Many teams want to use ChatGPT for work they already do in Salesforce: summarising a case, drafting an email after a call, making a first classification of an incoming request. The language model itself is rarely the hard part. The connection is. Anyone who wants to call a language model from a Flow soon thinks of Apex, named credentials and a developer who spends a week on it. It does not have to be that way.

Why a Flow is the logical place

In Salesforce, a Flow already decides when something happens and with which data. A record is created, a user works through a screen, a scheduled job runs overnight. If the AI step lives in the same Flow, you do not need a separate system that knows when to start. The result also comes back as an ordinary variable that you can show, store or pass on.

This applies to all three Flow types: Screen Flows for manual work, Record-Triggered Flows for automation on changes, and Auto-Launched Flows for background processes.

Building it in four steps

  1. Decide what the Flow should produce. A concrete outcome, such as “a follow-up email of at most 120 words”, works better than “something with AI”.
  2. Retrieve the data you need. A Get Records element or a record variable supplies the fields the model needs, for example account name, industry and the latest note.
  3. Build a prompt template. Fixed instruction text, combined with merge fields from your Flow. The instruction stays the same, the data differs per record.
  4. Use the result. The response is a variable in your Flow. Show it on a screen, write it to a field or send it in an email.

Where it can go wrong

A Flow that calls a language model behaves differently from a Flow with only formulas. Keep a few points in mind:

  • Give the model only the data it needs. More context does not always improve an answer, and you also send more data to an external party.
  • Keep a person in the loop for outgoing text. In a Screen Flow you can show the result first, so a colleague can adjust it before it goes out.
  • Treat outside input as untrusted. Text from a web form or email may contain instructions that try to override your prompt. Use such text as data, not as an instruction, and do not let the Flow take actions on its own based on the output.
  • Choose your model deliberately. The model you pick determines which privacy and data retention policy applies. Settle that before customer data passes through it.

How we solve this with the AI Prompt Builder

Our AI Prompt Builder is a Managed Package that makes a language model available as an invocable action in your Flow. Admins build prompt templates with merge fields, without Apex. You can choose between OpenAI, Anthropic, Google Gemini and Azure Co-Pilot, and you can connect your own LLM as well.

We are specific about cost: a fixed monthly fee of €150 including 1,500 callouts, then from €0.10 per call. If you connect your own language model, there is no per-callout charge. Resolve IT handles installation, configuration and training for your team.

Want to know which of your processes suit an AI step in a Flow? Get in touch and we will look at a first use case that delivers value quickly.