Auditrek vs Holistic AI: what we can prove, and what to ask them
If Holistic AI and Auditrek are both on your shortlist, the short answer is that Holistic AI's own site describes governing AI across an enterprise, with discovery, testing and runtime guardian agents, while Auditrek does one narrow thing: one SHA-256 hash-chained, append-only record per inference event, verified server-side and rendered into a report mapped to EU AI Act Article 12 and 13 fields.
Auditrek is not open yet, so anything on this page about what it does is what it is being built to do.
Why this comparison is shaped the way it is
Holistic AI and Auditrek turn up on the same shortlist because they are reached by the same search: a compliance officer at a mid-market regulated firm has been asked, by an internal auditor or a board risk committee, how the organisation would evidence what its AI systems decided and when. That single question has at least two very different software answers, and the choice between them is not a feature contest. It is a question about which gap you are actually filling.
Holistic AI's own site is the only authority on what Holistic AI builds, so the third column of the table below carries quotations from holisticai.com, retrieved on 16 November 2026, and nothing else. Where we hold no quotation on a dimension, the cell carries the question we would put to their team instead, phrased so the answer is checkable and datable. We make no claim about their product beyond those words, and none at all about their reliability, security or support: those are not ours to characterise, and a cell filled in from memory is exactly the kind of untraceable assertion we would refuse in an audit file of our own.
That is not a rhetorical device. If you are running a structured vendor evaluation, an empty cell you know is empty is more useful than a confident cell you cannot trace. Take the questions, put the same ones to both vendors, and write down the answers with the date you got them.
The category split worth understanding first
As a general observation about this market rather than a claim about any named vendor, AI governance tooling tends to divide along the unit of record. One family works at the level of the use case or the model: register the system, assess its risk, attach policies and evidence, test and monitor it, report on the portfolio. Holistic AI's homepage describes work of that shape, in its own words, "Discover every model, agent, and application. Assess AI risk, enforce policies, and bring your teams together around continuous governance." The other family works at the level of the individual inference: capture what the model was asked, what it answered, how confident it was, whether a human overrode it, and when.
Both families produce documentation. They produce documentation about different things, and an examiner asking to see the record for one specific decision on one specific date is asking the second family's question. Auditrek is built for the second. If your gap is portfolio-level assessment, policy enforcement, model testing and stakeholder evidence collection, Auditrek is not that product, and we would rather you found that out on this page than three weeks into a procurement.
What Auditrek is built to do
Capture one record per inference event
A REST and webhook API accepts inference event payloads carrying input hash, model id and version, output, confidence, timestamp and a human override flag. Each record is stored as part of a SHA-256 hash-chained, append-only sequence per ai_system, so records have an order and each one commits to the one before it.
Bind the record's own timestamp into the hash
Under hash version 2 the canonical hashed payload covers sequence number, input hash, output, confidence, human override flag, pseudonymised subject reference and the record's created_at. Editing a stored timestamp therefore breaks the recomputation rather than passing it, which is a change from the first version of the chain format.
Prove the chain rather than assert it
A server-side verification endpoint walks the chain for a record's system from the first sequence, recomputes each hash from the stored previous hash and payload, and returns a tamper status with the first broken sequence number if there is one. It is a recomputation, not a stored flag that could itself be edited.
Stay out of the production inference path
Records arrive out of band after the decision has been made, over a generic REST and webhook ingestion layer with documented payload mapping guides and worked curl examples for AWS SageMaker and Azure ML. There are no bespoke platform SDKs at launch, which is a limitation and also the reason no model migration is required.
Refuse the data you should not be sending
A data-minimisation gate at ingestion inspects payloads for recognisable raw sensitive-field patterns, such as social security number formats, ICD-10 style diagnosis codes and unredacted credit bureau fields, and rejects the request with a 422 before any database write. Subject identifiers are pseudonymised by default.
Gate connection on a recorded DPA confirmation
No API key and no webhook secret is issued for an AI system until a dpa_confirmation with a signed_at timestamp exists, enforced server-side on the connect endpoint so it cannot be stepped around from the browser. Be clear about what that is: a self-attestation by a signed-in org member giving a signer name and a document reference, with no document upload and no revocation path today.
