AI quoting for MSPs without vendor lock-in

If your MSP backend tool is going to call an LLM on your behalf, two questions matter:

1. Whose key is it calling on?
2. What does that cost you per token?

For most “AI-enabled” SaaS in this market, the answer is: their key, and there’s a markup baked into the seat price. You don’t see the token bill, you see a “20% AI surcharge” on the invoice. Switching providers — say, from a US-hosted model to one in Frankfurt for compliance — isn’t on the table.

We took the opposite route.

Bring your own model

Every AI feature in MSPercury — and there are several — calls your Anthropic, OpenAI, or OpenAI-compatible endpoint. You drop a key into Workspace Settings, pick a model, and from then on:

• You pay the provider directly. No markup on our side.
• You pick the model. Want Claude Sonnet for executive summaries and a cheap local model via Ollama for catalog matching? Configure both.
• You can swap providers any time without losing your data. The prompts are documented; the tool just routes.
• A self-hosted LLM with an OpenAI-compatible API (vLLM, LiteLLM, your own proxy) works the same as the hosted ones.

If you don’t configure a model, MSPercury runs in its classic mode. Every non-AI feature stays intact.

What the AI actually does

Concretely, today:

• Executive summaries for CheckUp PDF reports (2–4 sentences, no ChatGPT clichés).
• Quote drafts from a public CheckUp’s answers — 3 to 6 services from your catalog with quantities and a tailored summary, dropped straight into the normal draft-quote workflow.
• First-contact emails to inbound leads, anchored on their weakest CheckUp categories.
• Catalog matcher that links findings to your service items so the “generate quote from CheckUp” CTA isn’t gated on manual mapping.
• Status-update structurer — three hectic bullet points become a polite customer-facing post with the right category enum.
• Service-report builder that turns “did / noticed / recommended” into clean field-report prose.
• Project-task generator that converts findings into a prioritised plan with hours and rationale.
• Quote-reply drafter — suggests the next operator reply in a quote thread, pre-filled in the composer for review.

None of those auto-send anything. The operator is always the one who clicks the button that puts text in front of the customer. The audit trail (who/when) stays clean.

Why this is the right call

A managed service provider sells trust. Trust is bilateral: your customers trust you, and you need to trust your stack. Putting an opaque LLM intermediary between your operator’s notes and your customer’s inbox is a trust hand-off most MSPs don’t want to make blindly. Letting you choose the provider, see the prompts, and pay the bill directly is the cleanest way to keep that hand-off honest.

For MSPs in regulated markets — Germany, Spain, anywhere with a strict residency requirement — the same architecture lets you point the routing at an EU-hosted provider, or your own GPU. No “AI feature gated to a US-only model” friction.

What’s next

We’re calling MSPercury open for signups. The full feature set is documented at mspercury.com. The Partner Network and the cross-tenant Marketplace land as Pro features after the public beta — for now they’re open to every workspace so we can dogfood them on real partner relationships.

If you run an MSP and you want to try the quoting + audit + customer-portal stack with your own LLM, you can sign up here. Feedback to info@it-flores.de is read by an actual human within 24h.