There's a moment in every AI project where someone gets excited and says: "We could build an agent for that."
Before you do, ask one question: should that task exist at all?
Most business processes weren't designed. They accumulated. A step was added after a mistake in 2019. A sign-off was added when a new manager arrived. A spreadsheet was created to feed a report that fed a meeting that nobody remembers asking for. Years later, everyone just does it.
Point AI at that process and you get an agent that runs an outdated habit faster. You've paid to automate the problem.
Replicating vs. redesigning
Replicating asks: "How do we do what we do now, faster?"
Redesigning asks: "What outcome do we actually need, and what's the shortest path to it?"
That's jobs-to-be-done thinking. Nobody wants a weekly report. They want to know whether they can make payroll, which deals are at risk, where to put the next dollar. The report was always just a delivery mechanism, and often a bad one.
The five steps we use
In the AI Workflow Reinvention Lab, teams work through a process in five steps. You can run a light version yourself.
1. Target the problem. Name one workflow. Define the trigger (what starts the clock) and the Definition of Done (the exact output, e.g. "PDF in the CEO's inbox by Monday 8 a.m."). Then name the stakes: why do we do this, and what happens if we don't? Don't solve it yet. Just name it.
2. Deconstruct it. List every step, who does it, and how many times it loops. Add up the active time per person, plus the "dead air" spent waiting. If the total doesn't match the real cycle time, you've found a leak.
3. Stitch it together. Compare everyone's version. You'll be surprised how differently people describe the same process. Look for the Swivels (data copied between systems), the Loops (chasing people) and the one person who holds it all together.
4. Blue sky it. Ask the uncomfortable questions:
- "If we had to deliver the Definition of Done in 10 seconds, how would it happen?"
- "If sales volume doubled and we couldn't hire, how would this work?"
Then go back through your steps and label each one Legacy ("we've always done it") or Critical (regulatory, or genuinely needed). The legacy steps go on a list called Steps to Kill.
5. Create the AI Blueprint. Only now do you bring in AI, and only for the steps that survived. Give each one a role:
- TALK (Co-Pilot): helps a human think: drafting, summarising
- DO (Worker): completes a task: fetching, formatting, updating
- RUN (Agent): makes a bounded decision and acts: checks, chases, sends
Some steps get no AI at all. That's a good outcome.
What usually happens
When teams do this honestly, two things happen.
First, a third or more of the steps disappear before any technology is involved. Approvals that nobody reads. Re-keying that exists because two systems were never connected. Reports that feed other reports.
Second, the AI that's left is smaller, cheaper and more likely to work, because it's doing a clear job with a clear trigger and a clear definition of done.
The test before you build
Before any agent gets built in your business, ask three questions:
- What's the Definition of Done, in one sentence?
- Which of these steps are legacy habits we could simply stop?
- Is the process rules-based, or does it run on gut feel? If it's gut feel, fix the procedure first. AI will only scale the chaos.
If you can't answer all three, you're not ready to build. You're ready to redesign.
Want to work through this with your leadership team? Bring the AI Workflow Reinvention Lab in-house, or join us at the AI for Executives Bootcamp in Madrid.