Your logs say what your AI did. They do not say whether it mattered.
Intent attribution is the practice of tying every action an AI takes back to the goal that was declared before the work started. A task identifier is created when the goal is stated and travels with every tool call and every artifact after it. Without it you have a trace. With it you can answer what your AI accomplished, toward which objective, and who decided it was worth doing.
Three things you cannot currently see.
Ask most companies running AI what it did last quarter and the honest answer is a shrug. Not because nobody is paying attention, but because the question has no place to be asked from. It shows up as three separate complaints that share one cause.
Black-box AI
You know the team uses AI every day. You could not produce a record of it if someone asked. Not a rough one. Any.
Silent failures
An agent runs on a schedule and hits the same wall every time. You find out three weeks later, because someone noticed the output stopped arriving.
Connector blind spots
Some portion of your integrations are erroring or dead right now. There is no dashboard that would tell you which, so the failure surfaces as a wrong answer instead of an alert.
The shared cause is that requests reach your tools without passing through anything that records them. Nothing is watching, so nothing can be reported.
Agent observability instruments the engineering layer, not the business one.
This is not a criticism of the tooling. The tooling is good and getting better fast. The OpenTelemetry GenAI semantic conventions have become the de facto standard for AI telemetry, and they cover model attributes, token usage, latency, tool failures, agent role, and prompt versions. Vendors have built capable flight recorders on top of them.
But those conventions are still experimental, with no 1.0 release as of 2026, and they carry no attribute for business intent (OpenTelemetry GenAI semantic conventions). Every attribute answers an engineering question. None answers the one an operator actually gets asked.
What did our AI accomplish this quarter, attributed to which initiative, at what cost, decided by whom?
That is not a dashboard problem. You cannot chart your way to it, because the data required was never captured. It is a missing primitive.
Observability tells you the agent ran. Intent attribution tells you the decision was worth running it.
The primitive is a declared goal, carried through every call.
The mechanism is unglamorous. Before work starts, the goal gets stated. That statement creates an identifier. The identifier rides along with every tool call made in service of the goal, including calls made by agents that the first agent spawned, and every artifact produced attaches back to it. When the work finishes, the task closes with an outcome.
Two things fall out of that, and both matter more than the record itself.
The first is that a vague request cannot proceed. If a goal has to be stated before work begins, fuzzy asks get sharpened at the point of entry rather than cleaned up at the point of output. Most teams try to govern the output. Governing the input is cheaper and it works better.
The second is that the record becomes readable as work rather than as log spam. A list of ten thousand tool calls tells you nothing. The same calls grouped under the goals they served tell you which work your AI is repeatedly doing well, which is the only honest basis for deciding what to automate next. That is why this page exists inside a method rather than beside it. How we work is built on the evidence this produces.
We hold ourselves to it. In the 84 days to 2026-08-09 our own operation ran 2,423 automated jobs and 25,720 tool calls across 133 tools, and every one of those calls was attributed to one of 1,463 declared tasks. That is what made it possible to see which fourteen processes had earned promotion into scheduled agents, rather than guessing at them in a planning session.
As of 2 August 2026, this stopped being optional for some operators.
Article 14 of the EU AI Act requires that high-risk AI systems be designed so that natural persons can effectively oversee them while in use, including understanding the system's limitations, monitoring its operation, and detecting and addressing anomalies and unexpected performance. The obligation falls on the provider to make oversight possible and on the deployer to assign it to people with the competence and authority to exercise it. Those obligations took effect on 2 August 2026 (EU AI Act, Article 14).
Oversight you cannot evidence is a difficult position to defend. The useful property of an intent-attributed record is that the evidence is a by-product of operating normally, not a reporting exercise somebody has to run before an audit. Operators who already have it will answer the question. Everyone else will be reconstructing it from logs that were never designed to answer it.
Own the record. Do not rent it.
Every action an agent takes becomes a record of who asked, what it touched, and why. That record is the most valuable artifact in the whole stack, because it is how you answer for what your automation actually did.
So the question is not whose observability dashboard looks nicer. It is whether the place that record lives belongs to you. We keep that layer on the customer's side deliberately. If we ever part ways, the trail stays with the business that generated it.
If your agent did something you had to explain to a client next week, where would you go to find out what happened?
Common questions
What is intent attribution?
Intent attribution is the practice of tying every action an AI takes back to the goal a human or agent declared before the work started. A task identifier is created at the moment the goal is stated, and it travels with every subsequent tool call and every artifact produced. Without it you have a trace of what happened. With it you can answer what was accomplished, toward which objective, and who decided it mattered.
Is that not what AI observability tools already do?
Not quite. The OpenTelemetry GenAI semantic conventions and the vendor tools built on them instrument the engineering layer very well: tokens, latency, tool failures, agent role, prompt versions. Those conventions are still experimental as of 2026 and they do not carry a business intent attribute. So the stack can tell you an agent ran and what it cost. It cannot tell you which business objective the run belonged to.
Why does an audit trail need to be owned rather than rented?
Because the record is how you answer for what your automation did, and that answer should not live somewhere a vendor can withhold, reprice, or discontinue. If an agent does something you have to explain to a client next week, the place you go to find out what happened should belong to you.
What does this have to do with the EU AI Act?
Article 14 requires that high-risk AI systems be built so that natural persons can effectively oversee them, including detecting and addressing anomalies and unexpected performance. Those obligations took effect on 2 August 2026. Oversight you cannot evidence is difficult to defend, and evidence is exactly what an intent-attributed record produces as a by-product of running normally.
Where this sits.
Visibility is one of four things that break when AI reaches business tools with nothing in between. The other three are access, consistency, and lock-in. How we work covers all four and what one layer in the gap fixes.
Related reading: who can your AI actually reach? on the access half of the same problem, and bring your own agent on why the governance layer matters more than the agent you pick.
Want to see what your AI did last month?
Book thirty minutes. We will look at what your current setup could and could not tell you, and what an attributed record would have caught.