Most AI work is sold on the answer: the summary, the draft, the recommendation, delivered faster than a person could produce it. The answer is the least durable thing an AI system makes. It is stale within the week, and within the quarter nobody can say how it was reached.
What lasts is what the working day leaves behind. What was checked that nobody asked about. What was learned, and who agreed it was true. What was stopped before it ran. And the record of all three. That is what you should be buying from an AI supplier, and it is the easiest thing to ask them to show you.
We run our own business on ORCA, and this is what a working day on it looks like. The same loop is the one we build and run for clients.
Before anyone asks: observe what is actually running
The day does not start with a plan. It starts with observation. Scheduled checks compare what each system is doing against what it is supposed to be doing: services answering, evaluations holding, scheduled work having actually run.
That last one is the check most operations skip, and it is the one that matters. A component that crashes raises an alarm. A job that was never scheduled raises nothing at all, and every dashboard stays green. Absence has to be watched for deliberately, because it will never announce itself. Our checks assert that each scheduled job actually ran; a quiet dashboard does not count as a pass.
Observation is the unglamorous phase of the loop and the one that decides the rest. Fast decisions made on a stale picture feel decisive and lose all the same. Most of what looks like an execution problem is an observation problem.
A system nobody can observe is not finished, whatever the delivery plan says. Monitoring is designed in when a service is provisioned, never retrofitted afterwards.
Orient: knowledge that compounds, because it is governed
When a question arrives, it is not answered from a blank page. It is grounded against what the organisation already knows: earlier decisions and the reasons for them, the approach that failed last time, the constraint someone wrote down months ago. Every answer carries its provenance, so a reader can see which sources contributed and judge them.
Then the day's learning is written back. This is where most AI memory goes wrong. A system that stores everything it is given accumulates, and accumulation is not the same as knowledge: it fills with half-true notes that later answers will cite with full confidence. In ORCA nothing reaches a shared store without a named approver accepting it. Personal data is detected and replaced at the boundary before anything is stored or sent for inference.
That approval step is the whole difference. Ungoverned memory degrades with use. Governed memory compounds: each day starts from a better position than the one before, because what was learned yesterday was checked before it was kept.
Decide: challenged before it executes
The expensive mistakes in AI operations are rarely wrong answers. They are wrong actions taken confidently: a change applied, a message sent, a record overwritten. So the decision is where the loop slows down on purpose.
Significant design decisions go through adversarial review, deliberately looking for the failure mode rather than confirmation. Work produced by one pass is reviewed by another whose job is to disagree with it, before it merges or ships. A system that writes its own work and approves its own work holds one opinion, however many steps it runs.
Every infrastructure change carries a documented rollback and a preview of what will change before it is applied. If we cannot describe how to undo it, it is not ready. And some decisions are not the system's to make at all. Publishing, spending and anything outward facing stop at a person who can say no, and the system is built to hand those over cleanly rather than route around them.
This costs time, and it is meant to. Disagreement before execution is cheap. Disagreement after it is an incident.
Act, and leave a record you could hand to a regulator
By the end of the day, the useful output is not the pile of answers. It is the trail: what was asked, what was answered, what was stored, who approved it, and what was refused. Every running thing traces back to why it exists, from the decision that started it, through the work item, the change and the release, to the metric that shows whether it worked.
That trail is not a report assembled when someone asks. It is how the work is structured, so it exists whether or not anyone ever asks. When a board, an auditor or a regulator wants to know why a system behaved the way it did, the answer is a document, not a recollection. That is the test the new rules apply. Colorado's draft rules hold that if you cannot explain it, you have not complied, and under the EU AI Act the record either already exists or already does not.
Four questions to ask any AI supplier
You do not need to read anyone's architecture to tell whether their AI operation runs this way. Ask about yesterday.
- What did your system check that nobody asked it to? If the answer is nothing, it is not being operated. It is being used.
- What did it learn, and who approved it? Knowledge that nobody accepted is a liability that has not surfaced yet.
- What was stopped before it executed? A loop that never says no is not challenging anything.
- Can you show me the record? Not a summary written for the meeting. The trail the work left behind.
A supplier who answers all four from their own operation has a loop. One who answers from a slide has a pitch.
We build this loop into your operation, or run it for you to agreed service levels: observation designed in from the first service, knowledge that is governed before it is kept, decisions that are argued with before they execute, and a record you own outright. If your AI work produces answers but no trail, talk to us about what it would take to change that.