governance as the gate to launch

three days ago, on 2 august 2026, the eu ai act's obligations for high-risk ai systems came into force.¹ if you build software that goes into a medical device, this is your category. article 6 makes an ai system high-risk when it is a safety component of a product that already has to pass a third-party conformity assessment, and medical devices already do.² you do not have to argue your way into the classification. the device drags the model in with it.

what lands with it is a short list of things the system has to be able to do. article 9 wants a risk management process that runs across the life of the system. article 12 wants record-keeping. article 19 wants the automatically generated logs kept. article 14 wants human oversight that a person can actually exercise.¹ read those together and they describe properties of a running system, not a binder assembled at the end.

the retrofit does not work

in the regulated builds we have run, the thing that stalls a launch is almost never the model's accuracy. it is that nobody can reconstruct how a given output was produced, who looked at it, and what they were looking at when they approved it. that reconstruction is either a property the system had all along or it is a story someone writes afterward, and the second one does not survive a reviewer.

logging is the clearest case. a team that adds logging in the last sprint gets logs of the last sprint. the decisions that matter happened months earlier, in the prompt that got rewritten, the threshold someone moved, the training set that quietly changed. none of that is recoverable later. it either got written down while it was happening or it is gone.

human oversight retrofits badly for a different reason. bolted on at the end it becomes a person clicking approve on an output they have no way to inspect. that satisfies a screenshot and nothing else. oversight a person can exercise means they can see the inputs, see what the system did with them, and disagree with it in a way the system respects.

what we do instead

we decide the review points before anything gets built. where does a human have to be in the loop, what are they looking at when they are, and what happens when they say no. those answers change the architecture, which is why they cannot come last.

then we name who signs. not a team, a person, per decision type. this is the least technical part of the work and the one that most often has never been settled. a model spreads responsibility thin by default, and organizations tend to discover that only when something needs owning.

and we treat the log as the product's memory rather than as an audit artifact. if the log is written for the auditor, it gets written once and it is thin. if it is written so the team can answer its own questions about why the system did something last tuesday, it is thorough, and the auditor is satisfied by a side effect.

why this is a launch gate and not compliance work

the deadline is not the interesting part. article 6's classification rule itself does not apply until 2 august 2027, so there is a year of runway on that specific point.¹ the interesting part is that the work is upstream. a governance requirement that changes where a human sits in the flow is a design requirement. one that changes what gets recorded is a data model requirement. handing either to a compliance function after the build is a request to redesign the thing.

for buyers this shows up as a question that used to come late and now comes early: can you show me how this gets defended. teams that designed for it answer in an afternoon. teams that did not spend a quarter assembling something that reads like an answer.