# How to build an AI assistant that answers from your company documents

URL: https://yarigai.ai/guides/ai-knowledge-assistant
Author: Deniz Sancar, founder of Yarigai (https://denizsancar.com)
Publisher: Yarigai, a done-for-you AI agency in London, UK (https://yarigai.ai)
Published: 7 November 2026. Updated: 7 November 2026.

## Short answer
An AI assistant that answers from company documents costs £15 to £18 a seat a month off the shelf in the UK in 2026, on ChatGPT Business or Claude Team. A custom build starts at **£6,000 plus £399 a month** on one London agency's price list. Anthropic says a document set under about 200,000 tokens, roughly 500 pages, can go straight into the prompt. Larger or fast-changing sets need retrieval, which finds the right passages for each question and cites them.

## At a glance
| Option | Price as of November 2026 | Fits when |
|---|---|---|
| ChatGPT Business | £15 a seat a month billed annually, £18 monthly, 2 seats minimum | Staff ask questions of shared files and connected apps |
| Claude Team | £18 a seat a month billed annually including 20% VAT, £22.50 monthly | Staff ask questions of shared project files |
| Microsoft 365 Copilot Business | £16.10 a user a month plus VAT, billed annually, on top of Microsoft 365 | Files already sit in SharePoint, OneDrive, Outlook and Teams |
| Google Gemini Notebook | Included with Google Workspace; Business Standard £11.80 a user a month excluding VAT | A team works from a fixed set of sources |
| Documents in the prompt | API tokens only | Under about 200,000 tokens (500 pages) that rarely change |
| Custom retrieval (RAG) assistant | From £6,000 setup plus from £399 a month, excluding VAT (Softomate, October 2026) | Customers ask, answers must cite a source, or access differs by person |
| Wrong answers in retrieval-based legal tools | More than 17% to more than 34% of test queries (Stanford, 2024) | Every build needs a test set before launch |

UK prices from each provider's page, checked 7 November 2026.

## What a document assistant is and the three ways to build one
A document assistant is an AI model that answers questions from your own files, such as policies, price lists, contracts and help articles, instead of from what it learned in training. There are three ways to give a model your documents.

1. A business workspace. Staff upload files to a ChatGPT, Claude, Copilot or Gemini workspace and ask questions there. No build is needed.
2. Documents in the prompt. A small build sends the whole document set to the model with every question.
3. Retrieval, often called RAG. A build searches the documents for each question and sends only the matching passages to the model, so each answer can link to its source.

Fine-tuning retrains a model on your data, and the research favours retrieval over it for facts. The guide to [custom LLMs, fine-tuning and RAG](/guides/custom-llm-vs-rag) covers the research and the costs.

## When documents fit in the prompt and when you need retrieval
Use the prompt for small, stable sets and retrieval for large or changing ones. Anthropic's Contextual Retrieval write-up of 19 September 2024 says a knowledge base under 200,000 tokens, about 500 pages, can go into the prompt whole. That works out at about 400 tokens a page.

Retrieval earns its place under any of three conditions. The set is larger than about 500 pages, it changes every week, or different people may see different documents. For retrieval, Anthropic's contextual method cut the share of failed searches by 49%, from 5.7% to 2.9%, and by 67% to 1.9% when combined with reranking.

"Fits in the prompt" is about accuracy and build effort, not cost. Sending a large set with every question gets expensive at volume.

### A worked example at 1,000 questions a month
Take a 300-page staff handbook and policy set, about 120,000 tokens at 400 tokens a page, and 1,000 staff questions a month on Claude Sonnet 5.5 at $2 input and $10 output per million tokens. Assume 400 output tokens an answer.

- With the whole set in the prompt, 120,000 × $2 ÷ 1,000,000 = $0.24 of input per question, plus $0.004 of output. That is $0.244 a question, or **$244 a month** before any caching.
- With retrieval, 1,000 tokens of instructions, five passages of 500 tokens and a 100-token question make 3,600 input tokens. That is $0.0072 plus $0.004, or $0.0112 a question, and **$11.20 a month**.
- Turning the handbook into searchable vectors with OpenAI's text-embedding-3-small, at $0.02 per million tokens, costs 120,000 × $0.02 ÷ 1,000,000 = $0.0024.

A 20-page FAQ of about 8,000 tokens is different. It costs $0.016 of input a question in the prompt, or $20 a month with output, and needs no search system at all. These token counts are assumptions, so measure your own documents.

