Your provider, your key
In Setup → AI, an admin picks a provider and pastes the company's API key:
| Provider | Default model |
|---|---|
| Mistral AI (EU) | mistral-small-latest |
| OpenAI | gpt-4o-mini |
| Anthropic (Claude) | claude-sonnet-5-5 |
| Ollama, on your own server | llama3.1 |
| Any OpenAI-compatible endpoint | your choice |
The model, the endpoint and the temperature can be changed; the key is encrypted at rest and never shown again; Test checks that it all works before anyone relies on it. Your AI bill stays between you and your provider, and the assistant stays off until a provider is set.
Emails drafted from the relationship
On a contact or a lead, Draft the email writes a first version for three purposes:
- a follow-up after the last exchange;
- a meeting request, offering two time slots;
- a first introduction.
The draft uses the five latest activities with the person — calls, emails, meetings — so it refers to what actually happened, stays under 140 words, starts with a subject line, and is written in the user's language: English, French, Spanish or Greek. On a deal, the follow-up draft takes the deal's stage and next step into account. You read it, adjust it, and send it from your own mailbox.
A brief before the meeting
On an account, Meeting brief prepares five points before you walk in: the relationship so far, the open deals, the service issues, the money owed — overdue invoices included — and what to raise, ending with a suggested goal for the meeting. Two minutes of reading instead of twenty of clicking through tabs.
Ask in plain words
Ask the CRM answers questions such as:
- "My open opportunities over 10,000 closing this quarter"
- "Hot leads from Lyon"
- "High-priority cases still open"
The model never writes a database query. It turns the question into a small plan — which object, which fields, which conditions — and SoliCRM checks that plan before running it: only an allowed list of fields per object, only simple comparisons, at most 50 results, and only records the user is allowed to see. A question it cannot turn into a safe search falls back to an ordinary keyword search.
Case summaries for the next agent
On a support case, Summarise case gives three lines — the problem, what was tried, where it stands — and the next step, from the case's activities and up to twenty chat messages.
And what is deliberately not AI
Some things are better as rules you can read than as a model's guess. The lead score and the deal risk are computed from explicit signals — rating, source, title, company size; overdue close date, no activity for 14 days, no next step — and every reason is shown. Nobody has to trust a number they cannot explain.
In short
- Mistral (EU), OpenAI, Claude, your own Ollama server or any compatible endpoint — your key, encrypted.
- Email drafts from real activity, in the user's language.
- Meeting briefs with open deals, issues and money owed.
- Plain-language questions turned into checked searches that respect access rights.
- Case summaries, and rules — not guesses — for scores and risk.