Bill Gates recently used the term Human Reserved for work we deliberately keep for people, even where machines could technically do it. He compares it with a nature reserve: you could build there, you choose not to, because too much would be lost. His examples are about jobs and care, about a robot that shouldn't be the one to bring you bad news, about who then gets to draw those lines. That last question he leaves open.

I find the term useful on a much smaller scale as well: the way a company puts AI and agents to work in its daily operations. That is my application, not Gates's. There it is rarely about which whole job should stay human. It is about which decision you don't hand over, even when the system can technically perform the corresponding action perfectly. For Gates the emphasis is on what we don't want to lose. Inside a company another question comes with it: who has the authority to take the decision, and who remains answerable for it afterwards?

Being able, being allowed, being authorised

With AI we almost always start with the first question: can the system do this? Compare sources, write a proposal, produce code, summarise a file, spot a deviation. By now the answer is usually yes, and where it is still no, that is often temporary.

The second question is whether it is allowed to do that in this environment. May it modify a file, update a status, prepare a draft message? That is a matter of access and settings, and most companies handle it reasonably, if only because IT asks for it.

The third question rarely gets an answer of its own: is the system authorised to decide on this? That is something other than access. A new employee can operate the ERP perfectly and have every right in it, and still not be allowed to grant a discount, send a quote above a certain amount or put a customer on hold. Nobody finds that strange. In a company, authority is always something given explicitly, through a mandate, a job description or an agreement. It doesn't follow from someone making the right call.

The same applies to an AI system, it is just looked at that way less often. A system can have enough information to make a good proposal without therefore being allowed to determine that the proposal is now the decision. It can correctly judge what is probably the best choice, and still not be mandated to grant an exception on the company's behalf, approve someone, publish something or enter into a commercial commitment.

It can be right and still lack the authority to decide.

That distinction becomes more important as models get better, not less. As long as AI is unreliable, human control feels self-evident and nobody has to defend it. Gates himself notes that the big shift only comes once AI works almost flawlessly, because then there no longer seems to be a practical reason to have someone keep an eye on it. At company scale the same risk sits there. Not necessarily that someone consciously decides to hand over decision-making authority, but that it gradually shifts along with the reliability, because it becomes easy and nobody has drawn the line explicitly.

A human in the loop still doesn't say where the authority lies

"Human in the loop" sounds reassuring, but on its own it says little. There can still be someone pressing a button at the very end while the system has already framed the question, narrowed the options, made the assessment and pushed one answer forward as the obvious one. Formally there was a human in between. In practice that person does little more than confirm what already seems decided, certainly the tenth time in a day, when the previous nine were correct.

That is not an argument against humans in the loop. It is an argument for asking the question differently: which decision was never delegated to the system, and does the person taking it still genuinely have a say in it?

What may the system propose? What may it carry out itself? Which step changes the status of information? Which step creates a consequence outside the company? And who has the mandate to approve that transition?

That should be settled before the automation, not at the moment someone wonders afterwards why the system did something.

A request is not yet an approval

In an Ambassador Journey we developed ourselves, a lot is automated. People submit their profile and product requests, the application checks whether the required data is present, keeps status and history, shows what still needs attention and prepares the next step. A daily operational overview stops open items from getting scattered across files and mailboxes again. AI and agents helped with research, architecture, development, testing and review, which lets the app take over a large part of the preparation and follow-up reliably.

The commercial layer is very much part of that Ambassador Journey. Assessment, product choices, approval and the final handover to Sales belong to the same process. It's just that those steps don't all have the same owner.

The Journey may record that someone has submitted a profile or request. It may check whether the information is complete and show which assessment is still needed. It may prepare an approved selection for handover.

But a request never becomes a business approval by itself. The app doesn't independently determine which commercial assessment applies, and it doesn't create a definitive Sales handoff as long as someone with the right authority hasn't explicitly taken that step. The request itself creates no order, contract or commercial commitment.

Software could technically assign a next status itself without much trouble. That is just not the same as being authorised to take the underlying business decision. So the boundary isn't only in what the app can do, but in which status transitions it may carry out on its own and which require an explicit human decision.

Some decisions contain context that doesn't fit entirely into input fields: a relationship, an exception, a commercial trade-off, the responsibility for what happens next. And in the end someone has to be authorised, on the organisation's behalf, to say yes or no, also towards the person who made the request.

Truth and authority are two different questions

I wrote earlier that better AI forces us to determine more precisely what is true: what is a hypothesis, what is a decision and what is the current project truth.

That remains necessary. But suppose it is all in order. The source is right, the information is current and the model interprets it correctly. Then the question of whether the system may act on it is still open. It isn't answered by the quality of the information.

One determines what counts as true. The other determines who or what may do something with it. They belong together, but decision-making authority can't be derived from correct information.

This goes by the name authority governance, which sounds grander than it is.

Fixed limits allow more automation

Human Reserved sounds like you want to protect something against automation. In practice it tends to work the other way round. Once it's settled which status transition, decision or external action requires human authority, the system can work far more independently up to that line, without asking permission at every intermediate step. Human attention goes to the few moments where it is really needed, and those moments are named.

That doesn't mean a system may never decide. An automatic reorder below a threshold amount is a decision too, and it has been delegated. It's just that this happened deliberately, with a limit attached, and everyone knows that's how it is.

That seems more useful to me than handing out as much autonomy as possible and reconstructing afterwards where the decision was actually taken.

What AI can do will keep shifting, and what is impossible today becomes an ordinary feature tomorrow. Authority doesn't move along automatically. Someone has to hand it over or deliberately keep it, and the difference between the two is usually no more than an agreement made in time.