AI providers and office keys
Use OpenAI, Claude, Gemini, Mistral or Ollama with the office's keys or your own, and decide the AI rules.
Updated on
This guide is for the advisory firm's team.
The assistant and invoice reading use OpenRouter with free models unless you change something. In AI Assistant settings you can also use your own OpenAI (ChatGPT), Anthropic (Claude), Google Gemini, Mistral or Ollama (a model on your own network) key.
Your provider and model
Under Your provider and model you choose the provider, the model and your keys (one per line, in order of preference). Test key and model checks they work without using up the month's limit, and See the models on my account suggests those your account offers. Keys are stored encrypted and are never shown again.
Office keys
Management sets up, under Your office's AI, the providers the whole team uses: assistant model, invoice-reading model, shared keys and the order they are tried in (arrows ↑ ↓). If a person enters their own keys, those come first; if they fail, the office's are used.
Office rules
- Only EU providers or our own network: leaves out those that send data outside Europe (Mistral and Ollama are fine).
- Allow personal keys: if you untick it, everyone uses the office's.
- When reading invoices with AI: the AI proposes and a person always reviews, or it decides on its own when it is sure (with the minimum confidence you set).
- Read the PDF text first: cheaper and faster; if something does not add up, it is read again from the image so reliability is not lost.
- Split PDFs containing several invoices.
Local OCR: reading scans without sending them out
Under Office rules there is a selector called Text recognition on the server (local OCR):
- Do not use it.
- As a backup (the usual choice): used only if the model cannot read the image (for example, a model that does not accept images). The text is recognised on the server and only that text is passed to the model.
- Nothing leaves the server: no document is sent to an outside provider. The text is read here (the PDF's own or OCR) and only a local model (Ollama) processes it; other providers stop being used for invoices.
OCR is slower and makes more mistakes than a vision model, especially with poor photos and accents. That is why an invoice read this way is always reviewed by a person, even when the AI decides on its own, and the review screen says the text comes from OCR.
Usage and cost
At the bottom you will see this month's queries, tokens and estimated cost by provider and use. It is an estimate with public rates: the real figure is your provider's.
Firm's voice
In AI Assistant settings → Firm's voice management writes how the firm speaks («use the informal you, short sentences, no jargon, close by offering a call»). The AI imitates it when drafting email replies; it does not affect summaries or other functions. Always review the draft before sending.
What data leaves, how long it takes and what fails
In AI Assistant settings management sees, besides the cost, each provider's average latency and failures (if one fails, the query moves to the next on the list). Below that, Data sent to external providers (30 days): how many queries left and how much personal data was masked (people, tax IDs, IBANs…). Only counts are kept, never the content. Images and PDFs sent for reading cannot be masked: for no document to leave the server, choose «Local only» in the OCR. In Security → Team access you see who has two-step verification.
Related guides
Plan limit notices
When we let you know you are approaching your plan's limit on users, storage, invoices or AI.
AI limits per function
Cap how many queries each AI function can use (invoice reading, assistant, email) on top of the firm's general limit.
Firm's monthly report
The summary the management receives each month: invoices, deadlines and portal use.
Guilda Work