AI development
We build AI into the systems you already run.
Answers from your own documents and reports, paperwork read for you, requests sorted and sent on, agents that do the follow-up — added to the software you already have, with the permissions and review a business needs.
Your team and your apps
Chat, copilots inside your screens, agents on a schedule
Orchestration
Agents, tool calls, guardrails, human sign-off
Connectors
MCP servers · a RAG index · your APIs
Your data
Database · documents · the tools you already use
What you can get
Start small, with one job AI can take off your hands.
Each of these can be picked on its own or added to a system. Choose one and the draft opens with it already selected.
Starting out
AI readiness review
Not sure where AI fits? We look at your work and your data, and write down where AI would help, what it would need, and where to start.
Assessment · data review · a written plan
AI in your existing system
Add AI to the software you already run — a new answer, a new button, a new screen — without replacing it.
Model integration · your APIs · Claude · OpenAI · open models
Getting answers
Assistant in your system
A chat inside your own screens that answers from your records, in your vocabulary — not a general chatbot in another tab.
In-app chat · tool use · your data
Knowledge base over your documents
Ask your manuals, policies and contracts a question, and get the answer with the passage it came from.
RAG · embeddings · citations
Ask your reports
Ask about your numbers in plain words — sales by region, invoices still unpaid — and get the answer with the report behind it.
Natural-language queries · reporting · permissions
MCP server
Let Claude, ChatGPT or your own assistant work with your database and tools, with the same permissions as the person asking.
Model Context Protocol · tools · resources
Taking work off your hands
Invoice and document reading
Invoices, forms and contracts become records instead of being retyped. Anything uncertain waits for a person.
Extraction · validation · review queue
Sort and route incoming requests
Emails, tickets and requests sorted by what they are about, and sent to whoever handles them.
Classification · tagging · routing
An agent for one job
An agent that does one job on your systems — reviewing, following up, preparing a report — and hands the decision to a person.
Tool use · multi-step · human sign-off
WhatsApp or Telegram agent
Customers or staff write a request in WhatsApp or Telegram, and an agent carries it out in your systems.
Messaging APIs · webhooks · agents
In a business like yours
The same pieces, doing different work.
Examples of what AI can do in the kinds of systems we build. Yours will differ — describe it in the draft.
Operations and inventory
Staff ask what is in stock and where, instead of opening several screens. Supplier invoices become records, and a person checks only the ones that do not match.
Schools and academies
Staff find the policy or procedure they need by asking for it. Enrolment forms are read into records, and families’ questions reach the right office.
Bookings and sales
Customers ask about availability on WhatsApp and get an answer from your calendar. Each enquiry reaches the right person with what it needs attached.
E-commerce
Support emails are sorted into order problems, returns and questions before anyone reads them. Your team asks which orders still have not shipped.
HR and payroll
Staff ask the handbook about leave and get the paragraph that answers it. Absence requests are read and sent to the manager who approves them.
Property management
Maintenance requests arriving by message are sorted by urgency and assigned. Contracts can be searched by what they say, not by file name.
How a project runs
Small first, tested, then for everyone.
An AI project earns its place in steps, so you can stop at any of them.
- 01
Review
We look at the work you would hand to AI and the data it needs, and agree what a good answer looks like.
- 02
Pilot
One job, for a few people, on your real data — small enough to stop if it does not earn its place.
- 03
Evaluation
The answers are tested against cases you already know the answer to. It goes further only if they hold.
- 04
Rollout
Permissions, logs and monitoring in place, then everyone who needs it — and tested again whenever the model changes.
For the technical reader
How it works underneath.
Four of the offerings above, shown working: the protocol, the retrieval, the agents and the extraction.
MCP servers
Your assistant, reaching your systems — with your rules.
An MCP server is the bridge between an AI assistant and the software you run. It decides what the assistant can see and do, and it does it on behalf of the person asking.
Someone on your team asks
“Which orders from last week still have not shipped?”
Your MCP server
Checks what that person may see, and calls the one tool that answers it.
Your systems
Database · documents · the tools you already use
The answer comes back — With the records it came from, so it can be checked.
RAG
Answers from your own documents, with the source.
Retrieval-augmented generation: your documents are indexed once, the relevant passages are found for each question, and the answer is written from them — citing where each part came from.
Your documents
Manuals, policies, contracts, records
Indexed
Split into passages and made searchable by meaning
Retrieved
The passages that answer this question, and no others
Answered
Written from those passages, with citations
Example
How many days does a customer have to return an item?
Thirty days from delivery, if the item is unused. Custom orders cannot be returned.
Agents
Agents with one job each, and a person who signs off.
A general chatbot does everything a little. An agent built for one job — with its own tools and the rules written for your project — does that job properly, and hands the decision back to a person.
- A request arrives
- Requirements agent
- The work is done
- Review agent
- A person signs off
Requirements
Turns a request into what is actually true in your system, the approach, and the questions nobody has answered yet.
Review
Checks every change against the rules written down for your project, before a person does.
Audit
Flags when the plan and the work have drifted apart, so a stale note is never read as current.
Document automation
Paperwork in, records out.
Documents are read into the fields your system needs. What the model is sure of goes straight in; what it is not sure of waits for a person, with the reason shown.
Example
Record created
- Supplier
- Northwind Supplies
- Invoice no.
- INV-20417
- Date
- 12 Sep 2026
- Total
- $1,240.00
Needs a person — The tax ID does not match the supplier on file.
Built to be trusted
Where your data goes, and who stays in charge.
An AI system sends part of your data to a model to get an answer. This is what happens to it, and what keeps a person in control.
Your data
Sent only to a provider you approve
Hosted models such as Claude or OpenAI’s process your data on their servers. You choose the provider, and nothing is sent before you authorise it.
Not used to train their models
Their business APIs do not train on your data by default. You see each provider’s terms before you decide.
Kept briefly, under their terms
A provider may keep requests for a limited time to check for misuse, as its terms set out.
Or kept on your own servers
Where data must not leave your infrastructure, an open model can run there instead.
Control
Same permissions as a person
The assistant can only reach what the person asking could reach themselves.
Every call logged
What was asked, what was looked up, and what was answered — kept and reviewable.
A person approves what matters
Anything that changes a record or leaves the building waits for a person.
Answers are tested
The answers themselves are checked against known cases before a release, like any other code.
After delivery
What keeps costing once it is running.
An AI system has running costs an ordinary system does not, and they grow with use. They are estimated with you before anything is built, in the conversation that follows the draft.
Model usage
A hosted model is paid for by use: every question asked and every document read has a cost.
Hosting the pieces
A search index, an MCP server or an open model needs somewhere to run.
Messaging platforms
WhatsApp’s business platform has its own charges and set-up. Telegram’s bot platform does not charge.
Keeping answers right
Models change and so do your documents. Answers are tested again, and documents indexed again, when they do.
Tell us what you would hand to AI first.
The draft takes a few minutes. Describe the work — the questions your team keeps answering, the documents you keep retyping — and you get a written scope back.
Start a draft