Home / Industries

Model risk evidence for AI decisions in financial services

If you run model risk in a bank, a lender or an insurer, the thin spot is rarely the validation report, it is the per-decision evidence behind the model once it is serving live traffic.

Auditrek is not open yet, so anything on this page about what it does is what it is being built to do.

Model risk management in financial services documents the model, not the decision

Most mid-market banks, lenders and insurers already run a recognisable model risk function: an inventory of models with owners and tiers, a development file per model, an independent validation cycle, ongoing performance monitoring and a challenger process. The supervisory guidance that shaped that work, SR 11-7, asks for documentation robust enough that someone outside the development team can follow what happened.

That machinery documents the model. It does not, on its own, document the decisions. When a validator or a supervisor asks what the scorecard returned for a specific application on a specific afternoon, which model version was serving at that moment, what confidence the model reported and whether an underwriter or an adjudicator overrode it, the answer usually has to be reassembled: application logs in a platform tool with a retention window someone else set, feature values in the warehouse, override reasons in the case management system, the model version in a registry. Nobody wrote a single record at the moment of the decision, so the evidence is a reconstruction, and a reconstruction is the thing that is hardest to defend in a room full of examiners. The gap is widely shared: in Kiteworks' 2026 annual survey of 459 security, compliance and technology professionals, only 33 percent of organisations had tamper-evident audit trails in place, and half could not produce a complete AI access record within one business day.

In Europe the same gap now has an article number against it, since Annex III of the EU AI Act places creditworthiness evaluation and credit scoring, and risk assessment and pricing in life and health insurance, in the high-risk category, and Article 12 requires that such systems technically allow the automatic recording of events over their lifetime.

The decision in front of you

The usual responses are each expensive in a different way. You can extend the data science platform to emit richer logs, which means change control on the serving path and a release the ML team has to own. You can buy a wide governance suite and absorb a policy module, a risk register and a workflow engine you did not ask for. Or you can keep the spreadsheet and hope the examination window is short.

Auditrek takes the narrow position deliberately. It is not model risk management software and it does not replace your framework: it is the cryptographic audit-trail layer that sits underneath one. It accepts the event after your model has already returned its output, writes it into a per-model chain that cannot be quietly edited, and turns a date range of those records into a report.

From inference event to a report for one model

  1. Register each model as its own system

    Systems are registered with a name, platform, model id and version, and records are kept per ai_system, so your PD scorecard, your fraud model and your pricing model each carry a separate append-only chain rather than sharing one undifferentiated log. That maps onto the model inventory your second line already maintains.

  2. Record the data processing agreement before anything connects

    A system cannot be connected until a DPA confirmation with a signer name, a document reference and a signed_at date is recorded against it. Until that exists, the server refuses to issue an API key or a webhook secret, and the refusal is enforced server-side rather than hidden in the interface.

  3. Post the inference event

    Capture runs over a REST and webhook API that accepts an inference event payload: input hash, model id and version, output, confidence, timestamp, and a human override flag. Because the event is posted rather than intercepted, the model, the feature pipeline and the serving stack stay exactly where they are. A published API documentation page sets out the payload fields, the required headers and a worked field-mapping example for AWS SageMaker and Azure ML inference logs.

  4. Let the chain link the record

    On write, the sequence number and previous hash are taken from the last record for that system, the new hash is computed over the previous hash and the payload, and the whole record is stored in a single transaction as an append-only entry. The evidence is the hash relationship between records, not a mutable flag on a row.

  5. Verify the chain on the server, record by record

    A verification endpoint walks the chain for that system from sequence one to the record you asked about, recomputes each hash from the stored previous hash and payload, and returns a tamper status of valid or broken together with the first broken sequence number. The decisions table also carries a per-record verify action with the result shown inline, so the check is something a reviewer can run in front of you.

  6. Watch the position from the dashboard

    The web dashboard shows decisions logged, systems connected, when the last report was generated, and tamper-verification status, written for a compliance officer to read directly rather than for an engineer to interpret on their behalf. It also carries a permanent notice that data is stored in a single region.

  7. Generate the report for one model and one window

    Pick one system and a date range and the generator produces PDF and HTML output mapped to EU AI Act Article 12 and 13 fields, the record-keeping and logging fields and the transparency and information fields, with one regulation and one format set at launch. Every report carries an embedded disclaimer stating that it does not constitute legal advice and has not been reviewed by qualified regulatory counsel.

What each decision record carries

Field in the payloadWhat it holdsWhy a model risk team asks for it
Input hashA hash of the input, not the raw inputTies the record to a specific request without pulling applicant or bureau data into the ledger
Model id and versionThe identifier and version that served the decisionAnswers which artefact was live, the question a validator asks first
OutputWhat the model returnedThe decision itself, recorded at the moment it was made rather than reconstructed later
ConfidenceThe confidence value reported with the outputSupports review of borderline cases and of where the model was least certain
TimestampWhen the event was recordedLets a report be scoped to an examination window or a monitoring period
Human override flagWhether a person overrode the model outputEvidence of human oversight in practice, counted and located rather than asserted
Subject referenceA pseudonymised reference to the subject of the decisionLets a case be followed through the record set without storing a raw applicant identifier

What matters in a bank, a lender or an insurer

The chain is per model, like your inventory

Records sit in a separate append-only chain for each registered system, so a review of one scorecard does not drag in every other model's traffic, and a report for one system is genuinely a report for one system.

Overrides stop being anecdotal

The human override flag is a first-class field on every record, so the question of how often adjudicators departed from the model in a given quarter is answered from the same ledger as the decisions themselves rather than from a separate spreadsheet.