Generate the report from the records, disclaimer attached
The report generator produces PDF and HTML output 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. Every report carries a fixed, visible footer stating that it does not constitute legal advice and has not been reviewed by qualified regulatory counsel.
Show the four numbers a compliance officer is asked for
The web dashboard shows decisions logged, systems connected, last report generated and current tamper-verification status, for a non-technical audience, with auth-gated multi-user access for an organisation so compliance, risk and engineering read the same screen.
Dimension by dimension
| Dimension | Auditrek | Holistic AI, in its own words (holisticai.com, retrieved 16 |
|---|---|---|
| Smallest unit of record | One append-only decision record per inference event: input hash, model id and version, output, confidence, timestamp, human override flag, chained per ai_system | The site says "Audit trail updated automatically" and "Full audit trail for regulators and internal review". It does not state the smallest stored object, so ask whether a row is a use case, an assess |
| Discovery across the estate | None. You register each ai_system yourself with name, platform, model id and version; Auditrek records what is sent to it and nothing else | "Discover every model, agent, and application" and "Discover AI across AWS, Azure, GitHub and Databricks" |
| Relationship to the serving path | Out of band. Events are posted over REST or webhook after the decision is made, so no model migration is required | Guardian Agents are described as observing agent actions, evaluating them against policies and intervening to block, redirect, remediate or escalate. Ask what that layer needs in or beside the serving |
| Integrity mechanism | SHA-256 hash chain per system, plus an endpoint that recomputes the chain server-side from sequence one and returns a tamper status and the first broken sequence number | Not covered by the material we quoted. Ask whether integrity is a recomputed cryptographic proof or a stored status, and who other than the vendor is able to recompute it |
| What binds the record's time | Under hash version 2, created_at is inside the canonical hashed payload, so a changed timestamp fails verification | Not covered by the material we quoted. Ask whether a record's own timestamp sits inside the integrity mechanism they describe |
| Report output | PDF and HTML for one system and one date range, mapped to EU AI Act Article 12 and 13 fields, single regulation and format set at launch | Not covered by the material we quoted. Ask which output formats exist today, who they are written for, and ask to see a sample |
| Framework coverage | One: EU AI Act Article 12 and 13 field mapping at launch | "Controls mapped to EU AI Act, NIST AI RMF and ISO/IEC 42001". How evidence attaches to a control is not covered by the material we quoted, so ask what counts as evidence and what a per-framework outp |
| Model testing and red-teaming | None, and none planned in this scope. Auditrek records decisions, it does not evaluate models | "40+ tests for bias, hallucination, privacy and robustness" and "Red team for prompt injection and jailbreaks". Ask which of those run on a schedule against production systems and which are one-off as |
| Language in the output | A fixed non-legal-advice disclaimer is embedded as a visible footer in every generated report, PDF and HTML alike | Not covered by the material we quoted. Ask what disclaimer or scope language appears on documents the platform generates |
| Sensitive fields at ingestion | Payloads with recognisable raw sensitive patterns are rejected with a 422 before any write, and subject identifiers are pseudonymised by default | Not covered by the material we quoted. Ask what happens to a payload containing an ICD-10 code or an unredacted credit bureau field |
| Contractual gate before data flows | No API key or webhook secret is issued for a system until a DPA confirmation with signed_at is recorded, enforced server-side; the confirmation is self-attested, with no upload and no revocation path | Not covered by the material we quoted. Ask what has to be signed, by whom, and whether the platform enforces it or trusts the buyer to remember |
| Published price | Starter at 1,500 euros a month for 2 connected systems and 50,000 decision records a month; Growth at 3,500 euros a month for 5 systems and 500,000 records a month; 30-day free trial, no card to start | No pricing is published on the page we retrieved, so treat it as sales-led until their team gives you a figure in writing |
| Access and hosting | Email and password sign-in for multi-user organisation access, hosted on Amazon Web Services in a single region, with Stripe, Amazon SES and Sentry as the services behind it | Not covered by the material we quoted. Ask about sign-in methods, hosting regions and the subprocessor list |
| How the deadline is framed | We do not sell on dates. The point of the records is that they can be produced on demand, whenever the question arrives | The site states "Obligations for high risk systems apply from 2 December 2027" |
| Availability today | Waitlist. Auditrek is not generally available yet, and this page describes what it is built to do | Not covered by the material we quoted. Ask for current availability and a realistic onboarding timeline |
What Auditrek will charge, published rather than quoted on request
Starter
€1,500 / month
For teams putting their first one or two AI systems under tamper-evident record keeping.
