When an Agent Is the Wrong Answer

This site is about AI agents and takes no money from anyone selling them, which makes this the article worth writing first: most tasks people are considering agents for would be better served by something simpler.

The test

An agent earns its cost when the task varies enough that a fixed script fails.

For a vendor-side perspective on handling multiple clients, see this page from Monitask.

If you can write down the steps, write down the steps. Automation is cheaper, faster, deterministic, auditable and does not need a governance programme.

If you cannot write down the steps because the situation differs every time — and there is real tolerance for occasional wrong answers — an agent may be right.

Everything below is a version of that test.

When automation is better

The process is stable. Same shape every time, with known variations.

You need the same result every time. Deterministic beats capable when consistency is the requirement.

Errors are expensive or hard to detect. Automation fails visibly; an agent produces a plausible wrong answer.

The volume is high and the margin per transaction is small. Per-call model costs matter at volume. See what agents cost to run.

Regulators will ask how a decision was made. "The script does this" is an answer. "The model decided" is a problem.

Buying an agent for a scriptable process is the commonest expensive mistake in this field. You pay for a loop, accept non-determinism, and inherit security and governance work — for something a scheduled job did.

When an assistant is better

A person is going to review the output anyway. If every result needs human approval, autonomy adds nothing and adds risk.

The task requires judgement that is yours to make.

It is one person's work rather than a process. Drafting, summarising, researching, thinking through. See agents for a one-person business.

You want the person to stay engaged with the material. Delegation to an agent removes the understanding along with the work, which matters for anything you will be accountable for.

When nothing is better

Worth saying, because "do nothing" is rarely on the options list.

The underlying process is broken. Automating a bad process makes it fail faster. An agent on a bad process fails creatively.

The data is not there or not clean. Data hygiene is repeatedly identified as the limiting factor — an agent working with incomplete, incorrect or siloed data is restricted by exactly those limitations. Fix the data first; you may find you no longer need the agent.

Nobody has defined what success means. Organisations that fail to set clear success metrics before deployment struggle to show value, and the project gets cut when results look ambiguous.

The problem is organisational. Approvals that take three days, unclear ownership, a team that does not talk to another team. Software does not fix these and it does provide something to blame.

You are doing it because everyone else is. Reasonable pressure, poor basis for a deployment.

Specific cases where agents go wrong

Anything where being wrong is catastrophic and undetectable. An agent that is right 97% of the time is excellent, and unsuitable if the 3% causes irreversible harm nobody notices.

Long unsupervised chains. Error compounding is arithmetic — 95% per step is 60% across ten. See what agents are still bad at.

Work requiring accountability for the judgement. Not because the agent cannot answer, but because responsibility for the answer has to sit with a person.

Tasks depending on unwritten organisational context — how things actually work here, what the client meant, which rule is loose this quarter.

Anything touching systems you cannot afford to have wrong, until the identity, permission and logging work is done. See the security problem.

Small-volume tasks. The setup, integration and oversight cost is fixed. Below some volume it never pays back.

The questions to ask before committing

Can I write down the steps? If yes, do not use an agent. For broader independent background, see NIST AI Risk Management Framework.

What happens when it is wrong, and will we notice?

What is the volume, and what does that cost per month at model prices?

What must it reach, and does that access exist? Integration is the reported primary obstacle for 46% of teams.

Who is accountable for its output?

What does success look like numerically, and who agreed that?

And: what is the simplest thing that could work? Ask it last, and take the answer seriously.

What this does not mean

Not that agents are overhyped and useless. They do something genuinely new: handle situations nobody scripted.

The point is narrower. The set of tasks where that property is worth its cost is smaller than the set of tasks people are currently pointing agents at.

And the projects that work share a profile: narrow, high-volume, repetitive, with recoverable errors and clean reachable data. See where agent projects fail.

The short version

If you can write down the steps, write down the steps. Automation is cheaper, deterministic and auditable.

If a person reviews every output anyway, autonomy adds risk and nothing else.

Fix the data and the process first — you may find the agent is unnecessary.

Low volume rarely pays back, because setup and oversight costs are fixed.

And ask what the simplest thing that could work is. Ask it last, and mean it.