The AI Obligation Sits With the Deployer
The obligation generally lands on the organisation deploying the AI system, not on the one that built the model. Most AI governance conversations assume the opposite.
That single fact reorganises who in a business has to do something about it, and it moves the work away from the team most people have assigned it to.
What the text actually says
The EU AI Act is explicit. Article 12 requires high-risk systems to log events automatically across their lifetime, and deployers must retain those logs, with a floor of six months under Article 26(6).
Timing has moved during the Digital Omnibus process and the Annex III high-risk date now sits at 2 December 2027. Check the current text with counsel rather than trusting a blog, this one included: dates in this area have moved more than once and will likely move again.
Singapore's IMDA went further in January 2026, publishing an agentic AI governance framework that expects agents to carry verifiable identities and audit trails rather than merely being logged.
The direction is not in doubt even where the dates are. Two regulators on opposite sides of the world arrived at the same requirement from different starting points, which is usually a sign the requirement is about the technology rather than about the politics.
Why this lands on the wrong desk
Most organisations have filed AI regulation under model procurement. Which vendor, which model, what is in the contract, what did their assurance pack say.
That work matters and it answers a different question. A deployer obligation is not discharged by choosing a compliant supplier, because the thing being asked for is a record of what you did with the system, which your supplier was never in a position to keep.
The practical test is simple. If the vendor relationship ended tomorrow and their systems went dark, would your organisation still hold the record of what was decided and by whom? If the answer is no, the obligation has not been met regardless of what the contract says.
It is less about AI than it sounds
If an AI-assisted process produces a decision affecting somebody, you have to show what the system did, what a human did, and who took responsibility.
Read that again without the word AI in it and it is the oldest requirement in regulated industry. Show the act, show the actor, show the authority. The AI part is the trigger that brings a process into scope; it is not what the obligation is about.
This is good news, because it means the capability is not exotic. It is record-keeping, with an unusually specific set of fields and an unusually long retention floor.
It is also bad news for anybody hoping a model vendor will supply it, for exactly the same reason. Nobody outsources their own books.
What the six-month floor really implies
A floor is not a target, and treating it as one is the common mistake.
Six months is the minimum the text requires. The period that matters operationally is however long it takes for a decision to become contested, and that interval is set by your business rather than by the regulation. In lending it can be years. In employment it is usually measured from the point somebody leaves.
So the retention question is not what the Act requires. It is what your own disputes look like, with the Act setting a line below which you cannot go whatever your answer.
Organisations that size retention to the floor will find themselves, at some point, lawfully in possession of nothing useful.
What a deployer has to be able to produce
Strip the regulation back to its operative demand and it is a short list, which is the useful thing about it.
For a given decision that affected a person: what the system was asked, what it returned, what a human did with that, who that human was, what authority they held to do it, and when. Plus the ability to find all of that from an identifier the complainant would know, such as their own reference number, rather than from an internal trace id.
That last point defeats more organisations than the rest combined. Logging is usually indexed by request, by service, or by time. A person asking about their own case has none of those, and an obligation you can satisfy only if the regulator happens to supply a trace id is not an obligation you have satisfied.
The model is bought. The record is built.
Firms treating this as a model procurement question are solving the wrong half.
What Article 12 obliges you to hold is what the system did, what a person did, and who took responsibility, kept for as long as the text says and retrievable by somebody who was not there.
That last phrase is the one to design against. Retrievable by somebody who was not there means the record has to make sense to a stranger, after the people involved have moved on, without a guided tour. Most internal logging fails that test comfortably while passing every other one.
Questions
Who is responsible for AI logging under the EU AI Act, the vendor or the deployer?
The obligation generally lands on the deployer. Article 12 requires high-risk systems to log automatically across their lifetime and Article 26(6) sets a six-month retention floor on the organisation deploying the system, not the one that built the model.
Does buying a compliant AI vendor discharge the obligation?
No. The record required is of what your organisation did with the system, which the vendor was never in a position to keep. The practical test: if the vendor went dark tomorrow, would you still hold the record of what was decided and by whom?
How long should AI decision records be kept?
Six months is the floor under Article 26(6), not a target. The operative period is how long it takes a decision to become contested in your business, which in lending can be years and in employment usually runs from the point somebody leaves.
What does an AI audit trail need to contain?
What the system did, what a human did, and who took responsibility, retrievable by somebody who was not there. That last condition is the design constraint: the record must make sense to a stranger after the people involved have moved on.