Auditrek vs Saidot: inference records and governance workflow
If you are choosing between Auditrek and Saidot, the two names appear for the same searches and compete for the same budget line, but they answer different questions about your AI systems.
Auditrek is not open yet, so anything on this page about what it does is what it is being built to do.
Two different units of record
You searched for a head-to-head, so here is the short version before the detail.
Auditrek works at the level of a single decision. It accepts inference event payloads over a REST and webhook API, each one carrying an input hash, model id and version, output, confidence, timestamp and a human override flag, and stores them as SHA-256 hash-chained append-only records per ai_system. From those records it generates PDF and HTML reports for a chosen system and date range, mapped to EU AI Act Article 12 and 13 fields, with a single regulation and format set at launch. The design goal is narrow and deliberate: sit out of band, take events your pipeline already emits, and never touch the production inference path or require a model migration.
Saidot is described in our market research as a European, Finland-based AI governance platform positioned explicitly around EU AI Act documentation and AI transparency for enterprises and public sector bodies. That is a positioning description and nothing more. Our research says nothing about their market presence, customer base or product maturity, so this page makes no claim on any of those, and neither should your shortlist until you have it from them.
Why we hedge the second half of that sentence
Our research file names Saidot's own site as the source of record for every capability statement, and notes that the wording was not captured verbatim in that run. We are not going to reconstruct a competitor's feature list from memory and set it in a table as fact next to our own build scope. So the Saidot column below contains two kinds of entry: what our research actually states about their positioning and pricing visibility, and, where it states nothing, a direct instruction to confirm it with them. If you are running a procurement process, that is the column you should be filling in from their materials anyway, not from ours.
Dimension by dimension
| Dimension | Auditrek | Saidot |
|---|---|---|
| Unit of record | One append-only record per inference event: input hash, model id and version, output, confidence, timestamp, human override flag | Not captured verbatim in our research; ask whether the record is per system, per assessment or per inference |
| How events arrive | Generic REST and webhook ingestion, with documented payload mapping guides for AWS SageMaker and Azure ML; no bespoke platform SDKs in the first release | Not captured verbatim in our research; confirm the ingestion and integration model on their site |
| Tamper evidence | SHA-256 hash chain per ai_system, plus a server-side endpoint that walks and recomputes the chain and returns a tamper status rather than reading a stored flag | Not captured verbatim in our research; ask how record integrity is demonstrated to a reviewer |
| Report output | PDF and HTML for one system and date range, mapped to EU AI Act Article 12 and 13 fields, single regulation and format set at launch | Our research describes a platform positioned around EU AI Act documentation and transparency; confirm the exact outputs and framework coverage with them |
| Language on the output itself | Every generated report carries a fixed footer stating that it does not constitute legal advice and has not been reviewed by qualified regulatory counsel | Not captured in our research; ask what their generated documentation says about legal review |
| Sensitive fields at ingestion | Payloads with recognisable raw sensitive field patterns are rejected with a 422 before any database write, and subject identifiers are pseudonymised by default | Not captured in our research; ask how special-category personal data is handled at intake |
| Contracting precondition | No API key or webhook secret is issued for an ai_system until a DPA confirmation with a signed_at timestamp is recorded, enforced server-side | Not captured in our research; ask about their data processing terms and when they are signed |
| Who operates it day to day | A compliance officer, on a dashboard showing decisions logged, systems connected, last report generated and tamper-verification status | Our research places them with the same compliance and AI governance buyer; confirm the intended operator with them |
| Pricing basis | Per organisation, billed on connected ai_system count through Stripe, reflecting starter and growth tiers | No public price list found in our research; treat as contact sales until they confirm otherwise |
| Sign-in and hosting | Email and password sign-in with multi-user org access, hosted on Amazon Web Services | Not captured in our research; ask about sign-in options, hosting and data residency |
| Availability | Waitlist today, so everything above is declared build scope rather than a launched track record | Not captured in our research; confirm current product status and availability with them |
What Auditrek does not do
Fairness cuts both ways, and the shorter list is ours.
Auditrek has no AI use-case registry, no policy pack library, no obligation-mapping workspace and no questionnaire workflow for collecting attestations from business stakeholders. It does not assess your models, test for bias or robustness, and it does not tell you whether a given system falls inside the high-risk category. It holds one regulation and one report format at launch, and its integration surface is a documented payload mapping rather than a native platform SDK.
It also does not judge your practices. The reports are structural: they record what a system did, when, at what confidence, and whether a human overrode the output. They do not assert that your programme satisfies any obligation, and the embedded footer says so in plain language. If what you need is an interpretation of the regulation, that comes from your counsel, not from a PDF.
When each one fits
The honest way to split this is by which sentence you cannot currently finish.
Choose a governance platform of the kind our research attributes to Saidot when the gap is breadth
If you do not yet know how many AI systems your organisation runs, who owns them, which are candidates for the high-risk category, and what documentation each one owes, then per-decision logging is premature. You need an inventory, an obligation view across a portfolio, and a way to pull structured answers out of product teams. That is portfolio work, it involves people more than pipelines, and a platform built around registry and documentation workflow is the right shape for it. Verify the specifics on their own pages before you shortlist.
