Contents
An internal AI assistant finds information in your documents and answers with its sources, without your data being used to train a model. A first assistant covering a defined set of documents starts at $2,300 (€2,000)excl. VAT, in 2 to 8 weeks, plus AI API usage costs.
At many companies, the information exists but nobody can find it: the procedure sits in a shared folder, the technical answer in an old email, the right version of the contract with a colleague who is out. An AI assistant connected to your documents answers those questions in seconds, provided it is built properly. Here is how I approach it, what it costs, and when a plain search engine is enough.
How a document assistant works
The principle is simpler than it sounds, and it does not involve training an AI.
- Indexing. Your documents are split into passages and indexed in a search database, on your premises or with your hosting provider.
- Search. When an employee asks a question, the tool looks for the most relevant passages, taking their access rights into account.
- Answer. Only those passages are sent to the AI model, with instructions to answer from them and cite their sources.
- Verification. The answer shows links to the original documents, so the user can check in one click.
This setup has a decisive advantage: when a procedure changes, you only update the document, and the assistant picks it up at the next indexing run.
Which documents work well
| Type of documents | Value | Precautions |
|---|---|---|
| Internal procedures, work instructions | High: frequent questions, stable answers | Remove outdated versions |
| Product sheets, technical catalogs | High for sales and support teams | One reference source per product |
| Terms and conditions, standard contracts | Useful to find a clause | Not a substitute for legal advice |
| Ticket or customer reply history | Useful for support | Strip unnecessary personal data |
| HR, payroll, executive documents | Possible but sensitive | Strict access rights, separate scope |
| Meeting notes, loose emails | Low: contradictory and dated content | Avoid at first |
The quality of the assistant depends directly on the quality of the documents. An up-to-date procedures folder produces good answers; three conflicting versions of the same procedure produce conflicting answers.
What it costs
These amounts exclude VAT and come from my pricing.
| Situation | What is delivered | Indicative budget | Timeline |
|---|---|---|---|
| First assistant on a defined scope | Indexing, question interface, cited sources, testing | From $2,300 (€2,000)excl. VAT | 2 to 8 weeks |
| Assistant built into a business application | Assistant + accounts and roles, connection to your tools, usage tracking | $6,800 to $22,400 (€6,000–€20,000)excl. VAT | 1 to 3 months |
In the estimator, built-in AI counts for $2,800 (€2,500)excl. VAT, accounts and roles for $1,700 (€1,500)excl. VAT, and connecting to other tools for $2,300 (€2,000)excl. VAT. API usage, billed by the AI provider, and hosting for the index come on top. The choice of model depends on your quality, cost and data location requirements: the page on choosing an AI model covers the criteria.
Essential safeguards
- Access rights respected: the search filters by each user’s permissions, on the same principle as role-based access control in an application.
- Data under control: only the useful passages go to the API, with a provider whose business terms exclude training on your data. See what data to share with AI.
- Sources always shown, and an “I can’t find this information” answer rather than a made-up one.
- A set of test questions, written with your subject-matter experts and rerun after each update.
- User feedback: a “helpful / not helpful” button that feeds the corrections.
When an AI assistant is not the right solution
- You have few documents, or they are poorly maintained: start by organizing and updating them. A well-structured internal wiki often solves the problem.
- The questions are about live figures (stock, orders, revenue): you need a dashboard or a query on your database, not a document assistant.
- A tool you already use offers AI search across your files and meets your need: try it before investing in development.
- The answers commit the company legally: the assistant can help find a text, not decide in place of an expert.
Where to start
- Pick a narrow scope: one department’s procedures, or the technical catalog, not “all company documents.”
- List about thirty real questions your teams ask often, with the correct answer: they will serve as tests.
- Run a one-week audit, then a prototype tested by a few users on their real questions.
The approach, from scoping to deployment with human oversight, is detailed on the AI integration page.
Frequently asked questions
Do we need to train an AI on our documents?
No. The usual approach is to index your documents, then send the model only the relevant passages for each question. The model itself is not changed, your documents stay with you and can be updated at any time.
Can the assistant make up an answer?
The risk exists. It is reduced by requiring the assistant to answer from the passages it found, to cite its sources and to say "I don’t know" when nothing matches. A set of test questions, validated by your experts, checks quality before and after rollout.
Could an employee see documents they are not allowed to access?
Not if the assistant is built correctly: the search filters documents based on the logged-in person’s permissions before anything is sent to the model. HR or executive documents stay invisible to those without access.
Which document formats are supported?
The common ones: PDF, Word, intranet or wiki pages, simple spreadsheets, content exported from a business tool. Poor-quality scans and highly visual documents need preprocessing.