When an Agent Is the Wrong Solution

Every serious operator now has a line item, spoken or unspoken, for "doing something with AI agents." The pressure is real and it is rational: the technology is capable, the competition is moving, and no executive wants to explain later why they sat still. But that pressure produces a predictable failure — buying an agent for a problem that never needed one. The result is AI theatre: a system that demos well, costs more than the work it replaced, and quietly makes the operation harder to reason about.

The uncomfortable truth is that many of the tasks queued up for agents are better solved by something older and duller. A scheduled script. An RPA bot. A changed form. A rewritten policy. Sometimes a hire. These options rarely make it into the pitch, because they are not exciting and nobody sells them to you. That is exactly why a disciplined buyer should hold them up as the benchmark. An agent should have to beat the boring alternative on the merits, not win by default because it was the only option in the room.

This is not an argument against agents. It is an argument for spending on them where they actually pay — on genuine judgment over messy, variable inputs — and refusing to spend on them where a cheaper, more predictable tool would do the same job with less risk. Knowing the difference is a commercial skill, not a technical one, and it is learnable.

What an agent is actually for

An AI agent earns its cost in a specific place: when a task requires interpreting inputs that are unstructured, variable, and impossible to fully specify in advance, and then choosing among actions based on that interpretation. Reading a free-text customer message and deciding how to route it. Reconciling two documents that never quite match. Drafting a first response that a person will review. These are jobs where writing down every rule is impossible because the rules always end in "it depends."

The moment a task stops depending on judgment and starts following a stable set of rules, the case for an agent weakens fast. If the inputs are structured, the logic is fixed, and the same input always produces the same correct output, you are describing software that already exists and has for decades. Bringing a probabilistic model to that work adds cost, latency, and a new failure mode — the confident wrong answer — in exchange for flexibility you do not need.

An agent should have to beat the boring alternative on the merits, not win by default because it was the only option in the room.

Five times the boring answer wins

Before commissioning an agent, walk the problem past these five cheaper solutions. If one of them fits, take it — and put the savings toward the work that genuinely needs judgment.

1. A script or scheduled job. If the task is deterministic — same inputs, same rules, same output every time — code it once. It is faster, testable, auditable, and it never hallucinates. Moving data between two systems on a schedule is not an AI problem.
2. An RPA bot. If the work is high-volume clicking through stable screens with structured data, robotic process automation does it more cheaply and predictably. Reserve agents for the cases where the screen or the data is too variable for a bot to follow reliably.
3. A form or interface change. Many "the AI should figure out the missing information" requests are really "our intake form lets people submit incomplete information." Fix the form and the problem disappears rather than getting automated. This is almost always the highest-return option.
4. A policy or process fix. When the pain is inconsistent decisions or slow approvals, the root cause is often an ambiguous policy, not a missing tool. Clarify the rule and you remove the ambiguity for humans and machines alike. Automating a broken process only makes it fail faster.
5. A hire, or a redesigned role. Some work is low-volume, high-stakes, and relationship-driven. The honest answer is sometimes a capable person, or freeing an existing one by removing the deterministic drudgery around them — which the script above may already do.

How to tell the difference in the room

You do not need to be technical to run this triage. You need to ask a short sequence of questions and listen to whether the answers point at judgment or at rules.

Can the rules be written down completely?

Ask the person who does the work today to explain their decision. If they can articulate a complete set of rules — even a long, fiddly one — the task is deterministic and belongs in a script or a bot. If every branch of their explanation eventually lands on "it depends on the context, and the context is always different," you have found real judgment. That residual, and only that residual, is where an agent belongs.

Where do the inputs come from?

Structured inputs from your own systems rarely justify an agent; conventional software handles them better. Unstructured inputs from outside your control — customer language, supplier documents, inbound email — are where interpretation is unavoidable. If the mess is inside your systems, fix the systems. If the mess arrives from the outside world, interpretation may be worth paying for.

What does a wrong answer cost, and who catches it?

An agent introduces the possibility of a fluent, plausible, wrong output. If there is no human review, no clear escalation path, and a wrong answer is expensive or irreversible, the honest position is that the task is not ready for an agent yet — not because the technology cannot do it, but because you have not designed the safety net that makes it responsible to try. Deterministic tools do not remove that need entirely, but they fail in more predictable, more catchable ways.

The decision in front of you

The next time an agent is proposed, do not ask "can an agent do this?" — the answer is almost always yes, and it is the wrong question. Ask "what is the simplest thing that solves this, and does the agent beat it?" Make the boring alternatives compete. Fix the form before you automate around it. Script the deterministic core before you wrap it in a model. Clarify the policy before you teach a machine to navigate the ambiguity.

Doing this well does not slow your AI programme; it concentrates it. Every problem you solve with a cheaper, more reliable tool is budget and attention freed for the genuine judgment tasks where agents create durable advantage — and where you will want them running cleanly, under supervision, on inputs you have already tidied up. The discipline of saying "not here" is what earns the credibility to say "yes, here" and mean it. The first decision is not which agent to build. It is whether this problem deserves one at all.