Configuring AI
Choose which model provider serves your workspace's AI, connect its key, customize prompts, and see what it costs.
Pact's AI features — drafting, summaries, meeting prep, agents, and the rest listed on the AI Hub — run on a model provider your workspace chooses. The same gates apply whichever provider serves them: the kill switch, data-sharing mode, budget, prompt-injection screening, and PII redaction all run before a request leaves Pact.
Supported providers
| Provider | How you connect it | What serves your AI |
|---|---|---|
| Anthropic | An Anthropic API key under Admin → Integrations → Anthropic | Claude, billed to your Anthropic account |
| OpenAI | An OpenAI API key under Admin → Integrations → OpenAI, then choose OpenAI under AI Hub → Model provider | OpenAI's equivalent model for each tier, billed to your OpenAI account |
| AWS Bedrock | AWS credentials and a region under Admin → Integrations → AWS Bedrock, then choose Bedrock under AI Hub → Model provider | Claude in your own AWS account, billed by AWS |
| Custom endpoint | Any OpenAI-compatible endpoint (a vLLM or Ollama server, a gateway) under Admin → Integrations → Custom LLM endpoint | Your model server |
Bedrock serves Claude models only. Other Bedrock model families are not supported.
A stored key does not change where your data goes
An OpenAI key is also used for semantic-search embeddings, so connecting one routes nothing else until an admin chooses OpenAI under AI Hub → Model provider. The same rule applies to Bedrock credentials. Switching providers asks for confirmation and names where your data will go and who is billed. If step-up verification is on for your workspace, it also asks you to re-verify.
Before you switch, Test on the Model provider card sends one short request through every gate to the candidate provider. It shows the model, latency, tokens, and cost, or the reason the provider refused.
Which provider answers
For each request Pact resolves the provider in this order:
- A custom endpoint that an admin has scoped to that specific feature.
- An OpenAI or Bedrock credential an admin chose under AI Hub → Model provider.
- The workspace's Anthropic key.
- A custom endpoint with no feature restriction.
If none of these is configured, AI features say that AI isn't set up for the workspace. If an admin chose OpenAI or Bedrock and that credential can't be read, the request fails with that reason; it does not quietly move to another provider. The Model provider card shows which provider the next request will actually use, read from the same resolver.
Key scoping
Keys are encrypted at rest and scoped to the workspace. A background job always uses the credential of the workspace that owns the job.
Customizing prompts
Every AI feature has a default system prompt and user template. A workspace admin can override them in Admin → AI → Prompt library. The library lists every feature in the prompt registry with its model tier and whether you've customized it.
- Edit opens the effective prompt and template (yours, or the default) for editing. Reset to default removes your override.
- Test runs the prompt once and shows the rendered message, the output, the model used, and the cost.
Where AI drafts appear
Draft with AI appears in the social and messaging inbox, the social reply composer, and a marketing campaign's content step. A draft is a starting point: you edit or regenerate it before anything is sent.
Lead scoring
Lead-scoring signal weights and thresholds are configured in Admin → Lead scoring.
Usage and cost
Every AI call writes a row to the usage ledger. Admin → AI → Cost shows daily spend by feature and provider, the period total, and a month-to-date projection. Budgets and the workspace AI kill switch are in Admin → Security → AI budget.
With your own provider key, the provider bills you directly at its listed rates.
Turning AI off
The AI kill switch in Admin → Security → AI budget stops every AI call for the workspace immediately. Everything that doesn't need a model — CRM records, imports, sequences, consent — keeps working.