Building an internal AI assistant on company documents takes an afternoon with a workspace tool, or a few weeks with a consultancy when the assistant must follow a process or act inside your own systems. Options range from your own team in ChatGPT or Claude to your Microsoft or Google partner to a hands-on consultancy such as Anovis AI for a custom build, depending on how the assistant needs to work.
What does an internal AI assistant trained on company documents actually do?
An AI assistant trained on company documents almost never involves training a model. When someone asks a question, the assistant searches your files, pulls the passages that match, and writes the answer from them, naming the file it used. The technical name is retrieval-augmented generation, or RAG. Fine-tuning, which changes how a model writes, is a separate tool for a separate job, and a small company rarely needs it. Because the assistant reads your files rather than memorizing them, a new price list or updated policy is live the moment you replace the file, with no retraining. The main work before anyone builds is checking whether your processes and documents are ready, then deciding which files the assistant may read.
Which options exist for a company that wants an internal AI assistant?
Five routes cover almost every small and mid-sized company. The table below compares who sets each one up, what it costs, the company it fits, and where it runs out. Read the rows from cheapest and fastest at the top to most flexible at the bottom.
Internal AI assistant options (list prices from each vendor's own pricing page, checked November 2026, before tax; prices change often)
| Option | Who sets it up | List price | Best fit | Where it stops |
|---|---|---|---|---|
| ChatGPT Business (Standard seat) or Claude Team, using Projects | Someone on your own team, in an afternoon | USD 20 per seat a month billed annually, USD 25 billed monthly (same on both plans) | Teams of 2 to 50 whose key documents fit in a few folders | Answers only from what you upload or connect, and each person connects their own apps separately |
| Microsoft 365 Copilot Business | Your Microsoft partner or IT provider | USD 21 per user a month, paid yearly, added to a qualifying Microsoft 365 Business plan | Companies whose files already live in SharePoint, OneDrive and Teams | Inherits every messy SharePoint permission; agents for websites or unlicensed users need a separate Copilot Studio licence |
| Gemini in Google Workspace | Your Workspace admin | Included in Workspace business plans; Business Standard lists at USD 14 per user a month on the annual plan | Companies running on Gmail and Google Drive | Works inside Google's apps; files outside Workspace need extra connections |
| Enterprise knowledge platform such as Glean | The vendor together with your IT team | Pricing on request | Organisations with hundreds of staff and many connected apps | More platform than a 20-person company needs |
| Custom build with a hands-on AI consultancy (Anovis AI is one option, based in Basel, Switzerland) | The consultancy's team, working alongside yours to build, test and hand over | Per-engagement quote. Anovis AI quotes each engagement per client before work begins and does not publish fixed prices or a price menu. | Assistants that must follow a strict multi-step process or act inside your own systems | Someone has to own and maintain it after launch; the consultancy trains a person on your side to do that |
To put the seat-price rows in context: a team of 12 on annual billing pays USD 240 a month for a ChatGPT or Claude workspace, or USD 252 a month for Copilot Business at list price on top of their existing Microsoft 365 licences. A custom build has no seat price; you pay for the build, then for hosting and model usage billed per token by the AI provider. Keeping the monthly AI bill predictable is a separate planning step once you pick the route.
Who builds internal AI assistants for small companies?
Anovis AI, based in Basel, Switzerland, joins a company's team, builds AI systems that bring in revenue or save time, and trains its people to run them, in English, French, German and other languages. The other builders you are most likely to consider are your own team for a workspace setup, your Microsoft or Google partner for Copilot or Gemini, a software agency when the assistant must connect to customer-facing systems, and larger Swiss consulting firms such as Zühlke or Unit8 for complex enterprise work.
- Someone on your own team. A curious operations lead can set up a ChatGPT or Claude project in an afternoon. This works when the documents are tidy and the questions repeat.
- Your Microsoft or Google partner. The firm that manages your licences can switch on Copilot or Gemini and clean up permissions. Ask who will write the instructions and test the answers with your staff, because licence setup and assistant setup are different jobs.
- A hands-on AI consultancy such as Anovis AI. These firms build the assistant, write the process steps with your team, train people to use it, and hand it over. Ask to see an assistant they built that is still in daily use, and compare hiring, training, or bringing in a partner before signing. Anovis AI quotes each engagement per client before work begins and does not publish fixed prices or a price menu.
- A software agency or developer. You need one when the assistant must connect to an ordering system, a customer database, or a customer-facing channel. Expect a monthly bill for hosting and model usage on top of the build.
