AI agent use cases, department by department

The question worth asking is not which department could use an AI agent. It is which processes in that department are written down, repetitive, and involve interpreting language. Those three conditions predict success far better than the department does.

The test that applies everywhere

1. Is the process written down as steps and rules? If not, write it down first. That step alone often removes the need for the project.
2. Which steps are rules and which need judgement? Rules become plain code, which is cheaper and correct every time rather than most of the time.
3. Does a judgement step read unstructured language? That is where a model earns its place. Numeric work does not.
4. Does any step spend money or contact a customer? Then it needs a person approving before it commits.

Customer support

The strongest case, because tickets are unstructured language and the process is usually already documented in a help centre.

Agents work well for triage and routing, for retrieving the relevant internal documentation, and for drafting a reply that a person sends. What they should not do unsupervised is issue refunds or make commitments, both of which are exactly the cases where a confident wrong answer costs real money.

Sales

Research and hygiene rather than selling. An agent that reads a prospect's site, recent news, and your own CRM history and produces a briefing before a call removes an hour of preparation without touching the conversation itself.

The other reliable win is CRM hygiene: turning call notes into structured fields, flagging stale opportunities, and spotting duplicate records. Unglamorous, and it is where the time actually goes.

Finance and operations

Be careful here, because finance is mostly arithmetic and arithmetic should be code. The agent's job is the exception handling around the arithmetic, not the arithmetic.

• Matching an invoice to a purchase order where the descriptions do not align, which is a language problem rather than a numeric one.
• Reading a supplier email and extracting the structured fields a system needs.
• Flagging anomalies for a human, with the reasoning shown, rather than deciding what to do about them.
• Never: approving a payment, or reconciling a ledger by asking a model to add things up.

Internal knowledge

The most broadly useful and the least discussed. Most companies have their answers scattered across a wiki, a chat history, and several people's heads, and new staff spend weeks finding out who to ask.

An agent with retrieval over your internal documents answers those questions directly. The requirement is that it cites its source, so the reader can check, and that it says it does not know rather than inventing something plausible.

Marketing

Classifying inbound enquiries, clustering large search terms reports into themes a person can act on, summarising call notes into recurring objections, and drafting variations for testing. All interpretation, all with a human editing before anything publishes or serves.

What consistently does not work

• Agents given a department rather than a process. Scope is what makes them shippable.
• Models doing arithmetic, thresholds, or pacing. Use code and keep the model out of it.
• Unsupervised authority over budgets or customer communications. The failure mode is fast and expensive, and the upside does not justify it.
• Any agent shipped without an evaluation set. Without known-correct cases, every prompt change is an opinion and nobody can tell improvement from regression.

Building these against a company's own systems is AI agent development. The marketing-specific subset, where the workflows are already documented, is AI marketing automation.