A business AI assistant wired into the tools you already use

It is an assistant that answers from your own documents, not from general knowledge picked up on the internet. You ask it a question in plain language — "what is the penalty clause in the Dupont contract?", "what lead time did we quote on quotation 2024-118?" — and it finds the passage concerned, quotes it, and says which document it came from.

Technically it is retrieval-augmented search: the relevant content is found in your files, then handed to the model that writes the answer. The practical consequence matters: the assistant only knows what you have given it, and it can always say where its answer came from.

A restricted set of documents, defined with you and never "the whole server". Contracts, quotations, meeting notes, internal procedures, product sheets, technical documentation: we pick the folders that are useful, set aside what is out of date, and agree a rule for keeping it current. An assistant fed on stale documents gives wrong answers confidently; that is the main risk, and it is handled through the scope.

The existing access rights are mirrored: somebody who cannot open a shared folder must not be able to get its contents by asking the assistant a question. It is a point we check explicitly at acceptance.

In the tools your teams already use: email, order management, shared folders, ERP, CRM or business software. No platform to learn from scratch — that is the leading cause of an AI project being abandoned in a small business. Depending on the case, the assistant takes the form of a panel inside your existing software, a search across your file space, or an answer drafted straight into an email.

When an interface has to be built, an order-management system connected or data stored properly, we leave configuration behind and enter development: that is our other trade, and it is costed separately.

A human, in every case where the output leaves the company. The assistant prepares, proposes and cites its sources; it sends nothing of its own accord to a client, a supplier or a public body. That rule is not window dressing: it is what makes the tool deployable without turning every answer into a commercial risk.

We measure the sign-off rate — the proportion of answers used untouched — because it says more than anything else about whether the assistant is useful. A low rate points to a badly chosen document scope, not to an inadequate model.

Like every AI project here, it starts with a proof of concept. The costing breaks down as follows: proof of concept on one use case, delivered in two to four weeks; rollout and integration with your tools; monthly subscription for hosting, monitoring and support; team training, per session.

Model usage is recharged at cost and generally stays marginal at the scale of a small business. What moves the figure is the number of document sources to connect and how finely the access rights have to be mirrored.