Connecting an AI provider, or running one yourself

Setting up AI drafting in Pental: bringing a provider key, pointing at a local Ollama instance, or leaving it off entirely. Where the key lives and what each route needs.

8 min read

AI drafting in Pental runs on your own provider account or your own hardware. There is no marked-up token and no per-seat assistant fee, because the platform never sits between you and the model. This guide covers configuring each route from Settings, AI Assist.

Route one: a hosted provider on your key

Create an API key in your own OpenAI, Anthropic or Groq account.
Paste it into the AI settings and choose the model you want to use.
Save. The key is encrypted on write inside your own database, using a key generated in that same database, and the plaintext column is nulled.

Requests then go to the provider directly, under your agreement, your data-handling terms and your billing. Nothing is proxied through Pental, so there is no shared quota to queue behind.

Route two: a local model

If you would rather nothing left your network, run Ollama on a machine inside it and point the portal at that endpoint. Firms doing this already tend to have suitable hardware: the GPUs bought for password auditing are more than capable of serving a drafting model.

Install Ollama on the host and pull a model appropriate to its memory.
Make the endpoint reachable from wherever the portal runs, over your own network or VPN.
Enter the endpoint URL and model name in the AI settings, along with the context size and token ceiling you want to allow.
Generate one finding as a test before relying on it for a live engagement.

Local inference has no rate limits and no per-token cost, so the platform can afford longer, more careful passes over an engagement. In practice that often produces better drafts than a rate-limited hosted route.

Route three: any compatible endpoint

A gateway you operate, a regional provider or a fine-tuned deployment all work if they speak a compatible API. Set the URL, model, request format, headers and token ceiling and the platform will use it like any other route.

Making the output sound like your firm

Every generated field has its own prompt, visible and editable in settings, with a reset to the shipped default. Put your house style, structure rules and banned phrases in the prompt once and every draft follows them. This is the difference between drafts that read like your firm and drafts that read like everyone.

Leaving it off

Configure no provider and the AI is simply absent: no buttons, no prompts, no degraded mode. Proposals, engagements, findings, QA, reports, the portal and invoicing all work without it.

Also worth reading

Setup