Practice: Mapping customer workflows and choosing what to automate with AI
Generated by Claude through GigAI 0.1.11.10 on 2026-10-11.
Main tool: BPMN 2.0 · BPMN 2.0 is the one notation in this lesson that both the customer's process owners and an engineer can read, so the same diagram carries the as-is discovery, the automation decisions, the human checkpoints and the to-be implementation. · Alternatives: n8n and Zapier build and run the automation directly from connector nodes but have no standard notation for the as-is process; Temporal writes the workflow as durable code that non-engineers cannot read; Miro and Lucidchart draw free-form diagrams quickly but with no execution semantics; Camunda Modeler is the editor used here to draw BPMN 2.0, and the Claude API is what an automated task would call, not a way to describe the process.
Every step links into the tool's own documentation; links were checked on 2026-10-11 · version 1
Setup (one time)
1. Install the free desktop Camunda Modeler on your laptop and create an empty BPMN diagram.
A real BPMN editor enforces the notation rules as you draw, which a whiteboard tool does not. The desktop modeler runs locally with no account and no cloud.
Explained in: Camunda Modeler | BPMN Process Modeler | Camunda (camunda.com), checked 2026-10-11
Done when: The modeler opens a blank BPMN canvas showing a single start event and the symbol palette.
2. Read the plans page to confirm which parts are free to use on a laptop and which need a paid plan, and decide to stay on the free local modeler for this path.
Knowing where the free tooling stops keeps the practice laptop-only and lets you tell a customer what executing the model later would cost.
Explained in: Camunda Pricing | Process Orchestration Plans | Camunda (camunda.com), checked 2026-10-11
Done when: You have a two-line note naming what you will use for free and what you are deliberately not using.
3. Read the basic concepts of events and keep the symbol guide open beside the modeler as your reference for every later step.
Start, intermediate and end events with their catching and throwing forms are the vocabulary every other symbol builds on.
Explained in: BPMN 2.0 Symbols - A complete guide with examples. | Camunda (camunda.com), checked 2026-10-11
Done when: You can say aloud the difference between a start, an intermediate and an end event and between catching and throwing.
Steps
1. Create one pool for the customer team whose work you are mapping and a collapsed pool for each outside party they exchange messages with, using notes from a real or role-played discovery interview.
Discovering the as-is workflow starts with who is involved and where the boundary of the process lies. Pools force that scoping decision before any detail is drawn.
Explained in: BPMN 2.0 Symbols - A complete guide with examples. | Camunda (camunda.com), checked 2026-10-11
Done when: The canvas shows one expanded pool and at least one collapsed pool, each named after a participant from your interview notes.
2. Create a lane inside the main pool for every role or system that touches the work today, including the ones people forgot to mention at first.
Lanes expose handoffs between roles, and handoffs are where as-is processes lose the most time. This is the core of discovery.
Explained in: BPMN 2.0 Symbols - A complete guide with examples. | Camunda (camunda.com), checked 2026-10-11
Done when: Every person or system named in your interview notes has exactly one lane and no lane is empty of a named owner.
3. Write each step of the as-is work as a task in the lane of whoever performs it, named verb then object, and add loop or multiple instance markers where the work repeats per item.
Writing the process down with BPMN means one task per unit of work with a clear owner. Markers record volume and repetition, which later drive the automation choice.
Explained in: BPMN 2.0 Symbols - A complete guide with examples. | Camunda (camunda.com), checked 2026-10-11
Done when: The diagram reads from start event to end event as a sequence of named tasks and a colleague can retell the process from it without your notes.
4. Configure an exclusive gateway at every point where the as-is work branches, labelled as a question with each outgoing flow labelled with its answer.
Decisions are where human judgement lives in the current process. Naming them precisely shows which are rule-based and which need interpretation.
Explained in: BPMN 2.0 Symbols - A complete guide with examples. | Camunda (camunda.com), checked 2026-10-11
Done when: Every branch in the diagram leaves a gateway labelled with a question and every outgoing flow has an answer label.
