Auditrek vs IBM watsonx.governance and OpenPages
You are comparing Auditrek, a cryptographic audit-trail layer for AI inference events, with IBM watsonx.governance and IBM OpenPages. This page sets out exactly what Auditrek records and how events reach it, why we do not describe IBM's products here, and how to decide between them without taking either vendor's word for anything.
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 to do
If you are running this comparison, you probably have two names on a shortlist that are nothing like the same size, and the real question underneath is whether a narrow tool covers the specific gap you have been asked to close.
Auditrek is a record layer, and nothing wider than that. It accepts inference events over a REST endpoint or a webhook: input hash, model id and version, output, confidence, timestamp and a human override flag. Each accepted event is written as an append-only record inside a SHA-256 hash chain scoped to a single ai_system, so the decisions a model made sit in sequence and each link is computed from the one before it. Ingestion is deliberately generic, with documented payload mapping guides for AWS SageMaker and Azure ML rather than bespoke platform SDKs, which makes the integration work a payload mapping exercise for whoever owns your serving code rather than a model migration. Deep native integrations are fast-follow work, not something shipped at this stage.
IBM publishes its own pages for watsonx.governance and for OpenPages, and both are linked in the sources below. We did not capture verbatim wording from those pages in this research pass, so this page does not describe what either product does, what it costs, or how it performs. Where the third column below is empty of detail, that is not a judgement, it is an absence of quoted evidence, and the right fix is to read IBM's pages and ask IBM directly.
That absence shapes how to use this page. What we can set out precisely is one side of the comparison: what Auditrek records, how events reach it, what the verification endpoint recomputes, what a generated report contains, and, just as usefully, what the product does not do. Put that column next to whatever you obtain in writing from IBM, and measure both against the gap you actually have. If that gap is the ability to show, record by record, what a model output was, when it happened, which model version produced it and whether a human overrode it, then you are looking at a record-capture problem. If it is broader than that, you are looking at a different shape of problem, and the Auditrek column will tell you so quickly.
Dimension by dimension, with the evidence we actually have
| Dimension | Auditrek | IBM watsonx.governance and IBM OpenPages |
|---|---|---|
| What gets recorded | Inference events posted to a REST or webhook API: input hash, model id and version, output, confidence, timestamp, human override flag. Stored as SHA-256 hash-chained append-only records per ai_system | Not quoted in our research; see IBM's own product pages in the sources. |
| How events get in | Generic REST and webhook ingestion layer with documented payload mapping guides for AWS SageMaker and Azure ML. No bespoke platform SDKs at this stage; native integrations are fast-follow, not build-t | Not quoted in our research. Ask IBM for the ingestion or connector contract for your serving stack. |
| Tamper evidence | A server-side verification endpoint walks the chain for an ai_system, recomputes each hash from the stored previous hash and payload, and returns a tamper status with the first broken sequence number, | Not quoted in our research; see IBM's own product pages in the sources. |
| Report output | PDF and HTML reports for one chosen system and date range, with fields mapped to EU AI Act Article 12 and Article 13 fields. One regulation and one format set at launch. | Not quoted in our research; see IBM's own product pages in the sources. |
| Disclaimer in output | A fixed non-legal-advice line is hardcoded into every generated report, stating the report does not constitute legal advice and has not been reviewed by qualified regulatory counsel. | Not quoted in our research; see IBM's own product pages in the sources. |
| Controls before a pipeline is connected | An ai_system gets no API key and no webhook secret until a DPA confirmation with a signed_at timestamp is recorded, enforced server side. Ingestion rejects payloads carrying recognisable raw sensitive | Not quoted in our research; see IBM's own product pages in the sources. |
| Who uses it day to day | A web dashboard aimed at a non-technical compliance officer: decisions logged, systems connected, last report generated, tamper-verification status. Auth-gated multi-user org access. | Not quoted in our research; see IBM's own product pages in the sources. |
| Scope | One layer: per-inference records and the reports drawn from them. Not a control library, not a risk register, not an enterprise GRC platform. | Not quoted in our research; see IBM's own product pages in the sources. |
| Commercial model | Per-org billing tied to the number of connected ai_system records, with starter and growth tiers, billed through Stripe. No figures published on this page. | We publish no figure we have not read on IBM's own pricing material. Check IBM's pricing page and your procurement channel. |
| Hosting and sign-in | Hosted on Amazon Web Services. People sign in with email and password. | Not quoted in our research; see IBM's own product pages in the sources. |
| Availability today | Waitlist. Not generally available, so treat everything above as what the product is built to do, not as a field-tested track record. | Not quoted in our research; see IBM's own product pages in the sources. |
How to run this evaluation without taking either vendor's word for it
- Write down the obligation you are being asked to evidence
Not the framework, the obligation. Article 12 concerns record-keeping and logging, Article 13 concerns transparency and information provided to deployers, and the consolidated text is linked in the sources. If the question in your next internal review is whether you can reconstruct individual decisions, the shape of the answer is record capture. If the question is whether you have an owner, a control and a review cadence for each AI use case, the shape of the answer is something wider than a record layer.
- Inventory where the events already exist
Most teams already emit something at inference time, in application logs or a model endpoint's own capture. List each system, the serving platform, and the fields you can extract today. That list determines how much mapping work either option needs, and it is the single biggest driver of how long an integration actually takes.
