54% of executives admit *gaps in third-party AI governance*

The gap nobody wants to talk about

Think of it like this: you hire a removal company to pack your china. They turn up with a van full of robots. One drops a plate. A year later, when the claim lands, the removers produce a log that says the robot never touched that box. Can you prove otherwise?

That is roughly where many organisations are with third-party AI. According to an IBM study conducted with Whistic, 54% of executives admit their governance frameworks for third-party AI systems and models have gaps, and 62% say third-party AI presents new risks their current processes don't address ^[1]. The problem is accountability: when a partner's AI does something on your behalf, and someone questions it, you need to show what actually happened.

What the rules ask for

From 2 December 2027, high-risk AI systems under Annex III of the AI Act must technically allow for the automatic recording of events over their lifetime ^[2]. That includes AI used for recruitment, credit decisions about people and certain law enforcement tasks. The logs are meant to make the system traceable: you should be able to see what it did, when and why.

But here is the thing: Article 12 says what to log. It says nothing about how to protect the logs. If your partner runs the system and keeps the logs, you are trusting them to hand over an accurate record if something goes wrong. That is fine until it isn't.

Where the gaps show up

The IBM study points to several trouble spots. First, visibility: 46% of respondents say they lack full visibility into the AI systems their third parties use ^[1]. If you do not know a partner has deployed AI, you cannot ask the right questions about how it logs decisions.

Second, contracts: many agreements do not spell out who owns the logs, how long they are kept or who can audit them. Third, capability: even when the contract is clear, the partner may not have the means to produce a log that holds up under scrutiny. A spreadsheet export is not the same as a tamper-evident record.

The new Product Liability Directive, which covers software placed on the market after 9 December 2026, makes this sharper ^[3]. Courts can order a defendant to hand over relevant evidence, and liability is strict for personal injury and certain property damage. If the AI that made the decision is your partner's, but the claim is against you, the question is whether you can get hold of the evidence.

Why logs kept by the partner are easy to challenge

Logs live where the system runs, and that is often at the partner's. The trouble is, a log the partner can edit is a diary, not evidence. It does not matter whether they would edit it; what matters is that they could.

Say a lender uses a third-party credit-scoring model and an applicant is declined. Six months later, the applicant asks for an explanation and hints at a discrimination claim. The lender asks the partner for the log. The partner sends a file. Did that file exist at the time of the decision, or was it written yesterday? If the partner is the only one who can answer that, the answer is trust us.

That is not a governance framework. It is a hope.

How a digital witness fills the gap

A digital witness does one thing: it proves a record has not changed since it was sealed. Here is how it works in a third-party AI context.

The partner's system writes a log entry for each decision: applicant reference, model version, output, timestamp and, if there is one, the name of the person who reviewed it. As each entry is written, the system sends a fingerprint of that entry to WitniumChain through the REST API, the SDK or the MCP server. The fingerprint is a short string derived from the entry; the entry itself never leaves the partner's systems. WitniumChain seals the fingerprint and returns a witness ID.

Months later, when the question comes, the partner hands over the log entry and the witness ID. You, the regulator or a court can check the fingerprint: calculate it from the entry you have been given and compare it with the one that was sealed. If they match, the entry is exactly what was written at that time. Change a single character and the fingerprints will not match.

That does not prove the decision was fair or lawful, or that the log is complete. What it does prove is that this entry has not been quietly rewritten since the day it was made. The partner can still keep the logs, but now anyone you share the record with can check that it has not been altered, without having to trust the partner or you.

How it would work

1. Partner's AI makes a decision and writes a log entry.
2. The partner's system sends a fingerprint of that entry to WitniumChain and gets a witness ID back.
3. Later, the partner hands over the entry and the witness ID. Anyone can check that the fingerprint still matches.

What this means for you

If your organisation uses third-party AI, here are four questions worth asking now:

• Does your contract say who keeps the logs, for how long and in what form?
• Can your partner produce logs in a way that shows they have not been changed?
• If a regulator or a court asked for proof of what the AI did on a given day, could you get it?
• Do your governance reviews cover how third-party AI decisions are logged and who can audit them?

The first 2,000 digital witnesses a month are free, so there is no reason a partner cannot start sealing log records as they go. The question is whether your contract makes them.

What to watch

The AI Act's logging requirement for high-risk systems under Annex III applies from 2 December 2027 ^[2]. That is just over a year. For systems under Annex I, the date is 2 August 2028. The new Product Liability Directive is already in force for software placed on the market after 9 December 2026 ^[3].

IBM and Whistic have said they will continue tracking third-party AI governance, so expect updated figures. In the meantime, the 54% who admit to gaps have at least said it out loud. The rest may just be quieter about it.