Most AI agent pilots do not fail because the model is weak. They fail because nobody defined the job precisely, the data the agent needed was not ready, and success was never written down as a number. This note is a practical view of where agents create value inside ordinary businesses and how to get there without a costly detour.

Start with a workflow, not a chatbot

An agent is only useful when it sits inside a workflow that already has a beginning, an end and an owner. "Answer customer questions" is not a workflow. "Resolve order-status tickets by looking up the order, checking the courier feed and replying in the customer's language, escalating if the parcel is more than two days late" is. The second version tells you which systems the agent must reach, what a correct answer looks like, and when a person has to take over.

The best first candidates share three traits: high volume, a known pattern in most cases, and a clear cost to the business when done slowly or wrongly. Order status, invoice matching, first-line IT support, quote preparation and internal policy questions are typical examples.

Assisted before autonomous

There is a spectrum between a person doing everything and a system doing everything. Start in the middle. An assisted workflow drafts the reply, fills the form or proposes the classification, and a person approves it. This produces two things at once: immediate time savings, and a labelled dataset of what people accepted and corrected. That dataset is what earns the right to remove the human step for the easy cases later.

Data readiness is most of the work

Agents read. If the documents are out of date, duplicated across three wikis or locked in scanned PDFs, the agent will confidently repeat the wrong version. Before building, we inventory the sources the workflow depends on, decide which are authoritative, and fix access so the system only sees what the requesting user is allowed to see. Retrieval with permissions and freshness is a plumbing problem, and the plumbing determines the answer quality.

Define "good" as a number

Write a test set before writing a prompt. Fifty to two hundred real cases with the expected outcome, scored automatically on every change. Track accuracy, the rate of escalations to people, latency and cost per case. When the numbers are visible, arguments about whether the agent is "good enough" become short.

A sequence that works

  • Discovery: map the workflow, the systems and the failure modes. Agree the metrics.
  • Prototype on real data in weeks, evaluated against the test set.
  • Production: permissions, observability, cost limits, fallbacks and integration where people already work.
  • Operate: monitor, collect corrections, retrain or re-prompt, and expand scope case by case.

Agents that follow this path are unglamorous and reliable. They save hours every week, they are trusted by the people who work alongside them, and they get better with use. That is what "useful" means.