5. Create parallel gateways around the work that happens at the same time today, with a matching join before the flow continues.
Real workflows are rarely strictly sequential, and showing concurrency correctly tells you which steps can be automated independently of the others.
Explained in: BPMN 2.0 Symbols - A complete guide with examples. | Camunda (camunda.com), checked 2026-10-11
Done when: At least one split and its join appear in the diagram and the modeler shows no dangling sequence flow.
6. Create message events wherever the process waits for or sends something to an outside pool, and connect them to the collapsed pools with message flows.
Waiting on an outside party is usually the longest part of the as-is cycle time. Marking it stops you from trying to automate a delay you do not control.
Explained in: BPMN 2.0 Symbols - A complete guide with examples. | Camunda (camunda.com), checked 2026-10-11
Done when: Every arrow that crosses a pool boundary is a message flow attached to a message event or task, and none is a sequence flow.
7. Create a to-be copy of the diagram and annotate each task with its volume, rule clarity and cost of a mistake, then move the chosen candidates to a system lane and tag each with one AI pattern: extraction, classification, drafting or single prompt call.
Choosing what to automate is a decision per task, not per process. High-volume, low-risk, well-specified tasks get a simple pattern, and the diagram should record why each was chosen.
Explained in: BPMN 2.0 Symbols - A complete guide with examples. | Camunda (camunda.com), checked 2026-10-11
Done when: Each task in the to-be copy carries a keep, assist or automate note and every automated task names exactly one AI pattern.
8. Create an ad-hoc subprocess for the one part of the work where the order of steps cannot be fixed in advance, and place inside it the tasks an agent would be allowed to pick from.
An agent pattern is only justified where the path truly varies case by case. The ad-hoc subprocess draws that boundary explicitly and keeps everything else deterministic.
Explained in: BPMN 2.0 Symbols - A complete guide with examples. | Camunda (camunda.com), checked 2026-10-11
Done when: Exactly one bounded region of the to-be diagram is marked ad-hoc and everything outside it follows a fixed sequence.
9. Create a human review task in a reviewer lane after each automated step whose mistake would be costly, followed by an approve or reject gateway, and attach a timer event so that an unanswered review is rerouted.
Placing human checkpoints means deciding where a person must approve and what happens when they do not respond. A checkpoint without a timeout quietly becomes the new bottleneck.
Explained in: BPMN 2.0 Symbols - A complete guide with examples. | Camunda (camunda.com), checked 2026-10-11
Done when: Every high-risk automated task is followed by a review task with a reject path back and a timer path forward.
10. Configure an error event on each automated task that sends failed or low-confidence results to a human handling task instead of the normal flow.
A model call can fail or return something unusable. The to-be design has to say where that work goes so the fallback is a person and not silence.
Explained in: BPMN 2.0 Symbols - A complete guide with examples. | Camunda (camunda.com), checked 2026-10-11
Done when: Each automated task has a visible error path that ends at a named human task.
11. Package the review and fallback pattern as a subprocess and replace the repeated copies in the to-be diagram with it, so that the top level shows only the main flow.
Implementing the to-be workflow in any tool is easier from a diagram whose top level fits on one screen. A reusable checkpoint block is built and tested once.
Explained in: BPMN 2.0 Symbols - A complete guide with examples. | Camunda (camunda.com), checked 2026-10-11
Done when: The top-level to-be diagram fits on one screen and the checkpoint logic appears once, inside a collapsed subprocess.
12. Reproduce the to-be diagram from a blank canvas using only your discovery notes, save both versions as files in a version-controlled folder, and compare them symbol by symbol against the guide.
Redrawing without the first attempt proves you understand the notation and did not just arrange shapes. Versioned files let the customer review each change to the process as a diff.
Explained in: BPMN 2.0 Symbols - A complete guide with examples. | Camunda (camunda.com), checked 2026-10-11
Done when: Two saved diagram files exist in version control, they describe the same process, and you have a written list of every difference with the reason for it.
Back to Mapping customer workflows and choosing what to automate with AI