When not to use an AI agent

1. The process underneath is broken

This is the most common failure and the least discussed. An agent executes a process faster and more consistently than a person. It does not repair one.

If enquiries currently get lost because nobody agreed who owns them, an agent will route them consistently — to the same undefined place. Fix the process on paper first. Sometimes that is the whole project, and it costs nothing but an afternoon.

2. The volume is too low

Below roughly 20 to 30 repetitions of the same task per week, the build plus the attention it needs afterwards costs more than the hours it saves. The arithmetic is not close, and no amount of capability changes it.

One exception: when timing matters more than volume. Five enquiries a week that arrive at 11pm and go cold by morning can be worth automating where fifty during office hours are not.

3. The work needs judgment, not execution

Agents are good at gathering, structuring and routing. They are poor at deciding things where being wrong carries consequences — legal advice, a medical recommendation, a credit decision, anything a regulator would want to review.

The useful pattern is narrower than the marketing suggests: let the agent assemble the case and hand it to a person. That handover has to be designed deliberately, including what the person sees and how they push back. Assumed handovers are where this goes wrong.

4. Nobody internally will own it

An agent needs someone who reads the escalations and adjusts the rules when the business changes. That is not a full-time job — perhaps an hour a month — but it needs a name attached to it.

Without an owner, the agent degrades quietly. Prices change and it quotes the old ones. A product is discontinued and it keeps offering it. Nobody notices for a quarter, and then the technology gets blamed for an organisational gap.

5. The rules cannot be written down

Some processes work because an experienced person reads a situation. If nobody can articulate the rule beyond “you get a feel for it”, an agent cannot learn it from a handful of examples either.

The test is cheap: ask the person who does the work to write down what they do. If the result is three pages of clear conditions, automate it. If it is “it depends”, do not.

6. The systems cannot be reached

An agent that cannot read and write where the work lives is a chatbot with extra steps. Some older on-premise software genuinely has no usable interface, and no amount of clever engineering makes that cheap.

Worth checking before anything else: can the systems involved be reached programmatically? If the honest answer is a nightly CSV export, the automation project is a different, larger project than the one being discussed.

So when does it work?

When all four of these hold at once — not three out of four, which is what disappointing projects look like from the inside:

• A defined process that repeats often enough to matter.
• Rules that can be written down by the people who do the work.
• Systems that can be reached programmatically.
• A person available for the cases the agent should not decide.

We publish this list because the alternative is worse for both sides. A pilot that was never going to work costs you 30 days and costs us a reference we cannot use. If reading this talked you out of it, it did its job.