There are two completely different jobs people mean by “AI for Salesforce”, and conflating them is what makes the pricing conversation so confusing.
Job one: AI that faces your customers
A service agent that resolves cases end to end. A sales agent that qualifies inbound leads overnight. An autonomous workflow grounded in Data Cloud that acts on records without a human in the loop.
This is what Agentforce is built for, and it is genuinely hard to build anywhere else — it needs deep native access to Salesforce data, sharing rules, and the platform’s own trust layer. If this is your requirement, buy Agentforce. Nothing external substitutes for it.
Job two: AI that helps the people who run the org
The admin trying to work out what fires when an Opportunity closes. The developer writing a trigger. The analyst who needs a dashboard before a 2pm call. Nobody outside the business ever sees this work.
This is where the pricing mismatch bites. Salesforce’s AI is licensed per user at enterprise rates — the Agentforce add-on and the per-user editions are priced for customer-facing deployment economics. Applying that to an internal team of five admins means paying customer-scale prices for what is, functionally, developer tooling.
What bringing your own key changes
The alternative is a tool that connects to your org through a standard Connected App and calls a model API you already pay for — Anthropic, OpenAI, OpenRouter, or a self-hosted OpenAI-compatible endpoint.
Three things follow from that:
- You pay the provider directly, at cost. No per-seat AI licence and no consumption credits bought in advance at a markup. You also see exactly what the usage costs, because it is on your own bill.
- No Salesforce AI licences required. Any edition with API access works. No Einstein, no Agentforce, no Data Cloud.
- You choose the model, per task. In orgadmin.ai six AI surfaces — chat, the SOQL assistant, Insights, the Development Studio, Tech Debt Finder, and Security Review — are each configured independently, so an expensive reasoning model can handle the security analysis while something cheap and fast handles query generation.
The trade-offs, plainly
Bring-your-own-key is not free, it is unbundled. A failed run still consumes tokens on your key, so set a spending limit. Your prompts go to the provider you chose under their API terms rather than staying inside the Salesforce trust boundary — which is a real consideration for some regulated orgs, and a non-issue for most.
And an external tool cannot build customer-facing agents. That genuinely is Agentforce territory.
How to decide
Ask who the AI is ultimately talking to.
If the answer is a customer, you are buying platform AI, and the per-user licensing is what it is. If the answer is your own admin, developer, or analyst, you are buying developer tooling — and it is worth asking whether it should be priced like a CRM licence.
Plenty of teams end up running both, because they do not overlap. Building Agentforce agents still means understanding your org, writing SOQL, and deploying metadata — and that work is exactly job two.