Redesign around the decisions in the process, not the screens people click through. List each decision, the facts it needs and where they come from, how hard it is to undo, and who answers for it. Then give each decision a depth, place a person where consequences are hard to reverse, and design the exception route before the happy path.
Why not automate the process as it is?
Most processes were shaped around the limits of people and old systems: a form exists because someone once needed to copy data between screens, a second check exists because the first one was unreliable. Automating each step faithfully keeps every workaround and adds an agent on top.
This is where agents differ from robotic process automation. A script follows fixed steps and breaks when the screen changes. An agent can handle variation, so the useful unit of design is not the step but the decision: what is being decided, on what evidence, and with what authority.
What are the steps in a redesign?
- Process owner: maps the decisions. For each one, write what is decided and what happens next. Many processes turn out to hold only a handful of real decisions among dozens of steps.
- Analyst: traces the facts behind each decision. Note whether each fact comes from a system of record, a named party, the customer, or an outside document. A decision that rests on an emailed attachment needs a different control from one that reads the ledger.
- Process owner and risk: rate reversibility. Can the action be undone cheaply, with effort, or not at all?
- Owner: assigns a target depth per decision, from read-only to execution within bounds. This is a target; the agent still has to earn it.
- Owner: designs the exception route. What happens when facts conflict, a document is missing or the case is outside scope? Who receives it, and how fast? The guide on handling exceptions an agent cannot process covers the patterns.
- Analyst: documents the new process so each decision, its depth and its owner can be read by an auditor. See how to document an AI agent's process.
What should the agent do and what should stay with a person?
| Kind of decision | Agent's likely role | Stays with a person |
|---|---|---|
| Gathering and checking facts | Reads and compares, depth 1 | Spot checks |
| Routine classification or routing | Executes within bounds, depth 3 | Reviews misroutes |
| Customer-facing wording | Drafts, depth 2 | Sends |
| Anything resting on an unverified outside document | Flags it | Verifies the source |
| Exceptions to policy | Drafts a recommendation | Decides |
| Actions that move money or cannot be undone | Prepares | Approves |
What does a redesign look like in practice?
Take the onboarding of new business customers at a commercial lender. The old process has 34 steps across three teams, including re-keying company details from an application form into two systems and a manual check that the names match.
Mapped by decision, it has seven: is the application complete, does the company exist as stated, who are the controlling persons, are sanctions checks clear, what is the risk rating, is anything missing, and is the file ready for approval. The re-keying disappears because the agent reads the company register directly. The name check becomes part of "does the company exist", and its source is the register, not the form.
The agent takes the first two decisions at depth 3 and drafts the risk rating at depth 2. Sanctions matches and final approval stay with people. The exception route sends any mismatch between the form and the register to a named analyst within the same day. Working through new customer onboarding files with an AI agent goes further into this flow.
Which controls make the new process safe?
- One owner for the whole process, not one per team. Good owners matter more once agents are involved, which is why agents raise the value of good process owners.
- A filed verdict for every decision the agent makes or prepares, naming the bound that decided it.
- Provenance on every fact, so a reviewer can see which facts came from a system of record and which from an attachment.
- A dated review of the design after the first month, using the held queue and exception volumes as evidence.
The broader sequence is in our guide to deploying an AI agent.
Questions people ask
Should we redesign before or after choosing the agent?
Before. The decisions, sources and reversibility are properties of your process, not of any tool. Doing the map first also tells you which capabilities you actually need.
Does a redesign mean fewer human checks?
Often fewer, but better placed. Checks that existed to catch copying errors can go. Checks at decisions that move money or affect customers usually stay, and become more focused.
How long does a redesign take?
For a single process, a few weeks of the owner's and an analyst's time. Most of it is mapping decisions and agreeing the exception route, not building.