# Why institutions need decisions their software can use

A workflow finishes. A field says "approved." Someone later asks what was decided, on what basis, and whether it still holds. The answer is in the ticket, the emails, the policy and the memory of the person who handled it. The software recorded that something happened. What made it a decision the institution could rely on stayed largely with the people.

## Most consequential acts rest on several decisions

A payment, a hire or a purchase can look like one approval. It rests on several questions. Each needs its own evidence and judgment, and settling one does not settle the others.

## People hold the decisions together

The people doing the work know which questions this case raises. They know what counts as enough: which quotes, which checks, whose sign-off. They know when a question has been answered and when it is still open. And they know what the answer allows: this order, to this supplier, up to this amount.

The institution sets a standard for when those judgments may be relied on, and the people doing the work apply it. That is what turns them into decisions the institution is prepared to rely on. The software records the evidence, the reviews and the process. What makes the result a decision the institution can rely on stays largely with the people.

## AI changes the shape of the work

Applying that standard depends on a person standing between the steps.

AI at work is often pictured as an agent taking over one person's job, with a person reviewing the result. The change goes beyond replacing people one-for-one. When gathering evidence and assessing a case get cheaper, more work is worth doing, and more of it starts without a person. More agents contribute judgment, and more agents, services and machines can act. The system that carries out an act may have taken no part in deciding it.

An agent can produce a convincing assessment while a required check is still missing. The institution has to tell that assessment apart from a decision it is prepared to rely on. As the work passes between agents and systems, the software has to be able to tell them apart too.

## Intelligence can answer the question. The institution decides what it takes to rely on the answer.

The obvious alternative is to reason again each time: give a capable model the policies, the evidence and the history, and let it work out what to do. That produces an answer, often a good one. It does not establish that everything the institution requires was considered, that the evidence was enough for the act, or that the act taken is the one the answer supports.

An answer can be right for the wrong reason. An agent approves a supplier on a misreading of the contract, and an unrelated fact makes the approval correct anyway. The institution should still refuse to rely on that reasoning. A correct outcome from a defective process teaches the institution that the machinery is reliable when it is not.

Judgment can come from a person, a model or a rule. What it takes to rely on that judgment belongs to the institution. A more capable model is entrusted with more consequential work, so that discipline has to apply to more work. Institutions did not remove signing authorities because the people holding them were capable.

## An answer becomes a decision when the institution's standard is met

As the work passes to agents and systems, the institution has to set that standard in a form its software can apply: which questions have to be answered before the work proceeds, what evidence each one needs, whose judgment counts, and what may stay uncertain.

When the standard is met, the answer becomes a decision the software may rely on. Until then the question stays open, and it is clear what is still missing. A decision can be no. Deciding not to proceed is different from still waiting for something required. Meeting the standard does not prove the answer right. It establishes what the institution can stand behind, and what a later person or system can check.

## Once established, the decision is something software can use

The decision keeps what was decided, what it rested on, and the conditions for relying on it later. The ticket closes, the agent session ends, the person leaves, the application is replaced, and the decision remains.

Other software can then check whether a required decision has been established and still applies, without working through the judgment again. Identity establishes who is acting, and permission sets the bounds within which they may act. Whether this action is the one that should happen depends on the decisions the institution requires.

## What we are building

We are building Ordinant so an institution can set what it takes to rely on an answer and establish the decision when that standard is met. People and software can then use the decision in the work and the actions that rely on it.

People, agents and machines can all contribute to the work. The decision belongs to the institution.

[What Ordinant does](/what-it-does)