When is an AI agent the wrong answer?

When the process is undocumented, the work is rare, the judgement is high, the stakes are serious without a design to match, or a system you already pay for does most of it.

Start from this question

Free · a few minutes · starts from “I think AI could help”

The short answer

An agent is the wrong answer when the process is not written down, when the work comes round rarely, when the judgement involved is high enough that a person should keep it, or when a system you already run does most of the job. It is also wrong — in its self-built form — when the information is sensitive or a mistake is expensive and nobody has designed the approvals, permissions and logging first. In those cases the honest choices are ordinary automation, AI assisting a person, or not building anything yet.

Agents are the most capable and most expensive thing in this category, which makes them the most commonly reached for too early. The useful counterweight is a clear list of the situations where they are not the answer.

None of these is a claim that agents do not work. They are the conditions under which something simpler works better, which is a different and more actionable statement.

Is it really this?

Worth checking before you spend anything. These are the signals that tend to tell the two apart.

Signs it is

The process is not written down anywhere.
The work happens monthly or less, so there are few chances to earn back the setup and few to notice it going wrong.
Nobody can say what it would cost if the agent acted wrongly.

Signs it's something else

The steps mostly repeat but need calls along the way, and the stakes are ordinary.
It runs daily or weekly, is documented, and a person is doing it mechanically.
There is an owner, a process and a clear definition of a good outcome.

What's usually behind it

Genuine enthusiasm about what agents can do, applied before the work has been examined.
A repetitive process that looks agent-shaped but is actually fixed-step automation.
Pressure to have an AI initiative, which selects for visibility over fit.

What people usually get wrong

Building an agent for work that fixed automation would do more reliably and more cheaply.
Treating an undocumented process as an agent problem, and encoding the mess.
Choosing a platform before establishing what a failure costs.

The options, and when each one fits

In no particular order. Which of these fits depends on your situation, not on the question you typed.

Ordinary automation
Fixed steps, little judgement, no need for anything to decide.
AI assisting a person
The judgement is the value and speed is the constraint.
A bounded agent
The steps mostly repeat, need calls along the way, come round often, and the stakes are ordinary.
Build it with help
The work suits an agent but the information or the consequences are serious.

When doing nothing is the right call

If more than one of these applies, not building anything this quarter is the right answer, and writing the process down is the step that changes it. That is a week of work and it makes the next decision straightforward.

What the diagnostic asks

Whether the work is repeatable and how much judgement it needs.
How often it comes round.
What happens when it goes wrong and how sensitive the information is.
Whether software already exists for this work.

“I don't know” is always an answer. Nothing is assumed on your behalf.

The routes it weighs

Every diagnostic weighs all ten routes. These are the ones that tend to matter for this question — a description of the rules, not a promise about your answer.

Process improvement
Change how the work flows — remove steps, name an owner, fix the hand-offs.
AI Assist
Keep a person in charge and give them AI for drafting, summarising, checking or analysis.
AI Agent
Delegate a defined, repeatable workflow to software inside boundaries you set, with human oversight.
Do nothing yet
Record the issue, set a revisit date, and decide what would change your mind.
Specialist support
Bring in a freelancer, consultant or specialist for a defined piece of work.

What changes the answer

Fixed steps with little judgement are ordinary automation, not an agent.
High judgement keeps a person in charge with AI assistance.
Sensitive data and expensive mistakes together make it a build-with-help job rather than a self-build.

Get your own read

A handful of short questions, then the strongest route for your situation, what's also credible, what we still don't know and what would change our view. No account, no setup.