Which Workflows Are Worth Giving to an AI Agent

Most failed automation projects did not fail on the technology. They failed because somebody automated the wrong thing.

The wrong thing usually looks great in a demo. It is visible, it is annoying, and everyone agrees it is a waste of time. Then you automate it and discover it happened four times a month, took nine minutes, and the person doing it was also quietly catching errors nobody had documented.

Here are the four tests we run before agreeing to build anything.

1. It happens often enough to matter

Frequency beats complexity. An agent that saves four minutes on something happening thirty times a day is worth far more than one saving two hours on something quarterly.

Count honestly. Not "it feels constant" — actually count a week. Most teams are surprised in both directions: the thing everyone complains about turns out to be rare, and something nobody mentions turns out to be eating an afternoon a week.

If it happens less than a few times a week, an agent is rarely the answer. A checklist or a scheduled script usually is.

2. The decision rules can be written down

This is the one that kills most candidates.

Ask the person who does the task to explain how they decide. If they can give you rules — "if the invoice is over five thousand it needs a second approval, unless it is a repeat supplier from the approved list" — you have something an agent can hold.

If the answer is "you just know", stop. That is not a bad answer; it usually means real expertise. But that expertise is not written down anywhere, and an agent cannot infer it from three examples. You would be automating a guess.

The useful middle ground: agents are good at handling the eighty percent that follows rules and escalating the rest. That is often the whole win. You do not need to automate the judgement — you need to stop the judgement being buried under routine.

3. Being wrong is survivable

Ask what happens when the agent gets it wrong at three in the morning with nobody watching.

Some answers are fine. A misrouted support ticket gets rerouted. A summary that missed a nuance gets corrected by whoever reads it.

Some are not. A payment sent to the wrong supplier. A message to a customer that should have had a human read it first. A production change nobody approved.

That does not rule those workflows out — it changes the design. High-consequence steps get the agent doing everything up to the decision and a human making it. Drafted, not sent. Prepared, not executed. Most of the time saved is in the preparation anyway.

4. The systems can actually be reached

The least glamorous test, and the one that most often decides the timeline.

An agent has to read and write somewhere. If your order data lives in a system with a real API, this is straightforward. If it lives in a desktop application from 2009 with no integration surface, or in a spreadsheet on someone's machine, the honest cost of the project is mostly plumbing.

Worth knowing before you start, not six weeks in.

What good candidates look like

Put together, the pattern is fairly consistent. The workflows worth automating tend to be:

• High frequency, low individual value. The thing nobody wants to be assigned.
• Rule-following with clear exceptions. Not judgement-free — judgement-separable.
• Reversible, or reversible up to the last step. Draft, don't send.
• Sitting between two systems. Most of the work is moving and reshaping information, which is exactly what agents are good at.

Watching an inbox and turning it into structured, routed, actionable work fits all four. So does reconciling two systems that were never designed to talk. So does monitoring something continuously and only speaking up when a threshold breaks.

The workflows to leave alone

For balance, the ones we regularly advise against:

• Anything where the real problem is that the process is broken. Automating a bad process gives you a faster bad process.
• One-off migrations. Write a script.
• Work where the person doing it is also the person catching the errors. Automate that and you lose the error catching silently.
• Anything requiring judgement you cannot articulate. Revisit it once you can.

Start with one

The most common mistake after picking the wrong workflow is picking too many. One workflow, running properly, teaches you more about what agents are worth to your business than a roadmap of twelve.

Pick the one that survives all four tests, build it, run it for a month, and see what it actually changed. That answer is worth more than any estimate.

---

If you have a workflow in mind and want a straight answer on whether it passes these tests, tell us about it. The scoping call is free, and if we think an agent is the wrong tool we will say so.