How to Prove Your AI Behaved

Your business uses AI. Your regulator wants evidence. Your insurer wants documentation. Your lawyer wants an audit trail. Most AI tools give you none of these. Here is what you actually need.

The problem is already real

Somewhere in your organisation, an AI system is making decisions. It might be screening CVs. Summarising contracts. Routing customer queries. Flagging transactions. Generating reports. Maybe all of these.

Now imagine one of those decisions goes wrong. A candidate is rejected and files a discrimination claim. A contract summary misses a liability clause. A customer receives advice that costs them money. A flagged transaction turns out to be legitimate and a client relationship is damaged.

The first question your lawyer will ask: what did the AI do, exactly? What inputs did it receive? What logic did it apply? What output did it produce? Can you prove that? For most organisations, the honest answer is no.

The obligation is already here

The regulatory framework for AI accountability is arriving on a concrete timeline, and for some obligations, it has already landed. The message across every jurisdiction is consistent: if you deploy AI, you own the consequences. Ownership means evidence.

Five things you need to prove

Strip away the regulatory language and the requirements reduce to five questions. Every regulator, insurer, auditor, and opposing counsel will ask some version of these.

1. What decision did the AI make? Not the summary. The actual output: the specific recommendation, classification, score, or action. Captured at the point it was produced, not reconstructed later.
2. What inputs did it use? Which data did the system access? What context was retrieved? Were there documents, database records, or prior decisions that influenced the output? If the system used retrieval-augmented generation, which passages were retrieved and how were they ranked?
3. What logic was applied? Which model version? What policy constraints governed the interaction? Were there safety filters, business rules, or human-in-the-loop checkpoints? What confidence level did the system assign?
4. Was there human oversight? Who reviewed the output before it was acted upon? Was the review meaningful (the person understood and evaluated the output) or nominal (they clicked approve)? If no human reviewed it, was the system authorised to act autonomously?
5. Can you prove the record is authentic? If someone challenges the evidence, can you demonstrate it has not been altered? A database record can be edited. A log file can be rewritten. A screenshot can be fabricated. You need a tamper-evident record where any modification is detectable.

Why logs are not evidence

Every AI platform produces logs. This creates a false sense of security. A log records that something happened. It does not prove the record is complete, that entries have not been modified, or that events are chained together so tampering is detectable.

When Cigna's AI system denied 300,000 health insurance claims over two months, spending an average of 1.2 seconds per claim, the legal question was not whether the system had logs. It was whether the decision-making process could be examined, reproduced, and verified.

Financial services regulators treat missing audit traces as books-and-records violations, with average fines of USD 5 to 10 million in 2024 and 2025. The bar for "we have logs" is rising. The bar for "we have evidence" is what the regulation actually sets.

What evidence infrastructure looks like

Building evidence into AI operations requires three layers that most platforms lack. Together, they answer all five questions: what decision was made, what inputs and logic were used, whether there was human oversight, and whether the record is authentic.

• AEE: Structured Records (Layer 1). A chain of discrete events: input received, context retrieved, policy evaluated, tool invoked, output generated, human review completed. Each event timestamped, attributed to a specific agent or user, and linked to the events that preceded and followed it. Every agent interaction is wrapped in a standardised envelope with explicit sender, recipient, correlation, and causality fields.
• VOLT: Cryptographic Integrity (Layer 2). Each event is hash-chained to the previous one. If any event is modified, inserted, or removed, the chain breaks. Not blockchain: no distributed consensus, no tokens, no overhead. The same integrity mechanism used in certificate transparency logs. Evidence bundles are exportable in formats aligned with SOC 2, HIPAA, and GDPR requirements.
• WARD: Independent Witnessing (Layer 3). A separate system monitors the evidence chain without storing content. If the primary record is compromised, the witness detects the alteration. Stores only hashes, never content. Witnessing occurs without creating additional copies of sensitive data, a property that matters for personal information, health records, or commercially confidential material.

How the layers connect: AEE captures what happened (the structured record). VOLT proves the record has not been tampered with (cryptographic integrity). WARD provides an independent witness that detects alteration without storing content. Each layer addresses a different class of challenge: AEE satisfies "show me the decision," VOLT satisfies "prove the record is genuine," and WARD satisfies "who else can confirm this?"

What your lawyer should be asking you

If your organisation uses AI in any capacity that affects people, money, or compliance, your legal adviser should be asking these questions now. Not after an incident. Not when the regulator writes.

• Do you have a record of every AI decision that affects an individual or a regulated process?
• Can you produce that record for a specific decision about a specific person, on demand?
• Can you demonstrate the record has not been altered since it was created?
• Do you know which AI model version was used and what policy constraints were in place?
• If a subject access request arrives tomorrow asking "what did your AI decide about me and why," can you answer it?
• Does your cyber insurance policy require AI governance documentation, and can you produce it?

The commercial opportunity for law firms

Every organisation deploying AI will need legal guidance on governance, compliance, and liability. The firms that understand AI accountability from the inside, because they govern their own AI use with the same rigour they advise clients to apply, will be the firms that win that work.

A solicitor who can say "we use AI in our own practice, and here is the evidence trail it produces" has a credibility advantage over one who says "you should probably have a policy." The former is advising from experience. The latter is advising from a template.

The market for AI compliance advice is growing in step with the regulatory timeline. EU AI Act, GDPR, sector-specific requirements from the FCA, PRA, ICO, and CQC. Employment law obligations. Insurance governance riders. Directors' duties. Every one of these creates advisory work. The firms positioned to capture it will be those that have already solved the problem for themselves.

Start with evidence, not with policy

Most organisations respond to AI risk by writing a policy. Usage guidelines. Acceptable use frameworks. Responsible AI principles.

Policies are necessary. They are not sufficient. A policy describes intent. Evidence proves execution. When the regulator asks what happened, the policy tells them what should have happened. Only the evidence trail tells them what actually did.

The order matters. Build the evidence infrastructure first. Capture AI decisions, inputs, logic, and oversight as structured, verifiable records. Then write the policy that describes the governance those records enable. This is backwards from how most organisations approach it. It is also the only approach that survives contact with a regulator, an auditor, or a courtroom.

Quox (quox.ai) is a source-available platform for AI agent orchestration and governance. Four open protocol specifications (AEE, AOCL, VOLT, WARD). Open standards for AI accountability.