A small robot beaming with pride as it holds up a completely blank sheet of paper, while an auditor with a clipboard waits, patient and unimpressed.

The compliance file, as most deployments would file it today. AI-generated image.

Key Takeaways

The rules never ask whether your agent is accurate. Article 26 wants named human oversight, six months of logs, and incident reporting. Not “is it good.” Can you say what it did.

85% per step is 20% end to end. Accuracy multiplies across a ten-step workflow. The vendor deck quotes a per-call figure. You deployed a chain.

The adoption statistics count different things. 72%, 31% and 23% are all published figures for enterprises “in production” with agents. I tried to trace each to a methodology and could not reach one.

Six months of logs is architecture. You cannot backfill the reasoning, the retrieved context or the discarded branch into an incident that has already happened.

There is a sentence people keep waiting for in the EU AI Act’s high-risk rules, and it is not there. Nothing in them asks your agent to be accurate.

Article 26 is short and unglamorous. Deploy a high-risk AI system and you owe: use in line with the provider’s instructions, human oversight by a named natural person with the competence, training and authority to exercise it, management of the input data you control, logs kept for at least six months, incident reporting to the provider and the market surveillance authority, and notification to affected workers where it is used in the workplace.

The dates have already moved once and the Commission can still bring them forward. I would not plan around them either way. None of this work happens in the quarter before a deadline — logs you did not design cannot be backfilled into an incident that already happened. Whatever the calendar ends up saying, this is the build window.

Three things this breaks

1. “The agent is 85% accurate” was never a statement about your workflow.

I keep coming back to this because it is the most expensive misunderstanding in agentic AI: per-step accuracy multiplies. Chain ten steps at 85% each and end-to-end success is 20%. At 90% it is 35%. At 95%, 60%.

A row of ten identical cheerful robots passing a gold parcel down the line. It is large in the first robot's arms and smaller at every handoff, until the tenth is holding nothing at all and looking down at its empty hands, still smiling.

Ten steps at 85% each. The parcel is what survives. AI-generated image.

You can ship on the per-call number and find the workflow number out in production. What you cannot do is document competent oversight of a system whose failure rate nobody has measured. “We tested the model” does not describe a chain.

2. The adoption statistics do not agree, and that is going to matter.

Look at what has been published this year. One widely repeated figure puts 72% of enterprises at production adoption. Another reports 31% with at least one agent in production. Another puts 23% actually scaling. IDC says 88% of proofs-of-concept never reach widescale deployment. Gartner expects over 40% of agentic projects cancelled by end-2027.

Before publishing this I tried to trace the first three to a primary source. I could not. Each one leads to an aggregator citing another aggregator. The 31% is attributed downstream to S&P Global Market Intelligence and McKinsey, but I never reached the instrument: no sample size, no definition of “in production”, no field date. IDC and Gartner at least have a named house behind their numbers.

I would rather show you that than assert it. The provenance thins the harder you pull on it. This is the failure I keep running into: a published number that does not transfer, quoted as though it does. Once “we are in line with industry adoption” is a line in a governance file, where it came from stops being academic.

3. Six months of logs is a system design constraint, not a storage line item.

Retaining logs is easy. Retaining logs that answer what did it do, and why is a decision you make before you ship. Most frameworks log the final response and the tool calls. Fewer log the intermediate reasoning, the retrieved context, or the branch that was discarded. Six months from now, an incident report asks which of those mattered.

A painted illustration of an enormous dim archive hall, shelves of filed records rising far out of frame on both sides. One small figure stands at the bottom of it holding up a lantern, whose light reaches only a few shelves in either direction.

Six months of logs, and one person expected to find the relevant hour in them. AI-generated image.

What actually helps

Measure the chain, not the call. Run the workflow end to end, count the failures by step, and write the number down where your team can see it. If it is not written down, the honest answer to a regulator is that you do not know.

Make your evaluator take the exam. If an LLM judges your agent’s output, you have inherited that judge’s error rate. My own harness sits the judge against a fixed calibration set before it grades anything, and prints its measured error rate in every report. It took an afternoon, and it is the reason I trust the scores that come out of it.

Design the log for the question, not the volume. Decide now which fields let you reconstruct a decision, because you cannot add them retroactively to an incident that already happened.

The bottom line

The regulation did not ask for better agents. It asked for agents whose behaviour you can account for, and that is a measurement problem most teams deferred while the demos were going well.

A flat cartoon illustration of the same small robot handing the auditor a thick, full bound ledger. The auditor looks down at it with approval, pen lowered.

The version where you kept the record. AI-generated image.

Whenever the date lands, it will not be the date that catches anyone out. It will be the six months of logs that had to already exist by then.

Sources: EU AI Act Article 26, Article 14 (human oversight), AI Act Service Desk. The adoption figures are drawn from separate published surveys whose definitions of “agent” and “production” differ; that disagreement is the argument of that section rather than an aside.