Designing Agent-Driven Workflows People Can Approve, Audit and Trust

AI agents are moving from answering questions to doing work. They draft and send emails, update records, reconcile accounts and file tickets across several systems in one go. That shift changes the design problem. When software acts on someone's behalf, the interface stops being about input and output and starts being about oversight: what is the agent about to do, what did it do, and how do I fix it if it got something wrong?

Teams adopt agents when they feel in control of them. These are the patterns we use to give people that control without burying them in confirmations.

Make the Plan Visible Before Anything Happens

For any task with more than one step, show the plan first. A short, readable list of the steps the agent intends to take, the systems it will touch and the data it will use lets a person spot a misunderstanding before it turns into a mistake. Let people edit the plan directly, removing a step or changing a recipient, rather than having to rephrase their original request.

Keep the language human. A line that says the agent will update twelve customer records in the CRM is far more useful than a list of API calls.

Approvals Sized to the Risk

Asking for approval on every action trains people to click through without reading. Asking for none removes the safety net. The answer is to match the level of approval to the consequence of the action.

• Low risk, reversible actions, such as drafting a document or tagging a record, can run automatically and appear in the activity log.

• Medium risk actions, such as sending an external email or changing customer data, should pause for a single clear approval with a preview.

• High risk actions, such as payments, deletions or anything regulated, should require explicit confirmation and, where policy demands it, a second approver.

Show the approval request with everything needed to decide: what will change, a preview of the result, and why the agent thinks this is the right step. Approving should never require opening another system to check.

An approval screen is only useful if a busy person can make a good decision from it in a few seconds.

An Audit Trail Written for People

Every agent action should leave a record, and that record should be readable by the people who rely on it, not only by engineers. A good audit trail answers who asked for the task, what the agent did at each step, which data it used, who approved what, and when. It should be searchable, filterable by person or system, and exportable for compliance reviews.

Design two views of the same history. A summary view for everyday use, showing the outcome and any items that need attention, and a detailed view for investigations, showing each step with its inputs and outputs. Both should link back to the records that were changed.

Show What the Agent Actually Did

When an agent finishes, the person who asked deserves a clear account of the result. Lead with the outcome, then list what changed, then flag anything that did not go to plan. If the agent skipped a step, made an assumption or hit an error, say so plainly near the top. Hiding partial failures in a long log is the fastest way to lose trust.

• Summarise the outcome in one or two sentences at the top of the result.

• List each change with a link to the affected record and a before and after view where it helps.

• Highlight assumptions, skipped steps and errors, with a suggested next action for each.

Undo, Pause and Take Over

People need a way out. Wherever possible, actions should be reversible from the same place they were reported, with a clear undo for each change or for the whole task. Long running tasks should show live progress with the option to pause or stop. And at any point, a person should be able to take over a task manually, picking up exactly where the agent left off.

Where an action genuinely cannot be undone, such as sending a message to a customer, say so on the approval screen before it happens.

Explain Confidence and Sources

Agents often make judgement calls: which invoice matches which payment, which customer a message is about. Show the reasoning in a sentence and the sources the agent relied on. Where confidence is low, route the item to a person rather than guessing. Over time, the patterns in those low confidence items tell the team where the workflow or the data needs work.

Designing With the Teams Who Will Supervise

The people who approve and audit agent work are rarely the people who commission it. Bring operations, risk and compliance colleagues into research early, test approval screens with real cases, and watch how people review an activity log under time pressure. Their feedback will shape the interface more than any pattern library.

We design interfaces for AI agents and the people who oversee them, from approval flows to audit views. If you are building agent-driven workflows, get in touch, browse our work, or explore our services.