- Connected AI systems: 2
- Decision records: 50,000 / month
- Hash-chained decision records with server-side tamper verification
- Audit reports mapped to EU AI Act articles 12 and 13, in PDF and HTML
- REST and webhook ingestion with mapping guides for SageMaker and Azure ML
- 30-day free trial, no card to start
Growth
€3,500 / month
For teams running several production AI systems at higher decision volume.
- Connected AI systems: 5
- Decision records: 500,000 / month
- Everything in Starter
- Up to 5 connected AI systems
- 500,000 decision records a month
- 30-day free trial, no card to start
When to choose which
Auditrek fits when the missing artefact is the record itself
The pattern we built for is specific. You already have an AI inventory somewhere, probably in a spreadsheet or in a GRC tool the firm licensed for something else. You can describe your models. What you cannot do at short notice is produce an ordered, tamper-evident sequence of what a given system decided between two dates, with a mechanism that lets someone else check the sequence was not edited afterwards. If a supervisory question or an internal audit finding lands on that gap, portfolio-level documentation does not close it, because the question is about events rather than about policy.
Auditrek also fits when the political constraint is that nobody will let you near the production inference path. Posting events out of band over REST or webhook is a smaller ask of an engineering team than adopting a layer that expects to observe or intervene in serving, and the DPA gate plus the ingestion-side rejection of raw sensitive fields gives your privacy counterpart something concrete to review before anything is connected.
Look at Holistic AI, and platforms of that shape, when the gap is upstream of the decision
If the work in front of you is finding the AI you did not know you had, classifying risk, mapping obligations across several frameworks at once, testing models for bias or robustness, red-teaming prompts, or intervening on agent behaviour at runtime, Auditrek does none of that and we are not going to imply otherwise. Holistic AI's own site describes discovery across AWS, Azure, GitHub and Databricks, controls mapped to the EU AI Act, NIST AI RMF and ISO/IEC 42001, more than forty tests for bias, hallucination, privacy and robustness, and Guardian Agents that observe, evaluate and intervene. Evaluate that on their materials and their demo, not on our summary of them.
The two are not mutually exclusive
Assessment and event capture answer different questions, and a firm can reasonably run both: one system holding the judgement about whether a use case is high risk, another holding the ordered record of what that use case actually did. Auditrek ingests over a generic REST and webhook API, so it does not ask you to remove anything you already run. If budget stretches to only one, the deciding question is which artefact your examiner, auditor or board has actually asked for. Ask them before you ask a vendor.
What Auditrek does not have, stated plainly
- No independent security audit or certification such as SOC 2 or ISO 27001.
- No independent cryptographic review of the hash chain implementation.
- No single sign-on. Sign-in is email and password only.
- Single-region hosting on Amazon Web Services, with no multi-region replication and no choice of data residency.
- No standard data processing agreement offered by Auditrek for its own customers to sign; the in-product DPA confirmation records your own agreement with your own processor, self-attested.
If your vendor questionnaire has a hard gate on any of those, that is a legitimate reason to rule us out now rather than at legal review.
A limit worth saying out loud
No tool makes an organisation compliant, ours included, and nothing on this page should be read as a statement that a report generated by Auditrek will satisfy any supervisory authority. Auditrek is designed to support documentation of EU AI Act Article 12 and 13 record-keeping and transparency obligations by holding structured, verifiable records and rendering them into a report. Whether your practices meet an obligation is a judgement for qualified counsel, which is why the disclaimer is hardcoded into every report rather than left to a sales conversation.
How to run the evaluation so the answers are checkable
- Write down the artefact you have been asked for
One sentence, in the words the auditor, examiner or committee used. A register of AI use cases with risk classifications and the log of what the model decided on 14 March send you to different products, and most shortlists go wrong at this step rather than at the demo.
