Automate the measured leak. Not the demo.
Automate the workflow with the biggest measured leak behind it — the most expensive thing your team does by hand, not the most impressive thing you've seen AI do. The test is arithmetic, not excitement: hours per week the task eats, times a loaded hourly rate, times 52 weeks. Run that number for every manual workflow you've got. The biggest one a system can own end to end goes first. Everything else waits.
That answer disappoints people. They arrive wanting the thing they saw in a demo — the chatbot, the voice agent, the dashboard that talks back. I understand the pull. But a demo is a showroom and your operation is a job site, and the first build has one job: pay for itself fast enough that you trust the second one.
Impressive demos make expensive first builds.
Demos are selected to look good in ten minutes. A chatbot on your website. A dashboard that answers out loud. Nothing wrong with either — later. As first builds they fail for boring reasons: you can't measure what they save, they touch your customers before they touch your costs, and by month three nobody owns them.
The right first build is the opposite. It moves the order data, chases the invoice, assembles the Monday report. Nobody screenshots it. It quietly hands payroll hours back every week — and it proves the whole approach with a number instead of a feeling.
| The demo build | The first build that pays | |
|---|---|---|
| What sells it | The wow in a ten-minute demo | A number: hours × rate × 52 |
| What it touches | Your customers, on day one | Your costs, where mistakes are cheap |
| How you measure it | You mostly can't | Hours back, visible in payroll |
| Month three | Usage fades; nobody owns it | Boring. Still running. |
| What it proves | That AI is impressive | That automation pays, in your numbers |
Four leaks it usually turns out to be.
Run the list exercise on a mid-size operation and the same four workflows keep landing at the top. If you don't know where to look, start here.
Data typed twice
An order lands in email, gets keyed into the ERP, then keyed again into the invoicing tool. Three systems, one set of facts, and a person paid to be the copy machine between them.
Collections that run on memory
Invoices get chased when someone happens to think of them. The aging report gets honest attention at month-end — and the cash sits out there the whole time.
The hand-assembled report
Every week someone senior exports from three systems, pastes into a spreadsheet, and formats. The numbers are stale before the meeting they were built for.
Approvals waiting on an inbox
The quote is ready, the PO is ready, the change order is ready — all of it parked behind one unread email. Minutes of work, days of waiting.
All four share the same anatomy: defined inputs, rules a careful new hire could follow, a clear definition of done. That anatomy is what makes them automatable — it's the pattern I build against in workflow automation engagements.
Score every leak. Then pick one.
- 01
List the manual work. Every task your team does by hand more than once a week — quoting, invoicing, reporting, scheduling, re-keying. No filtering yet; the list is the point.
- 02
Price each one. Hours per week × loaded hourly rate × 52. That's the annual leak. The ROI calculator does the multiplication for you — no email gate.
- 03
Pick the biggest leak a system can own end to end. Not the biggest number outright — the biggest one with defined inputs and a clear finish line. Partial automation leaves the babysitting in place.
- 04
Automate judgment last. Pricing exceptions, tense customers, credit calls — the steps that need a human stay human the longest. Automate the routine around them first.
That's the whole framework, and you can run it yourself this week with a spreadsheet and an honest hour. The operations audit is the done-for-you version — same math, run against your real numbers, with a ranked roadmap at the end instead of a hunch.
When the answer is don't. Yet.
Some workflows fail the framework, and forcing them is how automation gets a bad name inside a company. Three signs to leave one alone:
- The process changes weekly. Automating a moving target means rebuilding it every month. Stabilize it first; automate the version that holds still.
- Nobody owns it. A system without an owner becomes shelfware by quarter two. If no one on your team wants the output, don't build the machine.
- The volume is too small. An hour a month isn't a leak — it's a rounding error. A checklist is cheaper than a system.
And if you run the math and even your biggest leak is small — don't hire me. You don't have an automation problem yet. Keep the number, re-run it when volume grows, and spend the money where it works harder today.
Keep reading: What AI automation actually costs puts real numbers on audits, builds, and ongoing engineering. Or go straight to how the operations audit works.