What is an AI agent, and does my business need one?

The word is used loosely enough to be almost meaningless in sales material, so here is the distinction that actually matters.

Workflow versus agent

A workflow runs a fixed sequence. Invoice arrives, extract the fields, validate them, write to the database, alert a human if confidence is low. The steps are the same every time. AI might handle one step, the extraction, but the path is decided by you in advance.

An agent is given a goal and works out the steps itself. It has access to a set of tools and chooses which to use, in what order, and when it is done. The path is different each run.

Why the distinction matters commercially

Fixed workflows are predictable, cheap to run and easy to debug. When one goes wrong you can see exactly which step failed. You can test them. You can hand them to an operations manager and explain what they do in two sentences.

Agents are flexible and much harder to reason about. The same input can take different routes on different days. Costs vary per run. Testing is harder because there is no single expected path. Debugging means reading a decision trace rather than a log.

That is not an argument against agents. It is an argument for using them only where the flexibility is actually needed.

When an agent earns its complexity

• The path genuinely varies. Research tasks where the next question depends on the last answer.
• The inputs are open-ended. A support query that might need a database lookup, a document search, or neither.
• The rules would be unmanageably long. If writing the decision tree out would take a hundred branches, letting a model navigate it may be more maintainable.

When it does not

Most SME processes. Inbox triage, lead routing, document extraction, report assembly, onboarding collection and follow-up sequences all have a knowable shape. Building them as agents adds cost and unpredictability for no benefit.

A good test: can you draw the process as a flowchart? If yes, build the flowchart. AI can sit inside individual boxes, classifying, extracting or drafting, without deciding the route.

What we usually recommend

Deterministic workflow, AI at the specific steps where judgement is needed, human review where a mistake would be expensive. It is less exciting than “we deployed an agent”, and it is the version that is still running reliably a year later.

If your process really does need to make decisions rather than follow them, that is a real design conversation and worth having properly. Most of the time the honest answer is that the simpler build does the job.