The smallest viable AI workflow

Every AI conversation eventually reaches the same question: “So what should we do first?”

The honest answer is almost always disappointing to hear and powerful to do: the most boring repeated task in the business that has a clear owner and a countable output.

Bigger isn’t stickier

Ambitious first projects fail for structural reasons. They touch many systems, so they need many approvals. They promise transformation, so expectations outrun iteration. They have no natural owner, so they have no natural defender. When they stall, the lesson learned is “AI doesn’t work here,” which is the wrong lesson drawn from the wrong experiment.

Small workflows invert every one of those failure modes:

• One task means one process to understand, one system to integrate, one approval chain.
• One owner means someone’s week visibly improves, and improvements with visible beneficiaries get defended in budget meetings.
• One number means success is a measurement, not a mood.

The selection filter

Score candidate tasks against five questions. A good first workflow scores five out of five:

1. Does it repeat at least weekly? (Volume is what pays back.)
2. Does it consume skilled time? (Automating trivial work is still useful, but the ROI story is weaker.)
3. Is the input already digital and accessible? (If step one is “clean up the data,” the project just grew a quarter.)
4. Is human review affordable and natural? (First deployments should draft, humans should approve.)
5. Can a number be attached to it today, before the project starts? (Baseline first. No baseline, no proof, no expansion budget.)

Typical survivors of this filter: enquiry response drafting, meeting notes into CRM, first-pass research summaries, proposal assembly from past proposals, report data-pull and formatting. None of these make a keynote. All of them return hours every single week.

The two-workflow rule

Once the first workflow has run for a month with its number improving, the organisation has something more valuable than the hours saved: internal proof, in its own context, with its own people.

That proof is what unlocks the second workflow, which is usually slightly bolder, and the third, which often connects to the first. Compounding starts small and unglamorous. Transformation is what it looks like from the outside two years later.

Start with the boring task. Ship it. Count it. Then start the next one.