Raw sensitive fields are refused at the door

The ingestion endpoint inspects each payload for recognisable raw sensitive-field patterns, including national identifier formats, diagnosis code patterns and unredacted credit bureau fields, and rejects the request with an error before any database write. Subject identifiers are pseudonymised by default. Raw model inputs never arrive at all, only their hash.

No credentials until the DPA is recorded

The gate is structural rather than procedural: no API key and no webhook secret exist for a system until a DPA confirmation with a signed_at date is on file, and the check is enforced on the server so client state cannot route around it.

Tamper status is recomputed, not stored

Verification walks the chain and rebuilds each hash on the server, returning valid or broken with the first broken sequence number, rather than reading back a status column that an edit could have set. The state of your evidence is something you can look at before an examination rather than discover during one.

One artefact for two audiences

The generator emits PDF and HTML for the same chosen system and date range, so the version that goes into a board pack and the version an examiner reads are the same content rendered twice, not two documents that can drift apart between quarters.

Your ML team's change control stays closed

Capture is a REST call or a webhook carrying an already-completed inference event, with documented payload mapping guides for SageMaker and Azure ML. That is a smaller conversation with the platform owner than a change to the production inference path, which matters when the deadline is nearer than the next release train.

Language your filings already use

Reports are mapped to the Article 12 record-keeping and logging fields and the Article 13 transparency fields, and they stay descriptive: what the system did, when, with what confidence, and whether a person overrode it. They record facts and do not assess your obligations.

What Auditrek does not do, stated plainly

Evidence-first cuts both ways, so here are the edges before you spend a procurement cycle on us.

  • It does not validate models. There is no backtesting, no discriminatory power analysis, no challenger comparison and no model inventory workflow beyond registering the systems that are being logged. Your validation function keeps its work; Auditrek is built to hold the decision evidence underneath it.
  • It does not tell you whether you meet an obligation. Reports are descriptive and structural. Every generated report embeds the line that it does not constitute legal advice and has not been reviewed by qualified regulatory counsel. What the output means for your supervisory position is a question for your counsel.
  • One regulation and one format set at launch. The report generator ships mapped to EU AI Act Article 12 and 13 fields. If your near-term need is an SR 11-7 shaped pack, the underlying decision records are the same records, but that report template is not in the launch scope.
  • Single region, and no residency choice. The service is hosted on Amazon Web Services in a single region. There is no multi-region replication and no choice of data residency, and the dashboard says so permanently rather than in a footnote.
  • The basics are the basics. People sign in with email and password, access is auth-gated and scoped to your organisation, and billing runs through Stripe against the number of connected systems on your plan.

Questions

Is Auditrek model risk management software?

No. It is the audit-trail layer inside a model risk programme: capture of inference events over a REST and webhook API, SHA-256 hash-chained append-only decision records per system, server-side chain verification, a compliance dashboard and a report generator for one system and one date range. It holds no model inventory workflow, no validation tooling and no policy library, and it sits beside your existing model risk framework rather than replacing it.

Does it satisfy SR 11-7 or produce an SR 11-7 pack?

The report generator ships with one regulation and one format set at launch, mapped to EU AI Act Article 12 and 13 fields. The decision records underneath are the same records whatever framework you are answering to, but there is no SR 11-7 shaped template in the launch scope, and no report from Auditrek asserts that your practices satisfy any obligation. Every report carries the embedded line that it does not constitute legal advice and has not been reviewed by qualified regulatory counsel.

Do we have to change our serving path to use it?

No. Auditrek accepts an inference event that your system has already completed, posted as a REST call or delivered to a webhook endpoint, carrying the input hash, model id and version, output, confidence, timestamp and human override flag. Nothing sits in the inference path and no model migration is involved, which is usually a shorter conversation with the platform owner than a change to production serving.

How does it connect to SageMaker or Azure ML?

Through the generic REST and webhook ingestion layer, with documented payload mapping guides and a worked field-mapping example for AWS SageMaker and Azure ML inference logs on the API documentation page. There are no bespoke platform SDKs in the launch scope, so the integration is whatever already emits or forwards your inference events.

Can raw applicant or bureau data end up in the ledger?

The ingestion endpoint is built to refuse it. Payloads are inspected for recognisable raw sensitive-field patterns, including national identifier formats and unredacted credit bureau fields, and rejected with an error before any database write, and subject identifiers are pseudonymised by default. What is stored for each decision is the input hash rather than the input itself, the output, confidence, the override flag and a pseudonymised subject reference.

Who can connect a system, and what has to happen first?

A system cannot be connected until a DPA confirmation with a signer name, a document reference and a signed_at date is recorded against it; before that, the server returns a refusal rather than credentials. Access to the organisation's systems, decision records and reports is auth-gated, and people sign in with email and password, which is the only sign-in method.

What does it cost, and how is it metered?

Billing is per organisation and tied to the number of connected AI systems. Starter is €1,500 a month for two connected systems and 50,000 decision records a month; Growth is €3,500 a month for five connected systems and 500,000 decision records a month, with everything in Starter included. Both carry a 30-day free trial with no card to start.

Where is the data held, and what is not in place?

The application and its database are hosted on Amazon Web Services in a single region, with no multi-region replication and no choice of data residency. Decision records are append-only and kept for the life of the organisation that owns them. There is no independent security certification such as SOC 2 or ISO 27001, no single sign-on, and no independent cryptographic review of the hash chain implementation; we would rather you knew that from this page than from a questionnaire.

Sources

Join the waitlist

Leave your work email and we will send the payload schema, the verification endpoint documentation and a sample Article 12 and 13 report before you talk to anyone. Pilot cohort conversations start with your own pipeline, not a demo script.

One email when it launches, and one now to confirm your address. Unsubscribe any time. Privacy.