When not to use an AI agent

The default should be the boring option

An agent is a loop that decides what to do next. That flexibility is only worth its cost when the next step genuinely cannot be known in advance. If the sequence of steps is fixed, a pipeline is cheaper to run, easier to test and far easier to explain to an auditor.

We start every engagement by writing the workflow out as a flat list of steps. If nothing in that list branches on unstructured judgement, we build a pipeline and stop there.

Signals that an agent is the wrong tool

These patterns almost always indicate that a deterministic implementation will outperform an agent on cost, latency and reliability.

• The inputs are already structured and validated
• Every decision can be written as a rule a domain expert would sign off on
• The action is irreversible and cannot wait for a human approval step
• Correctness has a single right answer that can be computed directly
• The volume is high and the margin per transaction is thin

Signals that an agent earns its keep

Agents pay for themselves when the work is dominated by reading, reconciling and drafting across sources that never quite line up.

• Inputs arrive as documents, emails, images or scans in inconsistent formats
• The step order depends on what earlier steps discovered
• A person currently does the work by consulting several systems in sequence
• Drafting is expensive but reviewing a draft is cheap
• Tool use, including search, retrieval, calculation and geoprocessing, is central rather than incidental

The hybrid shape we ship most often

In practice the reliable design is narrow: a deterministic pipeline handles ingestion, validation and delivery, and an agent is invoked only for the judgement-heavy middle. Every agent output carries citations, and any action with a real-world consequence is gated behind an approval step a named person owns.

This keeps the failure surface small. When something goes wrong, you can tell immediately whether the pipeline mis-handled the input or the model produced a bad draft, because they are separate components with separate logs.