Teams usually come to us wanting to "automate everything." That instinct is right, but the order matters enormously. Automate the wrong step first and you've just made a bad process run faster. Here is the sequence we run on every engagement, in order, without skipping steps.
1. Map the workflow as it actually runs
Not the version described in the onboarding doc — the version people actually follow, including the workaround someone invented eight months ago that nobody documented. We shadow the process end to end before touching any tooling.
2. Measure before you touch anything
Volume, time per task, error rate, and where the handoffs happen. This step feels slow, but it's what tells you whether a workflow is worth automating at all. Some processes run twice a month and cost nothing to leave manual.
3. Remove steps before you automate them
The fastest workflow step is the one that doesn't exist. Before we write any automation, we look for approvals, duplicate data entry, and status checks that exist only because the process was never redesigned. Automating a redundant step just makes redundancy faster.
4. Automate what's left
Only now do we build. At this stage the workflow is lean enough that automation is straightforward — clear triggers, clear success criteria, clear exception paths. This is also where we decide what needs a rules-based automation versus what genuinely needs an AI agent making judgment calls.
5. Monitor, because workflows drift
- Set an error-rate threshold that pages a human automatically.
- Review edge cases monthly — they cluster and reveal blind spots.
- Re-measure quarterly. Volume and inputs change, and automations age.
Automation without monitoring is just a slower way to discover a bug three weeks late.
Teams that skip straight to step four end up automating their own mess and calling it progress. Teams that run the full five steps end up with workflows that keep getting leaner even after we've left the room.