AI agents
Agents that do a defined job and know when to stop. The interesting part isn’t what they can answer — it’s how cleanly they hand over when they can’t.
A lot of AI agent work is a demo that never survives contact with real users. It answers the easy questions impressively and the hard ones confidently and wrongly, which is worse than not answering at all — particularly for a business where a wrong answer about price, availability or eligibility costs you a customer.
We’re interested in the version that earns its place: an agent scoped to a specific, repetitive job, with a clear boundary and a clean handover to a person the moment it’s outside that boundary. That’s less impressive in a demo and considerably more useful on a Tuesday.
Fit
When this is the right job — and when it isn’t
We’d rather lose a project at this stage than sell you one you don’t need. The right-hand column is the honest half.
- You answer the same handful of questions dozens of times a week
- Enquiries arrive out of hours and go cold before anyone can reply
- Someone spends their morning triaging messages that could be sorted before they reach them
- You have a body of documentation that people would use if it were easier to search
- Your enquiry volume is low. A person replying quickly beats an agent, and always will
- The questions people ask are genuinely varied and high-stakes. Scope it down or don’t build it
- You want it to sound human and never admit it’s a machine. We won’t build that
Scope
What’s included
A defined boundary
What it handles, what it refuses, and what it escalates — agreed before anything is built, because this is where these projects fail.
Clean handover
When it’s out of scope it says so and passes the conversation to a person with the context attached, rather than guessing.
Grounded answers
Responses drawn from your actual content and data, so it isn’t inventing prices, availability or policies.
Something you can watch
Logs of what it was asked and what it said, so you can see where it’s helping and where it needs narrowing.
Questions
Straight answers
Will it make things up?
Any language model can. The mitigations are scope and grounding: keep the job narrow, answer from your real content rather than the model’s general knowledge, and make refusal-plus-handover the default when it isn’t confident. We build all three in and we’re honest that it reduces the risk rather than eliminating it.
Should it pretend to be a person?
No. Beyond the ethics, it backfires — people work it out, and discovering they were misled costs you more trust than the agent ever saved you.
Do I need this, or do I need automation?
If the task is well-defined and the input is structured, you want automation and it’ll be simpler and more reliable. Agents earn their place when the input is messy human language.