How does partitioning a UML activity diagram by responsibility expose where a cross-team approval workflow stalls?

## What a partition actually asserts A **partition** — drawn as a labelled lane, and commonly spoken of as a swimlane — groups the actions of an activity diagram by the party responsible for performing them: a team, a role, or an automated service. It is an attribution overlay and nothing more. Moving an action from one lane to another changes *who does it*; it never changes *when* it happens, because ordering lives entirely in the control flow edges between actions. Establish that distinction first, because it is what makes the rest of the reading legitimate. Since lanes carry no ordering, every fact you read off the lanes is a fact about responsibility — and the most useful one follows immediately: **every control flow edge that crosses a lane boundary is a handoff**, a point where work stops belonging to one party and starts belonging to another. Handoffs are where a workflow queues, because the receiving party is almost never idle and waiting for it. ## Reading the crossings The procedure is mechanical and takes minutes: 1. Draw the process with one lane per responsible party and every action in the lane that performs it. 2. Mark each control flow edge that crosses a boundary — those are the handoffs. 3. Attach a measured wait to each crossing from whatever record the work leaves behind. 4. Attach a count: how many items travel each crossing, and how many travel it more than once. 5. Read the picture for the two shapes that matter — the crossing with the longest wait, and the crossing that repeats. ## A worked reading A veterinary practice-management product routes discharge approvals through three parties: the attending clinician, the billing desk and the practice owner. Over one 3-week iteration the team traced 31 approval requests. | Crossing | Requests | Median wait | |---|---|---| | Clinician lane to billing lane | 31 | 1.6 days | | Billing lane to practice-owner lane | 27 | 4.3 days | | Practice-owner lane back to clinician lane (rework) | 12 | 2.9 days | Median end-to-end time was 8.8 days, of which 6.4 days was time spent *between* lanes rather than inside an action. No individual action took longer than about forty minutes of real work. The picture makes two things obvious that a prose procedure had hidden for a year. First, the expensive crossing is into the practice-owner lane, and 27 of 31 requests make it even though only high-value cases need that approval — a decision node with an honest guard would keep roughly nineteen of them inside two lanes. Second, the rework edge is a *backwards* crossing, and every item on it pays two waits, not one. The stress test came from a two-week holiday shutdown, when the practice-owner lane had nobody in it. Throughput did not degrade gracefully; the queue in front of that single boundary absorbed everything, and the diagram had predicted exactly which boundary would do that. ## What partitions do not tell you - **Effort.** An action drawn in a lane may be four minutes of work or four days of it; the box is the same size either way. - **Waiting.** The diagram says a crossing exists, not how long work sits at it. The wait is measurement you bring to the picture. - **Authority.** A lane says who performs an action, not who is allowed to decide it or who is accountable if it goes wrong. - **The artefact.** Control flow says 'then', not 'carrying what'; that is object flow's job. - **The org chart.** Two lanes can be the same team wearing two hats, and one lane can hide three teams. Keep the grain of the lanes consistent or the crossings are not comparable. ## Object flow: the layer that shows what moves An **object flow** routes an edge through an object node — a rectangle, or a pin drawn on the edge of an action — and shows the artefact that passes between two actions: a signed form, a record, a decision. Control flow says one action follows another; object flow says what it hands over. On a multi-party workflow that second layer is where the deep defects live: the same artefact created independently in two lanes, an artefact that arrives in a form the receiving lane cannot use, or a receiving action that genuinely needs two artefacts when the diagram shows only one arriving. ## When the diagram earns its place A partitioned activity diagram beats a written procedure when the process has more than two responsible parties, when routing is conditional or concurrent, or when the team argues about what the process even is — a picture that everyone can walk with a finger settles that in one meeting. Prose is the better medium when the process is linear with a single owner, or when the difficulty is in the *rules* rather than in the sequence: eligibility thresholds and legal wording read far better as text than as guards hung on edges.