How to redesign a process for an AI to run

They dropped an agent into the middle of the process and kept doing exactly the same thing. Except now there is one extra step: someone copies what the agent says into the usual spreadsheet, just in case.

Carlos Andrés Ramírez · 2026-08-21

"Redesign before you automate" is the line everyone repeats. Vendors, boutiques, syndicated press releases. They say it with the confidence of someone handing over a truth, and what they actually hand over is a slogan. It sounds like a method and is not one, because it never says what gets decided first, who decides it, or against what criteria.

So companies do what anyone does when handed a slogan with no instructions: they interpret it however they can. They pull up the process map, find the box where the AI fits, and drop it in. Same approvals, same controls, same formal owner, except one step now runs on a model instead of a junior analyst. That is not redesign. It is a coat of paint over a skeleton built for a person with judgment, fear of getting it wrong, and a boss to answer to.

The symptom

How do you redesign a process for an AI to run part of it?

The people asking this run a process with real volume, not a demo. Billing, claims, vendor onboarding, reconciliation. And every search turns up the same slogan under a different logo. Nobody says what criterion makes a task delegable, what threshold triggers escalation to a human, who signs the output, or how that step holds up in front of an auditor who shows up eight months later asking who approved what.

One step got automated, but the review sitting right after it is still built for a human's error rate, not a model producing a hundred outputs an hour.
Nobody wrote down which decisions the agent can make on its own and which have to go back to a person before they execute.
The audit trail assumed a human signature, and now that signature either disappeared or gets stamped by someone who read nothing before signing.
Exceptions escalate to whoever happens to be online, not to whoever actually has the authority to rule on that exception.
The metric claiming the process "improved" measures output speed, not the cases now failing quietly because nobody reviews them.
The process map still has all the old boxes, with the word AI scrawled inside one of them by hand.

The problem underneath

Redesign is not drawing the same process with one new box.

Real redesign starts before anyone looks at what technology to buy, and it starts with a question almost nobody asks out loud: of the tasks that make up this process, which are delegable and which are not? The test is not "can AI technically do this." It almost always can. The test is whether the outcome can be undone at no cost when it is wrong, whether the error costs little or costs a lawsuit, and whether the input data is structured enough for the agent to act without interpreting a grey area.

There is something even less comfortable underneath that. Redesigning a process means redistributing who holds authority inside it, and that touches people with a name and a title to defend. The middle manager who currently approves every exception loses that power the moment the escalation rule gets written down and automated. That is exactly where real redesign stalls and cosmetic redesign never does: in the conversation about who stops deciding what they decide today.

Dropping an agent into a process built for a person to run is not redesigning the process. It is fitting a copilot onto a process that is going to keep crashing exactly the same way, just faster.
BECOME

The framework

What has to be decided before you touch the process?

Reversibility: If the task's outcome can be undone at no cost when it is wrong, delegate it. If undoing it means a refund, an apology to a customer, or a legal correction, the agent proposes and a person executes or approves before the effect becomes irreversible.
Cost of error: Not the average cost, the reasonable worst case. A figure and a threshold written down before anything gets delegated: below it, the agent decides alone; above it, the decision goes back to a named person, not a generic review queue.
Data structure: If the input is a form with clear rules, the agent acts. If it requires interpreting an exception, a contract nuance, or a situation nobody anticipated, that task is not delegable yet, no matter how well the model handles it in testing.
Signature: Who is accountable for the outcome to whoever questions it afterwards: a name and a title, not "the system" or "the team." If nobody is willing to put their name against that box, the task is not ready to be delegated.
Trail: What gets logged so that, months later, someone can reconstruct why the agent decided what it decided: what data it saw, what rule it applied, whether it escalated and to whom. Without that, an auditor has nothing to audit, just an outcome.

Take the process you were about to redesign and, before looking at any tool, list every task inside it. For each one, write an answer to the framework's five questions. The ones that come back green on all five get automated first. The ones that fail even one keep a human in the loop, and that list, not the pretty diagram, is the actual redesign plan.

Frequently asked questions

How do you redesign a process for an AI to run part of it?

Before touching any tool, break the process into tasks and decide, task by task, which are delegable against four tests: whether the error is reversible, what the worst case costs, whether the input data is structured, and who signs the outcome. The redesign is that written decision, not the box where AI gets inserted into the diagram.

What tasks can a company delegate to an AI agent?

Ones that combine low cost of error, reversible outcomes and structured input data: classifying a request, extracting data from a standard document, or drafting something a person then reviews. Tasks with costly errors, irreversible outcomes or frequent exceptions keep a human in the loop, even when the model is technically capable of handling them.

Who signs the output of a process an AI partly runs?

The same person who would sign it if a human team had run it: whoever has authority over that process and answers for it to their boss or to an auditor. Delegating execution to an agent does not delegate accountability; if nobody is willing to put their name on the outcome, that task was not ready to be automated.

How do you audit a process an AI partly runs?

With a trail that logs, for every decision the agent makes, what data it saw, what rule or criterion it applied, and whether it escalated to a human. Without that trail built in from the design stage, an auditor arriving months later has only a final outcome with no way to reconstruct how it was reached, which turns any review into an act of faith.

Let's redesign the process