off before deployment, with control functions involved and accountable senior managers named.
Sources
  1. 01FCA: AI in financial services: good practice and areas for improvement (19 October 2026)
  2. 02PRA: SS1/23 Model risk management principles for banks
Related reading
Dawid Kotur
Written by

Dawid Kotur

CEO and co-founder, Curvestone

Dawid co-founded Curvestone in 2024 after a decade working at the intersection of financial services and applied machine learning. He writes about the strategic direction of regulated-industry AI, the FCA's evolving approach to model risk, and the operational changes UK lenders are making in response to Consumer Duty. He sits on the FCA Smart Data Accelerator advisory cohort.

LinkedIn

Compliance that thinksahead. Automatically.

Join mortgage networks, lenders, and legal firms using Curvestone to review cases at scale.