Home / Comparisons

Auditrek vs Vanta: control evidence vs AI decision records

If your team already licenses Vanta and someone has asked whether the EU AI Act is just another framework to switch on, the distinction that decides it is the unit of evidence: evidence about how your programme is run, versus a record of what each AI system actually decided.

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

What each product is built around

Two tools can both appear under the heading "AI Act readiness" in a budget request and still be doing unrelated work. It is worth separating them before you compare them.

Auditrek is a decision-record layer. It accepts inference event payloads over a REST endpoint or a webhook: input hash, model id and version, output, confidence, timestamp, and a human override flag. Each payload is written as an append-only record inside a SHA-256 hash chain, one chain per ai_system, where a record's hash is computed from the previous record's hash plus its own payload. A separate verification endpoint recomputes that chain on request from sequence one and returns a tamper status of valid or broken, along with the first broken sequence number, rather than reporting a flag that was stored earlier and trusted since. From those records, Auditrek generates a PDF or HTML report for one chosen system and one date range, mapped to EU AI Act Article 12 and 13 fields, with a fixed footer on every page stating that the report does not constitute legal advice and has not been reviewed by qualified regulatory counsel. One regulation and one format set at launch.

Vanta is in this comparison as the incumbent, not as a feature list. If it is already on your licence schedule and already through your security review, it is the first place anyone looks when an AI Act line item appears, and that is a reasonable instinct. We did not re-read Vanta's own product and framework pages while writing this page, so we are not going to characterise what their framework library covers today or which AI-related frameworks are listed in it. Their site is in the sources below, and that is the right place to read the current list. So the table that follows does something slightly unusual for a comparison page: the Auditrek column states mechanisms we can point at in our own code, and the Vanta column states the question we think you should put to them directly.

Dimension by dimension

DimensionAuditrekVanta (not verified by us, ask them)
Unit of evidenceOne append-only record per inference event, holding input hash, model id and version, output, confidence, timestamp and a human override flag, hash-chained per ai_system.Not verified here. Ask what the smallest unit of evidence is in their AI-related frameworks: a control, a policy, an assessment, or an individual model decision.
Integrity checkA verification endpoint walks the chain from sequence one, recomputes each hash from the stored previous hash and payload, and returns valid or broken with the first broken sequence number.Not verified here. Ask whether evidence integrity is a stored status or recomputed on request, and what the recomputation covers.
Report outputPDF and HTML for one chosen system and date range, mapped to EU AI Act Article 12 and 13 fields, with a non-legal-advice disclaimer embedded in every report. Single regulation and format set at launchNot verified here. Ask which frameworks are listed today and what the exported artefact contains.
Where it sitsOut of band. Payloads arrive through a generic REST and webhook ingestion layer with documented mapping guides for AWS SageMaker and Azure ML. No model migration, and nothing added to the production iNot verified here. Ask what has to be installed, and whether anything touches the serving path.
Sensitive fields at ingestionPayloads carrying recognisable raw sensitive patterns are rejected with a 422 before any database write, and subject identifiers are pseudonymised by default.Not verified here. Ask how special-category data inside inference payloads is handled at the point of collection.
Gate before data flowsNo API key and no webhook secret is issued for an ai_system until a dpa_confirmation with a signed_at timestamp is recorded, enforced server side rather than in the interface.Not verified here. Ask what contractual step, if any, gates data ingestion.
Who the screens are forA dashboard for a non-technical compliance officer: decisions logged, systems connected, last report generated, tamper-verification status. Multi-user org access, sign-in with email and password. HostNot verified here. Ask who the primary user of any AI module is, and which screens an engineer has to live in.
Commercial shapePer-org billing tied to the number of connected ai_systems, across starter and growth tiers, handled through Stripe. Auditrek is on a waitlist today.Not verified here. We will not restate a price for another vendor without a quote from their own page.
What the output assertsThat a system produced these outputs, in this order, with these confidences and overrides, and that the chain recomputes. Nothing about whether your practices satisfy an obligation.Not verified here. Ask what their reports assert about your legal position, and what they explicitly do not.

The questions that usually settle it

Can you retrieve one decision?

If an internal audit or a supervisory examination asks what a model returned for one case on one date, a control register answers who owned which policy, not what happened at inference time. Auditrek's records are per event, carrying model id and version, output, confidence and the human override flag, so the answer is retrieved rather than reconstructed from screenshots and exports.

Is integrity asserted, or recomputed?

A stored verification column is a statement someone made in the past. The Auditrek endpoint recomputes each hash in the chain from the stored previous hash and the record payload at the moment you ask, and names the first sequence number where the recomputation disagrees. That difference matters most in the one conversation where it is questioned.

Who has to change their pipeline?

Auditrek receives payloads out of band over REST or a webhook, with documented mapping guides for AWS SageMaker and Azure ML, so the serving path stays as it is. If the answer to this question is "the ML team reworks inference", the project stops being a compliance decision and becomes an engineering negotiation.

Which obligation lands this quarter?

Management-system work and per-inference record-keeping are different deliverables. If your near-term milestone is an audit against a standard, a broad framework platform is the natural home for it. If the milestone is producing per-decision records for a high-risk system, that is the shape Auditrek captures and the shape the Article 12 and 13 report fields are mapped to.

What happens to the sensitive fields?

