When an AI system tells me a task is finished, I want to know where to look.
A useful answer might explain a payment, prepare a change or describe the next step. That still leaves a practical question: did anything happen outside the conversation? If it did, I should be able to find the result without asking the model to tell the story again.
In my earlier essay about what an agent may do without asking me, I wrote about the boundary before an action: a specific plan, a rehearsal and the checks required before it goes ahead. This is the question on the other side. Once the action has happened, what should it leave behind?
The answer needs somewhere to live
Accounting makes this easy to explain. A message saying that money moved is different from the corresponding entries in the books. The amount, the accounts and the recorded result have to agree. A clear explanation helps someone understand the transaction; it cannot stand in for the transaction.
That distinction was central to the accounting case study I presented at IEEE IRI. The model interprets the request, deterministic code writes the accounting result, and structural checks govern that write. I discussed the boundary in the four questions that came back after the talk. The published paper gives the technical account.
I keep coming back to it because the everyday test is straightforward. After the conversation ends, can the person responsible for the books inspect the result in the books?
What changed, and what did not?
I want the completion report to name the action precisely. Reading a balance, preparing a draft and posting an entry are different outcomes. Each can be useful. Confusing them makes the next decision harder.
A preview that changed no records should say so. A draft waiting for review should still be called a draft. If a request was refused, the explanation should name the reason and leave the user clear about whether anything was written. These are ordinary product details, but they are the details an operator needs when deciding what to do next.
The same discipline applies when I review work outside accounting. A local change, a private commit and a checked public release describe different states. I do not want the word finished to hide the distance between them.
A record has to help someone act
The next thing I look for is a record that can be found again. A reference to the affected entry or document is more useful than another paragraph of reassurance. It gives the person reviewing the work a place to start.
I also want the record to make the remaining decision visible. Who can review it? Does it need confirmation? Has the system stopped because a required condition is missing? An accounting check can establish that an entry balances without deciding whether the business chose the right account or intended the purchase. The person responsible still needs enough information to make that judgement.
This is a standard I want products to meet, rather than a claim that a reference number proves everything about an action. The record and the explanation have to be checked against what actually happened.
What happens when the result needs correction?
A successful example is only half the demonstration I want to see. I also want to understand an interruption, an uncertain result and a correction.
If a response is lost, repeating the request may be the wrong next step. Something may already have happened. I want the product to help establish the state before another attempt, rather than make the operator guess from the silence.
If the result is known but wrong, I want a correction process that fits the kind of record involved. Correcting the books is different from editing a sentence. The earlier result, the authority to change it and the effect of the correction all matter. Fluent conversation does not settle those questions.
The test I would use
Before I call an AI workflow complete, I want four answers: what changed, where I can inspect the record, who can act next and how we would recover or correct it.
Those questions do not require every product to become an accounting system. They do require the demonstration to reach the result someone will rely on. I want to close the chat and still know what happened.