We Built Our Own AI Assistant. Now We Build Yours.

Most businesses have now had the same conversation twice. Someone suggests adding AI. Someone else asks what it would actually do. The idea quietly dies, or worse, a generic chatbot gets bolted onto the website and starts confidently telling customers things that are not true.

We took a different route. Rather than sell an idea we had not tested, we built an AI assistant for our own business first, ran it in public on this website, and fixed everything that broke. That assistant is Steve. The framework underneath it is In4raGP1. Over the coming months we are turning that framework into a service: AI Assistant Implementation, delivered for businesses, schools and multi-academy trusts that want something more useful than a chatbot.

Meet Steve, the assistant running on this site

Steve is the chat window in the corner of this page. He is not a novelty. He does a specific set of jobs, and he does them because we needed them done.

• Answers questions about our services using a hybrid approach — fast deterministic answers for the things we know we will be asked, and a language model for everything else
• Speaks and listens. You can talk to Steve and Steve talks back, with a live transcript on screen so nothing depends on you catching every word
• Captures enquiries properly. When someone wants to be contacted, the details reach a real inbox with the context of the conversation attached, not a bare email address
• Prepares quotes. Steve can start a quote, which flows into a documented approval process with a PDF, a countersignature step and an archived copy
• Knows when to stop. If a conversation needs a person, Steve hands it to a person rather than guessing

That last point matters more than any of the clever parts. An assistant that recognises the edge of its own knowledge is worth considerably more than one that does not.

What In4raGP1 actually is

In4raGP1 is our internal name for the operating framework we developed while building Steve. It is not a product you install. It is the set of decisions that separate an assistant that works from one that embarrasses you.

Grounding before generation

The assistant answers from a defined, maintained description of your business. When a question falls outside that, it says so. We rewrote Steve's own instructions specifically to remove claims about capabilities he did not have — because an assistant overstating what it can do is a customer service problem, not a technical one.

A designed handover

Every assistant will eventually meet a question it should not answer. The interesting design work is what happens next: who gets notified, what context travels with the enquiry, and how quickly a human picks it up.

Voice as an option, not a gimmick

Two-way voice is genuinely useful for anyone working hands-free, anyone who finds typing slow, and anyone using a phone one-handed. It is also easy to do badly. Speech recognition mangles technical terms and product names constantly, so we built a repair layer that corrects the predictable mistakes before they reach the model.

Built on infrastructure you already trust

Steve runs on serverless infrastructure at the edge, with the assistant logic, the enquiry routing and the document pipeline as separate, individually deployable services. Nothing is a black box we cannot inspect.

The part most businesses have not thought about yet

There is a second reason to take this seriously, and it has nothing to do with chat windows.

Increasingly, the thing arriving at your website is not a person. It is an AI agent, working on a person's behalf, trying to establish what you do, whether you are suitable, and how to transact with you. If your site only makes sense to a human with a mouse, those agents form a poor impression of your business and move on.

We made In4ra readable to machines as well as people — a published capability description, a machine-readable summary of what we do, and structured endpoints an agent can query. An independent assessment scored the result 93 out of 100 for agent-readiness. We also published an AI Use Policy setting out honestly how the assistant works and how agent-driven commerce is handled.

This is early. That is exactly why it is worth doing now — the businesses that are legible to AI agents in 2026 will be the ones those agents recommend in 2027.

What the AI Assistant Implementation Service will cover

The service follows the order we did it ourselves, because that order is what stopped us wasting money.

1. 1Scope before software. We establish what the assistant is for — the questions it should answer, the enquiries it should capture, the things it must never attempt. Where personal data or safeguarding is involved, this is where those requirements get written down, not discovered later.
2. 2Ground it in your business. We build and structure the knowledge the assistant answers from, so its responses reflect your services, your prices and your language rather than a generic industry summary.
3. 3Build the handover. We connect the assistant to the places your team already works, so an enquiry becomes an action rather than a notification nobody opens.
4. 4Deploy, measure, correct. We put it live, review real conversations, and fix what the transcripts show is wrong. This step never fully ends, and any supplier who tells you otherwise has not run one of these in production.

Optional, depending on what you need: two-way voice, quote and document workflows, and the agent-readiness work described above.

What it is not

We would rather set expectations now than manage disappointment later.

• It is not a replacement for your team. It handles the repetitive questions so your people handle the ones that need judgement.
• It is not a one-week job that then runs itself. The assistant needs reviewing as your services change.
• It is not appropriate everywhere. For some organisations, particularly where children's data or regulated advice is involved, the honest answer is a narrower assistant or none at all. We will tell you that.

Who this suits

Businesses fielding the same twenty questions every week. Organisations where enquiries arrive out of hours and go cold before anyone sees them. Schools and trusts wanting a front door for routine questions from staff and parents, with clear boundaries about what it will and will not discuss. Any team where a quote takes three days because it needs three people to look at it.

Interested early?

The service is in development and we are talking to a small number of organisations first, so the early builds are shaped by real requirements rather than our assumptions.

If you want to see the framework working before you decide anything, it is already on this page — open the chat window and ask Steve something difficult. Then get in touch and tell us what you would want your own assistant to handle.

No obligation, and no AI strategy deck. Just a conversation about whether this would actually save you time.