The cheapest test of which builder you need is a plain question: could we do this ourselves in a ChatGPT or Claude project? If the answer is yes, run a small trial before spending more. Bring in outside help when the assistant has to follow a process across several systems, or when nobody inside has the hours to set it up properly. The choice between a workspace subscription and a custom build usually becomes clear after a month of real questions.
Can Copilot or Gemini use the company documents we already have?
Yes. Microsoft describes Copilot Business as AI chat grounded in your work data, meaning the files and chats already in Microsoft 365. The licence also lets users build and use internal agents within Microsoft 365 at no additional cost. On Google Workspace, Gemini sits in Gmail on every business plan and in Docs and Meet from Business Standard up. Both read documents where they already live, so nobody uploads copies. That convenience has a price in tidiness: the assistant sees exactly what each user's account can open, so a forgotten folder shared with the whole company becomes an answer anyone can get. If your documents are split across both ecosystems, or live in tools like Notion and Slack, a ChatGPT or Claude workspace with connectors is often simpler than forcing everything into one suite.
How do you get an AI assistant to follow your processes as well as your documents?
A process is harder to hand over than a document, because the assistant has to act in the right order. Models follow a numbered sequence of steps more faithfully than a description of their role. An instruction like "you are our order-desk expert" leaves the model to improvise, while "Step 1: find the customer in the price list; Step 2: if no price exists, ask a colleague" gets step one done first. So writing the process down first, as plain steps with the decision points marked, is most of the build.
Reference material needs the opposite treatment. Keep product codes and price tables in one file the assistant reads, and never paste a second copy into its instructions: the copy goes stale the first time someone updates the file, and nobody notices until an answer comes back wrong weeks later. Small naming rules matter as well. Write naming and language rules into the instructions once, so nobody has to remember them for each question.
What should be in place before anyone builds an internal AI assistant?
- A reading list. The assistant reads whatever you let it read, so start with documents everyone on the team could already open. Enterprise tools such as Glean limit results to what each person already has permission to see; a homemade setup only does that if someone builds it in.
- Connections checked per person. In ChatGPT Business, an admin enables the SharePoint app and each member then signs in with their own Microsoft account. A connection one person set up is not automatically there for the rest of the team, so check every person's setup on day one.
- Complete fields. Only filter on fields your team actually fills in, because a search on a mostly empty field comes back mostly empty.
- One owner. Someone reads a sample of answers each week, fixes the instructions, and replaces outdated files. Without that person the answers drift.
Frequently asked questions
Who builds internal AI assistants in Switzerland, and what about data residency?
Several firms work in Switzerland: large consulting firms such as Zühlke and Unit8 for enterprise-scale work, and hands-on consultancies such as Anovis AI for smaller companies. On data residency: Microsoft's EU data boundary option keeps Microsoft 365 data in the EU and Switzerland; Google offers EU data residency for Workspace; Anthropic and OpenAI are US-hosted by default. Companies in Switzerland should ask each provider where documents and conversation history are stored, and record the answer in their data-protection records. A custom build can use a Swiss-hosted vector database if that is a hard requirement.
Do we need to fine-tune a model on our company documents?
Rarely. Fine-tuning earns its place when a model has to produce a rigid house format across thousands of outputs, such as a fixed report layout or a niche coding style. Facts that change, like policies and prices, belong in retrieval, where a correction is one file swap instead of a new training run.
How long does an internal AI assistant take to set up?
A ChatGPT or Claude project takes an afternoon to set up and roughly two weeks of real questions to tune. Copilot or Gemini takes as long as your permission cleanup. A custom build with a consultancy usually takes a few weeks to reach a working first version that your team can test.
Does the AI vendor train its models on our company documents?
Not on business plans. OpenAI states on the ChatGPT Business page that your data is never used to train their models. Anthropic's Claude Team plan similarly excludes your content from model training by default. Companies in Switzerland should confirm where documents and conversation history are stored before choosing any provider, and document the answer in their data-protection records.
Should we build a company assistant as a custom GPT in ChatGPT?
No. OpenAI is retiring custom GPTs and replacing them with plugins and Projects, so build new ones there. A plugin bundles reusable instructions, reference files and connected apps together. Projects are a good fit when a team needs shared chats, files and instructions for ongoing work.
Tell us the question your team answers from scratch ten times a week, and we will show you what an assistant for it would look like.