- Ask every vendor for the smallest object they store
Not the dashboard, the data model. Is the row a use case, an assessment, a model version, an agent action, or an inference event? Everything else follows from that answer, including what a report can and cannot contain.
- Ask how integrity is established, and by whom
There is a real difference between a record marked unaltered in a database and a chain recomputed on demand that returns a status. Auditrek does the second, per system, from sequence one, with the record's own timestamp inside the hashed payload. Put the same question to every vendor and note the mechanism, not the adjective.
- Put the sensitive-payload question to the engineers
Hand them a fictional payload with a diagnosis code and an unredacted applicant identifier in it and ask what the platform does with it. Silent acceptance is a finding. Auditrek rejects it at ingestion with a 422 and pseudonymises subject identifiers by default.
- Get the data protection paperwork sequence in writing
Specifically, what must be signed before any pipeline can send data, and whether the platform enforces that or trusts the buyer to remember. Auditrek will not issue an API key or webhook secret for a system until a DPA confirmation with a signed_at timestamp exists, and we will tell you in the same breath that the confirmation is self-attested.
- Ask for a sample report and read the footer
Read what the document says about its own status. A report that claims to establish compliance is a liability for the party presenting it. Auditrek's reports state on every page that they do not constitute legal advice and have not been reviewed by qualified regulatory counsel.
- Get a price in writing, next to the limits it buys
A figure without limits is not a price. Ours is published with both: 1,500 euros a month for 2 systems and 50,000 records, 3,500 euros a month for 5 systems and 500,000 records. Ask any sales-led vendor for the same two numbers, and for what happens when you exceed them.
Questions
Where does the Holistic AI column come from?
From holisticai.com, retrieved on 16 November 2026, quoted rather than paraphrased. Where that page does not speak to a dimension, the cell holds the question we would ask instead. Their site is the authority on their product, so treat anything we have not quoted as unverified and get a dated answer from their team.
Are Holistic AI and Auditrek direct competitors?
They compete for the same budget line and often the same internal sponsor, the compliance or AI governance lead. On function they overlap less than the shortlist suggests: their site describes discovery, testing, red-teaming and runtime guardian agents, while Auditrek records and verifies individual inference events and renders them into a report.
Can we run a broader governance platform and Auditrek together?
Yes, in principle, because they hold different objects: assessments, controls and policy on one side, an ordered and verifiable record of individual inference events on the other. Auditrek ingests over a generic REST and webhook API with mapping guides for SageMaker and Azure ML, so it does not require you to remove anything you already run.
Does an Auditrek report make us compliant with the EU AI Act?
No, and we will not say otherwise. The report generator produces PDF and HTML output mapped to EU AI Act Article 12 and 13 fields for a chosen system and date range, and every report carries a visible footer stating that it does not constitute legal advice and has not been reviewed by qualified regulatory counsel. Whether your practices meet an obligation is a judgement for counsel.
Do we have to move or re-deploy our models?
No. Records are posted after the decision, out of band, over REST or webhook, with documented payload mapping guides and worked examples for AWS SageMaker and Azure ML. There are no bespoke platform SDKs at launch, so an unusual stack will need a small mapping layer written against the documented payload.
What does Auditrek cost?
Starter is 1,500 euros a month for 2 connected AI systems and 50,000 decision records a month. Growth is 3,500 euros a month for 5 connected systems and 500,000 records a month. Both carry a 30-day free trial with no card to start. Billing is per organisation, tied to the number of connected systems, and runs through Stripe, which works out VAT at checkout.
Where does the data live, how do people sign in, and what is not in place?
Hosting is Amazon Web Services in a single region, with no multi-region replication and no choice of data residency. Sign-in is email and password, with no single sign-on. Subject identifiers are pseudonymised by default at ingestion and payloads with recognisable raw sensitive fields are rejected before any write. There is no SOC 2 or ISO 27001 certification and no independent cryptographic review of the chain implementation.
Can our engineers verify the chain themselves rather than trusting the dashboard?
That is the point of the verification endpoint. It recomputes each hash for a system from sequence one using the stored previous hash and canonical payload, which under hash version 2 includes the record's own created_at, and returns a tamper status plus the first broken sequence number when the recomputation fails, rather than reading back a stored flag.