## Off-the-shelf options and UK prices
For staff questions, a business workspace costs less than any build. Ten ChatGPT Business seats cost 10 × £15 × 12 = £1,800 a year on annual billing.

| Product | UK price, November 2026 | How it uses your documents | Limits to know |
|---|---|---|---|
| ChatGPT Business | £15 a seat a month annual, £18 monthly | Projects, file uploads and connectors to Google Workspace, Microsoft 365, Slack and GitHub | 2 to 200 staff |
| Claude Team | £18 a seat a month annual including VAT, £22.50 monthly | Projects with shared knowledge files | When project knowledge nears the context limit, Claude switches to retrieval and raises capacity by up to 10 times |
| Microsoft 365 Copilot Business | £16.10 a user a month plus VAT, annual (£13.80 first-year offer to 31 December 2026) | Email, chats, files and meetings through Microsoft Graph | Needs an eligible Microsoft 365 Business plan |
| Google Gemini Notebook, formerly NotebookLM | Included in Google Workspace; Business Standard £11.80 a user a month excluding VAT | Notebooks built from uploaded sources | 50 sources per notebook on the standard tier, 100 on Plus and 300 on Pro; Business Starter gets basic access only |

Claude Team members can share a project with colleagues as "Can view" or "Can edit". Google's Workspace page gives Business Standard and Business Plus expanded Gemini Notebook access, with higher notebook, query and source limits. The guide to [ChatGPT Business, Claude Team and Copilot](/guides/chatgpt-vs-claude-vs-copilot-for-business) compares the three in more detail.

A workspace stops being enough when customers ask the questions, when an answer must cite an approved source every time, when the assistant has to act in another system, or when access rules are finer than a shared folder.

## How to build one, step by step
A custom assistant takes eight steps, and the first two need no code.

1. Write 30 real questions that staff or customers ask, with the right answer and the document it comes from. This is your test set.
2. Gather the documents, delete out-of-date versions and name one owner for each document.
3. Measure the set at about 400 tokens a page. Under 200,000 tokens and rarely changing goes in the prompt. Anything larger or changing weekly needs retrieval.
4. For retrieval, split the documents into passages, turn each into a vector and store them. OpenAI's hosted file search gives 1 GB of storage free, then charges $0.10 per GB a day and $2.50 per 1,000 searches.
5. Answer only from the retrieved passages, link every answer to its source passage, and send questions with no good match to a person.
6. Filter by permission before the search runs, so each user only ever retrieves passages from documents they can already open.
7. Run the test set and agree a pass mark before launch.
8. Re-index each document when it changes, and read the unanswered questions every week to find gaps in the documents.

### What a custom build costs
Softomate, a London agency, lists a RAG company knowledge assistant from £6,000 setup plus from £399 a month, excluding VAT, updated 1 October 2026. Year one costs at least £6,000 + (12 × £399) = £10,788. That is six times the £1,800 a year for ten ChatGPT Business seats, so a build has to do something a seat cannot. Softomate's timeline guide puts a simple chatbot with under 50 documents at 4 to 6 weeks and one with a CRM connection at 6 to 8 weeks. No independent survey covers these builds, so the figures come from a firm that sells the work.

## Accuracy, citations and permissions
Retrieval does not stop wrong answers. Stanford RegLab and HAI tested legal research tools built on retrieval in 2024. Lexis+ AI and Ask Practical Law AI gave incorrect information more than 17% of the time, and Westlaw AI-Assisted Research more than 34%.

Three rules keep an assistant honest:
- Every answer links to the passage it came from, so a reader can check it in one click.
- The assistant says it doesn't know when no passage matches, and routes the question to a named person.
- The test set runs again after every change of model, prompt or document set.

Permissions need the same care as accuracy. Microsoft says Copilot uses "the same controls and permissions that help protect your data in Microsoft 365", so it answers each user only from what they can already open. A custom build has to enforce the same rule at the search step. An assistant that indexes HR files alongside the staff handbook, with no filter, can quote salary letters to anyone who asks the right question.

## Data protection under UK GDPR
Business plans from OpenAI and Anthropic don't train on your data by default. ChatGPT Business lists "No training on your business data by default", and Claude Team lists "No model training on your content by default". Personal accounts work differently, so staff should not upload company files to them.

Under UK GDPR the business stays the controller of any personal data it loads into an assistant. The ICO's guidance on AI and data protection, last updated on 15 March 2023 and under review after the Data (Use and Access) Act, covers impact assessments, lawfulness, transparency and statistical accuracy. Run a data protection impact assessment before you index customer or staff records, and sign the provider's data processing terms.

Retrieval is easier to keep compliant than fine-tuning. A search index is a store of documents you can edit or delete, so removing one person's data means deleting their passages and re-indexing.

## How Yarigai builds assistants that answer from company documents
Yarigai builds support agents that answer from a client's docs, and document reading and filing, inside the client's own accounts and tools. The client owns the code, the data and the accounts from day one, and there is nothing to sign up for. AI and hosting bills are paid by the client at cost, and the monthly running cost is on the quote.

