AI agents in production — three use cases that work

The conversation about AI agents is loud, vague, and often disconnected from what an SME actually needs. We’ve been quietly shipping agentic workflows into client systems for the last year — sometimes as standalone products, more often embedded inside existing tools — and there are three patterns we keep coming back to. Each one is dull on a marketing reel and genuinely transformative inside the business that uses it.

1. Document-in, structured-data-out

The most common request: someone receives a steady stream of unstructured documents — receipts, invoices, CVs, supplier quotes, insurance schedules — and someone else has to extract a handful of fields and type them into a system.

Modern multimodal models do this very well. Drop a file in, return JSON in the shape your downstream tools want, and surface a confidence score so a human can spot-check the cases the model isn’t sure about. This is exactly the pattern behind the AI expense capture in the new LunaHR — the receipt comes in, the model parses it, and the reviewer’s job becomes “confirm” rather than “type”. The time savings on a high-volume team are large; the user-facing UI changes are small.

2. The triage agent

You have an inbox — support email, contact form, sales enquiries, IT helpdesk tickets — and the first thirty seconds of each ticket is spent figuring out what kind of thing it is and where it should go.

An LLM is very good at this. Wire one up to read each incoming message, classify it against a known taxonomy, route it to the right queue, and propose a draft response when the answer lives in your knowledge base. You don’t take the human out — you give them a head start. We’ve built variants of this for IT support helpdesks where the agent picks the right runbook and pre-fills the diagnostic steps before a technician sees the ticket.

3. The data-questioning agent

Almost every business has a database that holds the answer to questions the people who need them can’t write SQL to ask. “What are our top ten clients by revenue this quarter?”, “Which support tickets did we breach SLA on last month?”, “Show me a list of contracts expiring in the next 60 days.”

An agent with read-only access to a well-modelled view layer, given a clear schema description and a small library of approved queries, can answer these reliably. The trick — and it’s a real one — is in the guardrails: scoped data access, query allow-lists, and a UI that always shows the underlying SQL or filter so the user can verify what they’re looking at.

What ties them together

None of these are agents in the “set it loose and watch it work” sense. They’re narrowly scoped, have a clear contract with a downstream system, and are designed around the assumption that they will sometimes be wrong. That’s the boring secret. The agents that make it into production are the ones that earn trust on day one and quietly grow scope over months.

If there’s a process in your business that looks like one of these three patterns and feels like a candidate, we’d love to hear about it. We can usually scope a working prototype in a couple of weeks.