Audit-grade evidence, in seven AI verticals — and far beyond them.

It's a Tuesday morning at a Stockholm law firm. A partner is finalising a share purchase agreement and asks her AI tool to surface every clause where the buyer's standard template diverges from market norms. The tool returns a structured list with citations. The deal closes. Eight months later, the buyer is in litigation with the seller, and counsel for the seller subpoenas the diligence process.

Without witnessing: the firm produces logs from its AI vendor, who produces logs from theirs, who produces a screenshot. The chain of custody is us trusting them trusting them. In adversarial proceedings, that's typically not enough.

With Witniumchain: every output of the AI tool was hashed and sealed in real time. The firm produces a verification URL. Opposing counsel runs it. The chain — independent of the firm and of Witnium — confirms that exact output was produced at exactly that timestamp, on exactly those inputs, by exactly that model version. The discussion moves on.

The integration takes an afternoon. One API call per output, one verification URL per response, one defensible evidence chain across the platform.

Hiring screens, performance scoring, and termination-support tools are explicitly named as high-risk under Annex III. The reason is straightforward: when an AI-influenced decision materially affects someone's employment, that person has a right to understand and challenge it.

A Helsinki-based recruitment platform serves several thousand corporate customers across the Nordics. Each candidate flows through automated screening before a human reviews the shortlist. When a rejected candidate files a complaint with the Discrimination Ombudsman, the question becomes: what did the model see, what did it score, and which human signed off on the decision?

With Witniumchain, every screening decision is witnessed — the candidate's anonymised application fingerprint, the model's output score, the threshold applied, the human reviewer who confirmed or overrode it. The platform doesn't have to defend its logs; it points the regulator at the chain.

Claims adjusting is being transformed by AI faster than most observers realise. A claim is filed, photos are uploaded, an AI estimates damage, a settlement offer is generated. Speed has improved dramatically. Trust has not.

When a claimant disputes a settlement — particularly in jurisdictions where insurance ombudsmen take consumer complaints seriously — the insurer needs to demonstrate not just that the AI worked, but exactly what it worked on. Were the submitted photos altered between intake and adjudication? Did the model see the medical report or only the summary? Was the damage estimate based on the actual damage description or an autocomplete artefact?

Witniumchain gives insurers a single forensic timeline per claim: every input the model saw, every output it produced, every human override, sealed in real time. The same record satisfies internal audit, the ombudsman, and — increasingly — the reinsurance partner who wants to verify that the underwriting AI is operating within its declared parameters.

Credit scoring is named in Annex III. AI-assisted underwriting and AML decisions sit one regulatory shock away from being treated identically. DORA, separately, imposes operational-resilience and forensic-readiness obligations on financial entities and their critical ICT providers. Witniumchain sits squarely in the intersection.

A consumer lender uses a model to score loan applications. The model is well-documented, the development team is competent, but six months from now a borrower will default, a regulator will ask why the loan was approved, and the answer "our model thought it would work out" will not be sufficient. The lender needs to demonstrate that the inputs the model saw were the inputs the borrower provided, that the score the model produced was the score the underwriter saw, and that any human override is logged with a reason. That entire chain is what Witniumchain provides as a primitive.

Clinical decision support, triage, and diagnostic AI inhabit Annex I (product-embedded) and Annex III (high-risk) at once. They also operate in an environment where the cost of an unverifiable decision is, occasionally, someone's life.

A hospital uses a diagnostic AI to flag chest X-rays for radiologist review. The system performs well in the validation studies, the deployment goes smoothly. Two years later, a malpractice case alleges the AI missed a finding that should have been caught. The hospital's defence depends on being able to demonstrate: the model version in use at that time, the image hash submitted, the model's actual output, the radiologist's review and confirmation, and the time elapsed between each step. Without an external chain, all of that is reconstructed from internal logs the hospital itself controls — which, in adversarial proceedings, is treated as documentation rather than as evidence.

Witniumchain externalises that chain. The hospital is no longer asking the court to trust its own records.

Annex III names AI used for "determining access or admission" to educational institutions, for "evaluating learning outcomes," and for "monitoring and detecting prohibited behaviour of students during tests" as high-risk. The implications are not yet fully absorbed by the ed-tech sector.

A Nordic ed-tech company offers AI-assisted essay grading to upper-secondary schools across three countries. Students whose grades have material consequences for university admission are entitled, under both the AI Act and existing GDPR Article 22 provisions, to understand and challenge the automated decision. The school is the deployer, the ed-tech company is the provider, and both have obligations.

With Witniumchain, every graded essay is witnessed: the essay fingerprint, the model output, the rubric applied, the teacher's review status. When a student appeals a grade, the appeal process operates on cryptographically verified evidence rather than on the company's word. This is, increasingly, what schools' procurement teams are asking for.

Benefits eligibility, asylum and migration processing, and law-enforcement-adjacent AI are uniformly Annex III. Public-sector deployers operate under the additional constraint that their procurement processes increasingly require demonstrable independence of the audit trail from the vendor providing the AI.

A government agency uses AI to triage incoming applications for a benefit programme. Some applications are approved automatically, some flagged for human review, some rejected. When a rejected applicant requests reasons under administrative-law obligations, the agency must produce not just an explanation, but evidence that the explanation matches what actually happened.

Witniumchain provides that evidence as a primitive. The agency does not have to defend its internal logs. It points the appellant — and, if necessary, an administrative court — at a chain it does not control.