Episode 11 · September 10, 2026 · Hosted by Sebastian
We Drove Our Own Expense Process Close to Full Autonomy
Our own finance operation, taken apart and redesigned from scratch around AI. We modelled the ideal on a green field: snap a receipt and the agent labels, classifies and matches it against payments, with weekly runs instead of a month-end scramble. The diagram became the spec for our developers, and the process has been running live ever since.
We redesigned our own expense process around AI. It has been running live ever since.
The big takeaway
Giving the agent its own lane turned autonomy from a vague ambition into an explicit design decision.
The moment "pay unpaid invoices" moved into the agent lane, the real question surfaced: do we actually want an agent that sends money? That line, what the agent may do and what it may not, is the most valuable thing in the whole diagram.
What we walked through
- 01
Design it on a green field
Before: receipts on paper, handwritten notes, and finance chasing everyone at the end of the month before the tax advisor deadline. We ignored how it worked and modelled the ideal: snap a photo of the receipt, drop it in a chat, done. The agent labels, classifies and matches it against payments.
- 02
Weekly runs instead of monthly panic
An agent runs every week, collects missing receipts and unpaid invoices, and reminds the right people. The end-of-month scramble simply stops existing.
- 03
The agent lane as the red line
We broke our own best practice and gave the agent a dedicated lane next to the humans. It makes the boundary of autonomy visible, and it is where the hard conversations happened.
- 04
The diagram as the spec
We handed the diagram to our two developers as the specification. Almost no back-and-forth, almost no open questions. The explicitness did the work a requirements document never does.
- 05
Agents that document themselves
Each agent reports what it does back into the BPMN model. The audit trail builds itself, and when something misbehaves you can see exactly where.
- 06
Train on the standard cases, route the rest
Expense rules get complicated fast. Train the agent on the 50 to 80 percent that is routine, and send everything else to human review. Add review steps as compliance demands.
Discuss this episode
Ask a question, share how you would model it, or suggest the next topic.
More episodes
All episodes →Automating a Quote Process Step by Step: From Baseline to Roadmap
We took our own offer process, the one we use to quote every customer, and walked it from as-is to to-be live. Not in one leap but in milestones: an agent drafting offers, offers published on a portal, and contracts triggered on acceptance, each one a change small enough to hand to an implementation team.
Most BPMN Diagrams Get This Wrong
Why mixing departments, systems and roles in the same view quietly breaks a process landscape. We tied each view to a modelling level: on the business level, lanes show responsibilities; on the system level, systems become pools and message flows show their interactions. And we landed on one principle: consistency beats correctness.
Where Should AI Go First? Prioritizing Processes and Tasks With Attributes
How to prioritise AI initiatives from your process models instead of case by case. We ran two passes: on the process level, two simple attributes and a filter produced a shortlist; on the task level, value potential against feasibility picked the first initiatives. In our example that left five concrete opportunities.