AI agents
An agent is not a chatbot with tools bolted on. It is a job description with a computer attached: it owns an outcome, it runs without being asked, and it knows when to stop and hand over.
What it is
We deploy autonomous, conversational agents that hold a real job in your business: answering customers at 3am, working a lead list, or turning a folder of documents into an answer. They run on your channels, in your language, against your own systems.
Three are live today. The Lead Generation System works outbound end to end. The Support Specialist resolves enquiries, refunds and escalations. The Data Analyst answers questions over your data instead of your inbox.
How we build it
Write the job before the prompt
We start with the outcome and the boundary: what the agent may decide alone, what it must escalate, and what it must never touch. Most failed agent projects skipped this and ended up with a demo nobody was willing to trust.
Instrument it before you improve it
Every run is logged: tool calls, retrievals, refusals, handovers. You cannot tune what you cannot see, and an agent without traces is a rumour with a nice interface.
Ship it narrow, widen it on evidence
Version one handles one slice of the job with a human reading the output. We widen the boundary only where the logs show the agent has earned it.
The nine types of agent
“AI agent” is sold as a single thing. It is nine things, and they are not close relatives. Below is the taxonomy we work from, and against each one, whether it is a line we take.
- We build this
Fixed automationThe digital assembly line
Rules, no learning, no memory. Something triggers it, it does the one thing, it stops. Reliable exactly until it meets a case nobody wrote down.
Enterprise Workflow AutomationWatch it work - We build this
LLM-enhancedSmarter, but not Einstein
An LLM reads the ambiguous input and works out what was actually meant, then a fixed boundary decides what it is allowed to do about it. Stateless: every task starts from zero.
Multilingual AI AgentsWatch it work - We build this
ReActReasoning meets action
Thinks, acts, looks at what came back, thinks again. Loops until the outcome is reached or it runs out of moves. This is the point where an agent stops being a function call.
Lead Generation SystemWatch it work - We build this
ReAct + RAGGrounded intelligence
The same loop, with retrieval bolted on: it goes and reads your documents instead of answering from memory. The difference between a confident guess and a cited one.
Air-Gapped RAG VaultsTry it live - We build this
Tool-enhancedThe multi-taskers
Reaches into APIs, databases and software, and switches between them mid-task. Where the agent stops describing the work and starts doing it in your systems.
Lead Generation SystemWatch it work - We are building this
Self-reflectingThe philosophers
Reviews its own reasoning after the fact, can explain why it chose what it chose, and changes what it does next time. Accountability is the feature, not the explanation.
Nova, our own concierge agentTalk to it - We are building this
Memory-enhancedThe personalized powerhouses
Carries what it learned about a person, a project or a preference across sessions, so the second conversation is not the first conversation again.
Nova, our own concierge agentTalk to it - Not our line
Environment controllersThe world shapers
Acts on the physical world rather than on a screen: robotics, IoT, infrastructure, anything where a wrong move has a cost you can touch.
- Not our line
Self-learningThe evolutionaries
Trains and retrains its own models rather than being prompted by one. Research-grade work, and it needs a data position most businesses do not have yet.
These are types, not a strict ladder. The order runs roughly from rigid to self-directed, but self-learning is not simply “above” environment control. They are different bets, and a business rarely needs both. If your problem sits in rows 8 or 9, we will say so rather than quote you for it.
Taxonomy from Mastering AI Agents (Galileo, 2025). Descriptions and the “our line” column are ours.
Who it is for
A good fit if
- Teams answering the same few questions hundreds of times a month.
- Outbound motions where research, qualification and first contact eat the day.
- Support queues that are growing faster than headcount.
Not a fit if
- Decisions where a wrong answer cannot be undone and no human reviews it.
- Processes nobody can describe in writing yet. Automating a mystery produces a faster mystery.
See it
A reference build
The Support Specialist
The support agent on our demos page is the same shape we deploy for clients. It reads the enquiry, decides whether it is allowed to resolve it, acts, and when it cannot, it hands over with the full context rather than dumping a blank ticket on a human.
- Resolves, refunds and escalates without a human in the loop
- Hands over with the transcript and the reasoning attached
- Works in the languages your customers actually write in
This is a system we run, not a client case study. We do not publish client numbers we cannot show you the workings for.
Talk to someone who builds this
No qualification call, no discovery deck. Describe the work and we will tell you whether this is the right line for it.