AI agents that do the work, not just answer questions
A chatbot answers. An agent decides and acts. The difference is tool access, and most of the engineering is in what you do not let it do.
Why most AI pilots never reach production
- It demos beautifully on clean questions and falls apart on the messy ones real customers ask.
- It has no access to your systems, so it can describe the refund policy but cannot issue a refund.
- Nobody defined what happens when it is unsure, so it guesses confidently and someone has to apologise.
- There is no log of why it did what it did, which makes it impossible to trust or improve.
A real one, building itself.
Not a diagram of an idea. This is the shape of a system we have actually shipped for this service, assembling and then running a job.
Six things, all of them yours.
Scoped agent
A narrow agent that does one job well beats a general one that does everything unreliably. We scope hard.
Tool access
Read and write against your real systems, with each tool call permissioned and logged.
Guardrails
Explicit limits on what it can do unsupervised. Refunds over a threshold, anything legal, anything irreversible: human first.
Confidence routing
Below a threshold it escalates rather than guesses, with the full context attached so the human is not starting cold.
Your voice
Tone tuned against your real past replies, not a generic assistant persona.
Full transcript logs
Every decision, tool call and input retained, so you can audit why it did something and improve it.
What we build this on.
Chosen per project, not per habit. If your team already runs something that works, we build on that instead.
Four steps, no surprises.
Numbers we actually see.
Typical ranges from our own builds, not industry averages. Yours will depend on your process.
AI Agents, answered.
What is the difference between an AI agent and a chatbot?
A chatbot generates text. An agent has tools it can call, so it can look up an order, issue a refund, book a slot or update a record, then decide what to do next based on what it found. The engineering effort is mostly in constraining that power, not in the conversation.
How do you stop it making things up?
Grounding and gating. Answers come from retrieval over your actual documents rather than the model's training, and anything below a confidence threshold escalates to a person instead of guessing. We also keep the agent narrow, because a scoped agent has far fewer opportunities to invent.
Which model do you use?
It depends on the job. We benchmark on your real data rather than picking a favourite, and often mix them, using a cheaper fast model for classification and a stronger one for drafting. Models are swappable in the way we build, so you are not locked to one vendor.
Will customers know they are talking to AI?
That is your call, and we will build it either way, but we recommend disclosure. In practice the agent handles the routine volume and hands the genuinely difficult conversations to a person, which is a better experience than a human buried in tickets.
What does it cost to run?
Token cost is usually the smallest line item, typically single-digit dollars per thousand interactions at current pricing for a well-scoped agent. The real cost is the build, and the tuning in the first month.
Find out what this is worth to you.
A 20-minute call where we map your process and tell you honestly what is worth automating first, and what is not worth touching.
What pairs with this.
Workflows
We design, build and host n8n and Make workflows that run your operations end to end, with error handling and alerting built in from the start.
Bespoke Dev
Custom-coded automation for processes no-code tools cannot handle: heavy data volumes, complex branching logic, legacy systems and strict compliance rules.
Integrations
We connect your CRM, inbox, forms, billing and spreadsheets so data moves between them automatically, including tools with no native integration.