On the weight of unfinished work
Somebody
Handled It
The moment AI feels different is the moment you no longer have to supply every next step.
You can get a great deal of help and still be carrying the entire job.
Someone sends you the answer. Now you have to apply it. They prepare the file. Now you have to get it into the right system. They make the change. Now you have to remember to check whether it worked.
Each contribution is useful. Yet the unfinished thing keeps returning to you. You are the place where its context lives, where its loose ends collect, where the question What happens next? must be answered.
Then, occasionally, a different kind of message arrives. It explains what was understood, what was done, how the result was checked, and what still needs a decision.
You read it, and something releases.
That is the feeling I have been trying to name about AI.
01 / A small repair
The message that made it visible
The setting was a restaurant in Mount Pleasant, South Carolina. I was working on Aki, an AI receptionist. There was an ordinary, consequential problem: when a requested time was full, a caller could end up in the turn-away records without a useful phone number for a callback.
There were other unfinished matters around the project, too. One depended on the restaurant providing staff names and email addresses. I gave the AI agent helping me with the project an instruction that amounted to two words: Fix it.
Here is the report I received:
“I took your ‘fix it’ to mean the Aki problem, since the shared login can’t change until the restaurant sends real names and emails. When a caller asks for a full time slot, Aki now saves their phone number with the turn-away record, just as it does for bookings, so the Callbacks tab will have numbers to call back. It’s live, and the deployed code matches the local file exactly. I tested it with a fictional caller by phone and by text, and both saved the number. My test records are deleted. That was the last known problem to fix before Aki can go on the main number.”
There are two AI roles in this scene. Aki is the receptionist. The message came from the agent working on the software behind Aki. The receptionist was being prepared to handle calls; the other agent was handling a repair.
What struck me was the distance between my instruction and the point at which the work came back to me.
A small experiment in the feeling
Who is carrying the next step?
Imagine you asked for the callback problem to be fixed. Read each illustrative reply, then notice what you would have to do next.
The difference is how much unfinished work the reply hands back.
02 / The space between instructions
It carried the middle
The first sentence of that report deserves attention: “I took your ‘fix it’ to mean…”
Two words were enough because they sat inside an ongoing project. The agent had context about what was unresolved and what was blocked. It selected an actionable problem and made its interpretation visible. In another situation, the right next step might have been a question. Here, it reported why it had proceeded.
Then it described a change to the live system, tests through two channels, and cleanup. The little phrase about deleting test records mattered. Cleanup belongs to the work, even when nobody remembers to request it separately.
Those are the details that usually fill the middle of an assignment. Between “fix this” and “it works,” somebody must choose the next action, inspect the result, notice what remains, and keep going.
Much of the AI conversation focuses on the quality of the first answer. This experience drew my attention to what happens after that answer would normally end.
Software has acted without us for decades. Scheduled jobs, deployment pipelines, and other automation already make changes and run checks. The important shift here is the agent’s ability to use project context and intermediate results to choose a path through work whose every step I have not specified.
Anthropic’s engineering account of agents describes this feedback loop: a system takes actions, receives information from its environment, and uses that information to determine how to proceed, including when to seek human input. The technical mechanism is a loop. The human experience can be a handoff.
The task has somewhere else to continue.
03 / The work behind the work
Remembering to remember
A task occupies more than the minutes spent doing it. It can also occupy a small place in your attention until you know how it will be finished.
We already build supports for this. A calendar remembers an appointment. A note by the door reminds us to take something. Research by psychologist Sam Gilbert found that people used external reminders more when memory demands or distraction increased, and that those reminders improved performance in the tasks studied. He calls this offloading intentions.
That research concerns reminders, not AI delegation. But it helps locate the familiar human need beneath this new experience: we do not want every future action to depend on keeping it alive in our own minds.
A reminder can bring an unfinished task back to you. A capable agent, given the context and authority to act, can sometimes bring you a finished result instead.
This is why faster output does not fully explain the feeling. Ten excellent answers can still leave you with ten things to shepherd. One well-completed assignment can let you turn your attention elsewhere.
I felt as if somebody had handled it because I recognized a pattern I already knew from working with dependable people: the work had continued without requiring me to become the next instruction at every turn.
The word somebody describes that recognition. It does not establish that the system is conscious or has become a person. An agent can occupy a role in a working relationship without having the inner life we associate with a colleague.
04 / What makes relief justified
“Done” must point to something
There is a trap inside this feeling. A convincing completion message can create relief even when the work is incomplete.
The message reproduced here is the agent’s report of the repair. Its confidence alone cannot establish that every claim is true. The meaningful evidence is in the deployed system, the test results, and the saved records the report describes.
Anthropic makes this distinction explicit in its guidance on evaluating agents: the transcript records what the agent said and did; the outcome is the resulting state of the environment. A booking claim and an actual reservation are different things.
For the restaurant repair, the practical question is whether a caller turned away from a full slot now leaves a usable callback number. Matching the deployed code to a local file helps confirm what was shipped. Testing by phone and text checks two routes through the experience. Removing test records restores the working environment. Together, these checks make a completion claim inspectable.
They still do not prove that the entire receptionist is ready for every real caller. The phrase “last known problem” has a boundary built into it. The decision to put Aki on the main number remains a human decision.
This is where the word responsibility needs care. We can delegate the practical responsibility for carrying work forward. We remain accountable for what we authorize, what we accept, and whom the system affects.
Microsoft Research describes appropriate reliance as avoiding both excessive trust in faulty AI outputs and rejection of useful ones. In this setting, that means learning which work an agent can carry, under which conditions, with which evidence.
Trust becomes specific. This task. These permissions. These checks. A clear point at which the system must return to a person.
That specificity is what makes it possible to let go of some supervision without pretending that judgment has become unnecessary.
05 / The feeling
A place to put the weight
If you have managed a project, cared for a household, or run a small business, you know the difference between receiving help and having something handled.
The second experience changes what you must keep track of. You still care about the outcome. You may still need to review it. But you no longer have to hold every unfinished step in place yourself.
That is the possibility I saw in an unremarkable software repair for a restaurant. An agent interpreted a request within a shared context, reported carrying it through implementation and testing, and returned with a result connected to the larger purpose.
No grand announcement accompanied it. I read a message about a phone number being saved correctly.
But the message described a different relationship to the work. I could ask what needed to be accomplished and receive an account of what had happened. The distance between those two points had become something I could begin to entrust to a system.
We tend to picture powerful AI as a spectacle: a machine doing something astonishing while we watch.
The change may arrive much more quietly. You look up from a message and realize that, for this one thing, you no longer need to ask what happens next.
The feeling is relief.
Somebody handled it.
1 thought on “Somebody Handled It”