Choose Auditrek when the gap is depth on a named system
If the inventory already exists, often in a spreadsheet, and the unanswered question is what a specific model decided for a specific case on a specific date, Auditrek is built for that. A supervisory examination, a customer complaint, an internal audit sample or a disputed adverse outcome all end at the same place: a request for the record of one decision, with the model version that produced it, and a reason to believe the record was not edited afterwards. That is why verification recomputes the hash chain server-side instead of trusting a stored flag, and why records are append-only per ai_system.
They are not mutually exclusive
The two questions sit at different altitudes. A registry answers which systems exist and what they owe. An inference-level ledger answers what system seventeen output at 09:12 on 14 March and whether a human overrode it. Mid-market teams often solve the first with a tool they already license and the second not at all, which is the gap Auditrek was designed around. If you are documenting the compliance-officer side of that work, our page for the compliance officer who has to produce the evidence covers the workflow, and the compliance dashboard page covers what the operator sees.
One trade-off worth naming
Inference-level capture means data flows from your pipeline into a vendor system, and in credit, hiring and clinical contexts those payloads can carry special-category personal data. Auditrek's answer is to refuse it at the door: recognisable raw sensitive patterns are rejected with a 422 before any write, subject identifiers are pseudonymised by default, and no credentials are issued for a system until a DPA confirmation is on record. Ask every vendor on your list, us included, to describe its equivalent controls at the level of the mechanism, and keep the written answers with your processing record.
Five questions to put to both vendors
- What exactly is one row in your system
Ask each vendor to describe a single stored record field by field. If the answer is a system profile or an assessment rather than an inference event, you are buying portfolio documentation, not decision evidence, and you should know which one you are short of.
- How does a reviewer test that a record was not altered
Ask for the mechanism, not the adjective. Auditrek's answer is a per-system SHA-256 chain and an endpoint that walks it from sequence one and recomputes each hash, returning a tamper status and the first broken sequence number. Ask for the equivalent walkthrough elsewhere.
- What does the generated document say about legal review
Read the footer of a sample output. Auditrek hardcodes a disclaimer on every page of both the PDF and HTML render paths stating that the report is not legal advice and has not been reviewed by qualified regulatory counsel. Ask to see the equivalent wording on any sample you are shown.
- What happens to a payload containing a raw diagnosis code or bureau field
Send the awkward test case in a pilot. Auditrek rejects it at ingestion with a 422 before any database write. Whatever the other answer is, get it in writing, because it becomes your processing record.
- What has to be signed before anything connects
Auditrek will not issue an API key or webhook secret for a system until a DPA confirmation with a signed_at timestamp exists, and the check is enforced server-side. Confirm the sequencing with any vendor before your pipeline sends its first event.
Questions
Is Auditrek a replacement for Saidot?
Not in the general case. Auditrek covers one layer, the per-inference record and the report drawn from it, and has no use-case registry, policy packs or obligation-mapping workspace. If your programme needs portfolio documentation across many systems and stakeholders, that is a different job from the one Auditrek is built for, and the two can sit side by side.
Why does this page not list Saidot's features in detail?
Because our research records their own site as the source for every capability statement and notes that the wording was not captured verbatim. We will not print an unverified feature list for a competitor. The Saidot column tells you what to confirm on saidot.ai, which is where a procurement file should get it from anyway.
Which EU AI Act articles does Auditrek's report generator map to?
The generator produces PDF and HTML output for one system and a chosen date range, mapped to EU AI Act Article 12 and 13 fields, being record-keeping and logging, and transparency and provision of information to deployers. One regulation and one format set at launch. The output is descriptive, records what happened, and states on every page that it is not legal advice.
Do I have to move or change my models?
No. Ingestion is a generic REST and webhook API with documented payload mapping guides for AWS SageMaker and Azure ML, so events are sent out of band and the production inference path is untouched. There are no bespoke platform SDKs in the first release.
How is Auditrek priced, and how do people sign in?
Billing is per organisation, tied to the number of connected ai_systems and handled through Stripe, reflecting starter and growth tiers. Sign-in is email and password, with auth-gated multi-user org access. The product is hosted on Amazon Web Services.
What stops sensitive data ending up in the ledger?
The ingestion endpoint inspects incoming payloads for recognisable raw sensitive field patterns and rejects them 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.
Can I prove to an examiner that a record was not edited later?
Records are append-only and hash-chained per ai_system with SHA-256, and verification is server-side: the endpoint walks the chain from sequence one to the record in question, recomputes each hash from the stored previous hash and payload, and returns a valid or broken status with the first broken sequence number. It does not simply read back a stored flag.
Can I try it today?
Auditrek is in waitlist mode, so there is no self-serve trial yet. The capabilities on this page are the declared build scope. If you want to be in the first group connecting a system, join the waitlist and say which platform serves your inference traffic.