- Ask both sides for the ingestion contract in writing
For Auditrek that is the REST and webhook payload shape and the SageMaker and Azure ML mapping guides. For IBM, ask IBM for the equivalent. Comparing two ingestion contracts side by side is far more informative than comparing two feature grids, because it tells you who on your team is doing the work.
- Test detection, not just storage
Immutable storage and tamper evidence are different properties. Ask what happens when a record is altered after the fact, and whether verification recomputes from stored material or reports a flag written at ingest time. Auditrek's verification endpoint walks the chain from the first sequence, recomputes each hash from the stored previous hash and payload, and returns the first broken sequence number.
- Generate one real report over a real date range
Pick one system and one month. A report you can read, hand to your own counsel and mark up is worth more than any demo. Note what the report cannot say: Auditrek's output carries an embedded line stating it is not legal advice and has not been reviewed by qualified regulatory counsel, because that judgement belongs to your counsel, not to a document generator.
- Price it against the count of systems, not the count of seats
Auditrek bills per org against the number of connected ai_system records, in starter and growth tiers, through Stripe. Model your own two-year system count before you compare that with any quote you receive elsewhere, because the growth curve of your AI inventory matters more than today's line item.
When to choose which
There is a version of this decision where the two are not alternatives at all, and that is worth naming before the trade-offs.
Reasons to keep looking beyond a record layer
- You already hold an enterprise relationship with IBM, or your risk function has standardised on tooling it has bought from IBM, and avoiding a new vendor, a new contract and a new security review is the deciding factor.
- Your requirement is broad governance work across many obligations, owners and review cycles, not per-inference record capture.
- Your procurement process favours an established vendor with an enterprise contracting path, and you have the budget and internal capacity to run it.
- You need capabilities we have not verified and therefore cannot describe. If your shortlist criteria include anything not in the Auditrek column above, read IBM's pages and ask IBM.
Reasons the Auditrek route may be the right one
- The gap you have been handed is narrow and specific: nobody can currently reconstruct what a given model output was, when, on which model version, and whether a human overrode it.
- You do not want to move models, retrain, or change the production inference path to get records. Events are posted to an API after the decision has already been made, using a documented payload mapping rather than a platform SDK.
- You are mid-market and the deciding constraint is how much internal effort a tool costs, not how many modules it has. A narrow layer has a narrow implementation.
- You want the sensitive-data controls to be structural rather than policy text: an ingestion gate that rejects payloads with recognisable raw sensitive fields, pseudonymisation of subject identifiers by default, and no credentials issued for a system until a DPA confirmation is recorded.
Running both
The two roles compose. A governance platform holds the register, the owners, the assessments and the control evidence. A record layer holds the per-decision trail that those controls assert exists. If you already have the first and are short of the second, the comparison is not really a choice, it is a sequencing question. The related pages in this pillar cover the dashboard a compliance officer looks at, the workspace access model for compliance, risk and engineering, and the wider picture for the person who has to produce the evidence.
Questions
Is Auditrek a replacement for IBM watsonx.governance or OpenPages?
We cannot answer that for you, because we did not capture verbatim wording from IBM's product pages and will not describe what they do. What we can say is what Auditrek is: one layer, capturing inference events over a REST or webhook API and storing them as SHA-256 hash-chained append-only records per ai_system, plus reports and a dashboard drawn from those records. It has no risk register, no control library and no enterprise GRC workflow. If your requirement is platform breadth, read IBM's own product pages, which are linked in our sources.
Do I have to migrate or change my models to use Auditrek?
No. Events are posted to Auditrek's REST endpoint or webhook after your model has already produced an output, using documented payload mapping guides for AWS SageMaker and Azure ML. There are no bespoke platform SDKs at this stage, so the work is mapping your existing event fields to the documented payload shape.
Why does the IBM column say so little?
Because we did not capture verbatim wording from IBM's product pages, and we will not paraphrase a competitor's capabilities or restate a price from memory. The two IBM URLs are in the sources so you can read the primary material yourself.
What does Auditrek cost compared with IBM?
Auditrek bills per org against the number of connected ai_system records, across starter and growth tiers, through Stripe. We publish no figure for IBM because we have not read one on IBM's own pricing material in this research pass; check IBM's pricing page or your procurement channel.
Does an Auditrek report mean our AI systems satisfy the EU AI Act?
No, and the report says so. The generator produces PDF and HTML output for one system and date range with fields mapped to Article 12 and Article 13 fields, and every report carries an embedded line stating that it does not constitute legal advice and has not been reviewed by qualified regulatory counsel. Whether your documentation satisfies a supervisory authority is a judgement for your counsel.
How does Auditrek prove a record has not been altered?
A server-side verification endpoint walks the chain for an ai_system from the first sequence to the record in question, recomputes each hash from the stored previous hash and payload, and returns a tamper status along with the first broken sequence number. It reports a recomputed result rather than a stored flag.
What happens to sensitive fields from credit, hiring or clinical pipelines?
Ingestion rejects payloads containing recognisable raw sensitive-field patterns before anything is written, and subject identifiers are pseudonymised by default. Separately, no API key or webhook secret is issued for an ai_system until a DPA confirmation with a signed_at timestamp is recorded, and that check is enforced server side.
Who logs in, and where does it run?
Auditrek offers auth-gated multi-user org access, with sign-in by email and password. It is hosted on Amazon Web Services.