Voice agents
Answers the business line in a natural voice, handles the questions it has been taught, books into the same calendar your team uses, and puts real callers through to a person. Every call transcribed into the same place as the messages.
AI Agents
Software that does the job, not just the talking.
A chatbot answers. An agent acts — it checks the calendar, writes the booking, updates the record, sends the follow-up. We build agents around the systems you already run, with clear limits on what they may do alone and a person able to step in at any point.
- Scoped to one job
- Human takeover
- Works your systems
One job, done properly
Every agent ships with the list of what it may do on its own, and what it must hand to a person.
The idea
An agent is a colleague with a very short job description
The agents that work are the narrow ones. Give a piece of software one job, the systems it needs to do that job, and an explicit rule for when to stop and fetch a human — and it will do it at 2am on a bank holiday. Give it everything and it will do all of it badly.
- One job per agent, scoped in writing before anything is built
- Read and write access only to the systems that job genuinely needs
- An escalation rule, so it stops rather than guesses
- A log of what it did and why, that a person can actually read
What we build
The agents businesses ask us for
These are the jobs that come up again and again in payments and local business. Each is built to fit your systems — none of them is a product you are switched onto.
Review agents
Asks for the review at the right moment — after a completed appointment or a paid order — and drafts a reply to every review that arrives, in your voice, for approval or automatic sending.
Follow-up agents
Goes back to the quotes nobody chased, the no-shows and the customers who have gone quiet, on the channel they first used. Books them back in when they answer, and stops the moment they ask it to.
Social agents
Drafts a week of posts from what the business actually has on — new products, offers, opening hours — schedules them for approval, and answers the comments and DMs those posts bring in.
Insights agents
Watches what your terminal and online checkout are telling you — busiest hours, what sells together, what has gone quiet — and sends a short plain-English brief each week.
Back-office agents
The internal jobs: triaging inbound email, chasing missing paperwork, reconciling a daily report, preparing the same summary somebody currently builds by hand every Monday.
How the work runs
Four stages, in order
Each one ends with something you can look at and a decision you can make.
- 01
Scope the job
One job, written down: what triggers it, what the agent may do alone, what it must escalate, and how we will know it worked.
- 02
Wire it up
Access to the systems that job needs — calendar, inbox, CRM, POS — and nothing else. Credentials encrypted, permissions kept narrow.
- 03
Rehearse it
The agent runs against real scenarios while a person reviews everything it proposes, before it is allowed to act on its own.
- 04
Release and widen
It goes live on the narrow version of the job. What it may do alone widens only once the log shows it earning that.
Built with
Tools chosen because they can be handed over
Nothing exotic, and nothing that locks the work to one vendor. An agent you cannot move, audit or turn off is a liability whatever it does.
Models
- Claude
- GPT
- Gemini
- Open-weight models
Orchestration
- Tool calling
- Retrieval
- Queues & schedulers
- Evals
Runtime
- Node.js
- Python
- PostgreSQL
- Webhooks
Around it
- Audit logging
- Human-in-the-loop review
- Rate limits
- Cost controls
Questions
AI agents, answered
A chatbot produces a reply. An agent produces an outcome — it uses tools, so it can check availability, write a booking, update a record or send a message. A chatbot tells the customer what your opening hours are; an agent puts them in the diary.
Scope. Each agent gets access to the specific systems its job needs and no others, and an explicit list of actions it may take alone. Anything outside that list is escalated to a person rather than attempted.
Always. Every agent we build assumes a person will need to step in — mid-conversation, mid-task — and hands back cleanly afterwards. An agent nobody can interrupt is not one we would ship.
That is the usual case. Most agents are built around an existing calendar, CRM, inbox or POS rather than replacing them. Where a system has no API we will say so before the work starts, not after.
It is logged, and the log is written to be read by a person rather than a machine. Corrections go into the agent's instructions and its test scenarios, so the same mistake has to get past a test next time.
Either. Some clients want it hosted and monitored by us with an agreed response time; others want it handed over with the code, the documentation and a walkthrough. Both are normal.