A project starts with a free 30-minute scope call on Google Meet, and the client keeps the notes whether they hire Yarigai or not. A blueprint of about a week, credited to the build, sets out a written plan, a clickable prototype, the monthly running cost, a fixed price and the payback maths. The build takes 2 to 6 weeks, with a demo of working software every week. A person approves anything that sends money, emails a customer or deletes data.

Yarigai's guarantee reads "If the build isn't profitable for you, we won't build it." If a ChatGPT or Claude workspace does the job, the blueprint maths will show it, and Yarigai won't take a build that doesn't come out ahead within the first year. More on [AI customer support](/services/ai-customer-support).

## Common questions
### Can ChatGPT Business answer questions from our Google Drive or SharePoint files?
Yes, through connectors. OpenAI's pricing page lists connectors to Google Workspace, Microsoft 365, Slack and GitHub for ChatGPT Business, which costs £15 a seat a month on annual billing in the UK. The page names the suites rather than Drive or SharePoint, so test the connector on your own folders first.

### Is NotebookLM free for business use?
Google now calls it Gemini Notebook, and its help pages list a standard tier with 100 notebooks and 50 sources per notebook, with higher limits on paid plans. Google Workspace includes it. Business Starter gets basic access, and Business Standard, at £11.80 a user a month excluding VAT, gets expanded access.

### Do I need a vector database to build a document assistant?
Not for a small set. Anthropic says documents under about 200,000 tokens, roughly 500 pages, can go straight into the prompt. Above that, a hosted option such as OpenAI's file search includes 1 GB free, then $0.10 per GB a day.

### How do I stop the assistant making things up?
Make it answer only from retrieved passages, link every answer to its source and route questions it can't answer to a person. Stanford found retrieval-based legal tools still gave incorrect information more than 17% of the time in 2024 tests. Test on 30 real questions before launch and after every change.

### Will staff see documents they shouldn't?
Only if the build ignores permissions. Microsoft says Copilot works within each user's existing permissions, and Claude Team projects can be shared as view-only or editable. A custom build should filter passages by the user's access before the search runs.

### How long does it take to build a custom document assistant?
Softomate puts a simple chatbot with under 50 documents at 4 to 6 weeks, and one with a CRM connection at 6 to 8 weeks. Yarigai builds in 2 to 6 weeks, and a single workflow can go live in about two weeks. Writing the test set and cleaning the documents take more time than the code.

## Sources
1. [Introducing Contextual Retrieval](https://www.anthropic.com/news/contextual-retrieval), Anthropic, published 19 September 2024.
2. [What are projects?](https://support.claude.com/en/articles/9517075-what-are-projects), Claude Help Center, checked 7 November 2026.
3. [Plans and pricing | Claude by Anthropic](https://claude.com/pricing), Anthropic, UK Team prices and API prices checked 7 November 2026.
4. [Plans and Pricing | ChatGPT](https://chatgpt.com/pricing/?type=team), OpenAI, Business features and UK prices checked 7 November 2026.
5. [Microsoft Copilot for Business: Enterprise AI Solutions](https://www.microsoft.com/en-gb/copilot/solutions/business), Microsoft UK, checked 7 November 2026.
6. [Learn about Gemini Notebook's plans](https://support.google.com/notebooklm/answer/16213268), Google, checked 7 November 2026.
7. [Compare flexible pricing plan options | Google Workspace](https://workspace.google.com/intl/en_uk/pricing), Google, UK prices checked 7 November 2026.
8. [Pricing](https://developers.openai.com/api/docs/pricing), OpenAI, embeddings and file search checked 7 November 2026.
9. [AI Automation Pricing for UK Businesses: What Every Service Costs and What ROI to Expect in 2026](https://www.softomatesolutions.com/blog/ai-automation-pricing-uk-businesses-2026/), Softomate Solutions, updated 1 October 2026, checked 7 November 2026.
10. [How Long Does It Take to Build a Custom AI Chatbot? A Real 4-8 Week UK Timeline with Milestones](https://www.softomatesolutions.com/blog/how-long-to-build-ai-chatbot-uk-timeline/), Softomate Solutions, updated 1 October 2026.
11. [AI on Trial: Legal Models Hallucinate in 1 out of 6 (or More) Benchmarking Queries](https://hai.stanford.edu/news/ai-trial-legal-models-hallucinate-1-out-6-or-more-benchmarking-queries), Stanford HAI, published 2024.
12. [Guidance on AI and data protection](https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/artificial-intelligence/guidance-on-ai-and-data-protection/), Information Commissioner's Office, last updated 15 March 2023.