An AI assistant on your own documents: what a UK SME gets from a knowledge assistant, and what it costs.
By Zain M · 7 October 2026 · 12 min read
A knowledge assistant is an AI that answers questions from your own documents and data, retrieving the relevant passages and citing them, rather than answering from the model’s general memory. For a UK SME it is the most useful shape of AI for staff questions, policies, product information and customer support, because it is grounded and checkable. There are three ways to get one: a shared project in Claude or ChatGPT with the documents attached, which costs nothing beyond the seats; a configured assistant inside the business plan with connectors to your drive; or a custom build at £8,000 to £25,000 where the documents live in your systems, the answers must be audited, or customers use it. The work that decides quality is not the model but the documents: one current copy of each, the superseded ones removed.
What retrieval-augmented generation is, plainly
A language model on its own answers from what it learned in training, which is why it can be confidently wrong about your firm. Retrieval-augmented generation, RAG, changes the order: when a question arrives, the system first searches your documents for the passages most likely to answer it, then gives those passages to the model and asks it to answer from them, citing what it used. The answer is grounded in your material, it shows its sources, and when the documents do not cover the question a well-built system says so instead of guessing. That last behaviour is the whole point for a business.
What UK firms use it for
Staff questions about policy, process and product: how do we handle a refund, what is the holiday rule, what does this contract clause mean for us. Customer questions from your own information: order status, returns, allergens, opening hours, service details. Sales and tender support: what did we say last time, what is our position on this, draft the answer from our previous responses. Onboarding: a new starter asking the assistant instead of interrupting the person next to them. And professional work: a practice asking questions of its own precedents, notes and guidance, with the sources shown for a qualified person to check.
The three ways to get one
The first costs nothing beyond the seats. A shared project in Claude Team or a custom GPT in ChatGPT Business, with the relevant documents attached and an instruction to answer only from them and to cite. It suits a team of any size, a modest document set, and questions asked by staff rather than customers. The second is the configured assistant inside the business plan with connectors to Google Drive or SharePoint, so it reads the live folder rather than a snapshot; it suits firms whose documents already live tidily in one place, and it needs the connector permissions decided deliberately. The third is a custom build: the documents are indexed from your systems, the assistant lives in your website, app or phone line, answers are logged and auditable, and customers can use it. That is £8,000 to £25,000 at fixed-price UK rates, with a running cost that scales with questions asked.
| Route | Cost | Suits | Limits |
|---|---|---|---|
| Shared project or custom GPT | Seats only | Staff questions, a modest document set | Snapshot documents; no audit trail; not for customers |
| Configured assistant with connectors | Seats only | Firms with tidy shared drives | Reads whatever the drive permits; permissions must be reviewed |
| Custom build | £8,000 to £25,000 plus running cost | Customer-facing, audited, inside your systems | A build, with the discovery step first |
The work that decides quality
The model is the easy part. The documents are the hard part, and they are the same problem in all three routes: one current version of each policy, price list and process; the superseded versions removed rather than left in the folder; a source of truth named for anything that changes often; and a sample of fifty real questions with known right answers to test against before anyone relies on it. A knowledge assistant on a messy folder answers confidently from the 2023 price list, and the failure is blamed on the AI. Our data readiness guide sets out the month of tidying, which pays back whether or not the assistant follows.
The data rules
A knowledge assistant reads your documents, and many of them contain personal data. On a business plan of Claude or ChatGPT the content is not used for training and a data processing agreement applies; on a custom build the documents stay in your systems and the model sees only the passages retrieved for each question, which is the most contained shape. Where the documents include client or patient information, the ICO’s DPIA screening applies, and where customers use the assistant, it says it is an AI and hands to a person when unsure. Connectors deserve a specific decision: they let the assistant see whatever the drive lets the user see, so a permissions review comes before switching them on.
What it costs to run, and what we have learned
Retrieval is cheap and the model call per question is small, because the answer is generated from a few passages rather than the whole library. A small model handles most questions well when it is given the right passages, which is why a well-built assistant routes to the large model only for the hard ones. From running these in production: the assistant that says “the documents do not say” is the one staff trust; showing the source next to every answer does more for adoption than any accuracy figure; and the maintenance that matters is keeping the documents current, not retraining anything.
Common questions
What is a knowledge assistant?
An AI that answers questions from your own documents and data, retrieving the relevant passages and citing them, rather than from the model’s general memory. It is grounded, checkable, and says when the documents do not cover a question.
What is RAG?
Retrieval-augmented generation: search your documents for the relevant passages first, then have the model answer from those passages with sources. It is how a knowledge assistant stays accurate about your firm.
Can we build one without a developer?
Yes, for staff use: a shared project in Claude Team or a custom GPT in ChatGPT Business with the documents attached, instructed to answer only from them and cite. It costs nothing beyond the seats.
How much does a custom knowledge assistant cost in the UK?
£8,000 to £25,000 at fixed-price suppliers for one that lives in your website, app or phone line, indexes documents from your systems and logs answers, with a running cost that scales with questions asked.
Is it safe with confidential documents?
On a business plan the content is not used for training and a DPA applies; on a custom build the documents stay in your systems and the model sees only retrieved passages. Where personal data is involved, the DPIA screening applies.
Why does a knowledge assistant give wrong answers?
Almost always because the documents are wrong or out of date: the 2023 price list still in the folder, two versions of a policy. Tidy the documents and test on fifty real questions before relying on it.
Can customers use it?
Yes, as a custom build that says it is an AI, answers only from your information, shows sources where useful and hands to a person when unsure.
Want your documents answering questions?
We set up the shared project for staff at no build cost, or build the customer-facing assistant inside your systems at a fixed price, with the documents tidied first.