Give agents useful work
AI agent examples for service firms: the work, the information and the human decisions
An AI agent uses a model to choose steps or tools while working towards a goal. Useful examples explain what it can do, which information it needs and when a person takes over. Here are three service-business jobs where that distinction matters, including the limits of what we have actually delivered or observed.
Judge an agent by its permitted work, evidence and stopping rules; a convincing answer on a screen is only one part of the job.
What makes an example an agent?
A system can use AI without letting the model choose its route. An email workflow might always classify a message, match it to a client and create a draft. An agent might decide which records to inspect next or which permitted tool to use because the first result is incomplete.
Anthropic’s explanation of agents and workflows makes this distinction useful: predictable routes and model-directed routes create different design and testing needs. A chat interface does not settle which kind of system sits behind it.
For each example, ask what changes between jobs. If the next step always follows an agreed rule, ordinary automation may be sufficient. If the system needs to choose a route, define the choices and the point at which it must stop.
Example 1: match a journalist request to a client
We delivered a PR workflow that sorts journalist requests, matches them to relevant clients and helps prepare a response. At launch, a person approved every outgoing message. The existing journalist workflow case study records the delivered work.
That is an AI-assisted workflow; the case does not establish that a model independently chose an open-ended sequence of tools. An agent version could be allowed to investigate whether a client has relevant evidence before recommending a match. That added behaviour would need its own evaluation.
| Requirement | Reason |
|---|---|
| Current client expertise and permitted claims | A relevant-looking name is insufficient evidence for a pitch. |
| The request, deadline and journalist’s requirements | A correct match can still be an unusable response. |
| Human approval of the outgoing message | The client representation needs an accountable reviewer. |
| A stop for missing evidence | The agent should not invent a client capability to complete the task. |
Example 2: investigate a change in a client metric
Said has described a capability in our current build where an agent can investigate why a metric has fallen. Treat this as a reported capability: we do not have a published observed investigation with an independently checked diagnosis.
A useful investigation would identify the metric and period, inspect authorised sources and separate observations from possible explanations. The agent might choose a second source after finding a change in the first. A person decides whether the proposed explanation is supported and what to do next.
This is separate from our delivered unified analytics work, which brought information from 12 platforms together. Connecting the numbers and independently establishing the reason for a change are different pieces of work.
Example 3: help a client prepare an informed content request
Said sees an opportunity for content businesses to offer former clients a smaller ongoing service using context already held about their business and competitors. The proposed self-service system would let the client do recurring work with that context.
An agent could ask for a missing objective, retrieve the relevant approved context and prepare a brief for review. The client and the service firm still need to agree who checks claims and who approves publication. This is a proposed pattern; no customer uptake or revenue result is claimed.
A task that asks only for an outline from approved inputs may need a fixed workflow. Give an agent extra freedom only when that freedom helps it handle a real variation, and test the resulting choices.
Give the agent a job it can finish or decline
Worked example
A bounded metric investigation
Invented teaching example. No client outcome is claimed.
GOAL
Explain which observed changes might account for a lower metric in an agreed period.
PERMITTED WORK
Inspect the client’s authorised reporting data and definition record.
Choose a second source when the first observation needs checking.
LIMITS
No changes to a campaign or budget. Stop if the metric definition or source period cannot be established.
HUMAN DECISION
A reporting expert reviews the evidence and distinguishes an observation from a causal claim.
TEST
Use an invented metric change with known underlying records, then repeat with an unavailable source.
Measure source correctness and appropriate stopping, before testing a live client job.
Further reading
Written by Said Lopez, founder of LeverageAI and author of Lean AI. Our guides draw on the method in the book and our work with clients.