Credit, hiring and clinical pipelines carry data well above the ordinary baseline for a web application. Auditrek's ingestion rejects recognisable raw sensitive patterns with a 422 before writing anything, pseudonymises subject identifiers by default, and refuses to issue credentials for a system until a signed DPA confirmation exists. Ask any vendor in this category the same three questions.

When to choose which

Stay with the platform you already license when the work is management-system shaped

If your next deliverable is an AI use-case inventory, named policy owners, control evidence gathered from technical and business stakeholders, and an audit against a standard your team has already chosen, then a platform your security review has cleared and your auditor recognises is a genuine advantage. A second vendor means a second procurement cycle, a second review, and a second set of access decisions. Procurement friction is a legitimate deciding factor in the mid-market, not an excuse, and buying a specialist tool for a problem that is not on your plan for this year is a cost without a return.

Choose Auditrek when the question has a date and a case number in it

The moment Auditrek is built for is narrower and harder to fake: someone asks what a specific system output on a specific day, which model version was serving, what the confidence was, and whether a human overrode it. That evidence either exists as records captured at the time or it does not exist at all, and it cannot be assembled retrospectively in a credible way. If your high-risk systems are already in production and nothing is currently writing per-decision records, the gap is not a policy gap, and a policy tool will not close it.

Run both when the programme is split, which is common

Nothing about Auditrek asks you to move the rest of your programme. Control and policy work stays where your team already does it, and inference records go into hash-chained storage built for that purpose, with reports produced per system and per date range when you need them. Auditrek connects per ai_system and bills per connected system across starter and growth tiers, so the scope of the overlap is something you set deliberately rather than by licence.

If you are earlier than the comparison stage and still scoping what evidence you actually owe, the use-case page for compliance officers who have to produce the evidence covers that ground, and the dashboard page covers what the day-to-day view looks like.

What neither product does

Neither Auditrek nor any other vendor in this category can tell you that your documentation will satisfy a national supervisory authority. That determination is made by people with a mandate we do not have, reading the regulation text and the guidance that applies to your system class.

So we have drawn the line inside the product rather than only in the copy. Every generated report, in both the PDF and HTML render paths, carries a fixed visible footer stating that it does not constitute legal advice and has not been reviewed by qualified regulatory counsel. Report templates stay descriptive and structural: they record what a system did and when, not what you are required to do. Our build includes a check that fails if certain absolute compliance claims appear in the product's own source or report templates, which is a blunt instrument, and that is the point.

What Auditrek will claim is mechanical, and you can test each part in an evaluation: records are appended in a SHA-256 chain per system, the chain is recomputed server side on request and returns a tamper status with the first broken sequence, reports are mapped to Article 12 and 13 fields for a chosen system and date range, and none of it requires a change to how your models serve traffic. Read the regulation text yourself at EUR-Lex, then ask both vendors to demonstrate against the specific obligation your legal team has written down.

Questions

Can the compliance automation platform we already have cover EU AI Act record-keeping on its own?

We cannot answer that for Vanta, because we did not verify their current framework list while writing this page. The question to put to them is concrete: which frameworks are in the library today, what the smallest unit of evidence is, and whether anything captures individual inference events with model version, confidence and human override. If the answer is control and policy evidence, that is a different artefact from a per-decision record.

Does Auditrek replace our SOC 2 or ISO 27001 tooling?

No. Auditrek captures inference events as hash-chained decision records per ai_system and generates reports mapped to EU AI Act Article 12 and 13 fields. It does not manage control sets, policy ownership or certification workflows, and it is designed to sit alongside whatever you already use for those.

Does Auditrek sit in our production inference path?

No. Records arrive out of band through a generic REST endpoint or a webhook, with documented payload mapping guides for AWS SageMaker and Azure ML. There is no model migration, and there are no bespoke platform SDKs in the current scope.

What exactly does a generated report contain?

PDF or HTML output for one chosen system and one date range, mapped to EU AI Act Article 12 and 13 fields, with a visible disclaimer on every page stating that the report does not constitute legal advice and has not been reviewed by qualified regulatory counsel. One regulation and one format set at launch.

How would we know if a record had been altered?

The verification endpoint walks the chain for that record's system from sequence one, recomputes each hash from the stored previous hash and payload, and returns a tamper status of valid or broken, naming the first broken sequence number. It is a recomputation at request time, not a stored flag being read back.

Our inference payloads contain sensitive fields. What happens at ingestion?

Ingestion inspects payloads for recognisable raw sensitive-field patterns and rejects the request with a 422 before any database write, and subject identifiers are pseudonymised by default. Separately, no API key or webhook secret is issued for a system until a DPA confirmation with a signed_at timestamp is recorded, and that check is enforced server side.

How is Auditrek priced against a platform licence?

Auditrek bills per org, tied to the number of connected ai_systems, across starter and growth tiers, handled through Stripe. We will not restate another vendor's pricing without a quote from their own page, so ask them for a current figure and compare it against the number of systems you actually need records for.

Who in our team uses which tool day to day?

Auditrek's dashboard is written for a non-technical compliance officer: decisions logged, systems connected, last report generated, tamper-verification status. Engineers touch the ingestion payload and the mapping guide once per system. Sign-in is by email and password, and org access is multi-user, which the workspace page in this section covers in more detail.

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.