An agent is a role, so it gets a lane
When people start thinking about an AI agent in their process, the first question is usually "which steps can it do?" That is the right question, and the diagram answers it better than a list. In BPMN, a lane is a role: a set of steps performed by the same actor. An agent that reconciles data, drafts a reply or prepares a decision is an actor like the handler or the approver, so it gets a lane like them.
That one decision does a lot of work. Every step in the agent's lane is a step the agent performs. Every arrow into the lane is a hand-over to the agent, with whatever data it needs attached. Every arrow out is a hand-over back to a person or a system. And every control step that stays in a human lane is a point where a person decides, which is exactly the guardrail the page on processes before AI agents argues for.
A lane is a role. Giving the agent a lane makes visible what it does, what it receives and where it stops.
Start from the As-Is, and keep it one switch away
The To-Be is not designed on a blank canvas. It is designed from the As-Is, the process as it runs today, captured with the people who run it. That matters for two reasons. First, the steps that are candidates for the agent are only visible once the real process is drawn, including the exceptions and workarounds that never made it into the handbook. Second, the people who will work next to the agent need to see what changes for them, and the easiest way to show that is the To-Be as its own diagram with the unchanged As-Is one switch away, lane by lane.
In patchley, the To-Be is a second diagram derived from the As-Is, and the two stay side by side. The As-Is is not overwritten; it remains the record of what was agreed, and the To-Be is the proposal of what should change.
What goes into the agent's lane
We have found a short set of questions useful for deciding which steps move.
- Is the input available as data? An agent can reconcile two records, check a document against a list or draft a text from a case file. It cannot phone a customer back about a missing signature. Steps whose inputs are documents, records and messages are candidates; steps whose inputs are conversations usually are not.
- Is the rule stated? A step that follows a stated rule ("if the amount is below the limit and the account is active, approve") can sit in the agent's lane. A step that follows "experience" cannot, until the experience has been written down as a rule or the step has been split into the part that is a rule and the part that is a judgement.
- What happens if it is wrong? Steps where an error is cheap to catch and cheap to correct move first. Steps where an error is expensive stay with a person, or get a control step after them.
The first To-Be usually moves fewer steps than people expected, and that is fine. The lane can grow in a later version; the diagram keeps every version, so the path from the first To-Be to the third is visible.
Where the control steps go
The agent's lane is surrounded by control steps, and their position is the design. A control before the agent's step decides whether the case goes to the agent at all: the gateway "standard case?" with its "no" path to a person. A control after the agent's step is where a person approves what the agent prepared before it has an effect: the four-eyes approval before a payout, the review before a reply is sent.
Both are ordinary control steps, marked as such in the diagram, with their objective, reference and evidence in the compliance analysis. That is deliberate. An audit of the To-Be process asks the same three questions about a control around an agent as about any other control, and the diagram answers them the same way. The post on control steps that survive an audit has the detail.
The agent's lane is bounded by control steps. Where a person decides is a modelling decision, and it is visible in the diagram.
What the lane does not do
A lane for an agent in a To-Be diagram is a design. It describes where the agent would work, what it would receive and where a person stays in charge. It does not make the agent run. Today, patchley captures the As-Is, derives the To-Be with you, and exports both as BPMN 2.0 XML for the platform where the automation is built; the step from the signed-off model to a running process is the Executions pilot, in preparation, not a released feature.
That separation is useful rather than limiting. The To-Be with its agent lane is a document the department, compliance and IT can agree on before anything is built. The example model with an agent lane on the BPMN-with-AI page shows the shape: three lanes, one for the agent, a gateway that routes the non-standard case to a person, and a control step before the result leaves the process. Agreeing that picture first is cheaper than discovering the missing control after the agent has been running for a month.