Strip away the hype and an AI agent is a piece of software that can take a task with some judgement in it and carry it out from start to finish, without a person driving each step. For a small business that rarely means a robot running the company. It means one well-defined job, done reliably, that used to need a person’s attention all day. Here’s what that looks like in practice, and where it falls flat.
The jobs they’re good at right now
The agents we build and see working do one of a few things:
- Qualifying and chasing leads. An agent reads an incoming enquiry, works out whether it’s a fit, asks the obvious follow-up questions, and only hands a warm, qualified lead to a person. The sales team stops spending its day on tyre-kickers.
- Triaging inbound messages. Email and tickets get read, sorted, and routed, with the routine ones answered straight away and the rest passed to the right person with context attached.
- Turning messy input into clean data. Pulling the figures off an invoice, summarising a long thread, extracting the key fields from a document. Reading unstructured text is something models do well.
- Drafting from a known pattern. First-draft replies, reports, or summaries that a person then checks and sends, which removes the blank-page time without removing the human sign-off.
All of these jobs share the same shape: specific, repetitive, with clear inputs and a clear definition of done. That’s where an agent earns its keep.
Where they fall flat
The limits matter just as much. The marketing won’t mention them:
- Anything needing real judgement or a relationship. Agents are weak on calls that depend on context they don’t have, and weaker on the ones that depend on trust.
- Tasks with no clear right answer. If you can’t define what good looks like, neither can the agent, and it will produce confident output that’s wrong in ways you won’t immediately catch.
- Work built on messy data. An agent pointed at contradictory sources gives you contradictions, faster. It’s the same reason most “we need AI” requests turn out to be process problems first.
What it takes to make one work
The agent itself is rarely the hard part. The work is in the wiring: connecting it to your systems so it can read and act, defining the task tightly enough that it behaves, and putting a human checkpoint wherever a mistake would be expensive. An agent that can’t reach your data, or that’s pointed at a vague task, looks impressive in a demo and does nothing useful on Monday.
That’s why the AI agents we build start with the job and the data, not the model. Once those are right, the agent is close to the easy bit.
A realistic first step
If you’re a small business curious about this, don’t start with “let’s add AI.” Start with the single most repetitive, rules-based job that eats someone’s day and has a clear definition of done. That one job is where an agent will pay off first, and getting it working well teaches you more than a grand plan ever will.
If you’ve got a job like that in mind, tell us what it is and we’ll tell you honestly whether an agent is the right tool for it yet.
Written by
Alex
CEO
I’m a software developer and CEO of Digital Royalty, helping growing teams scale their SaaS platforms without losing quality, visibility, or control. I focus on building structured, maintainable systems with clear processes, reporting, and accountability. With over a decade of experience across agency and in-house roles, I specialise in delivering long-term, scalable solutions that support complex, evolving products.
