PARAS MADAN
All writing

Latest in AI

Make AI Scenarios: Replace Retired Gemini Models and Reusable Prompts

Audit Make scenarios for retired Vertex Gemini models and OpenAI reusable prompt IDs, then test mappings before switching live workflows.

Two Make automations can fail for entirely different reasons even when both modules still look familiar in the scenario editor. One may be calling a model that no longer exists. The other may be pointing to an external prompt that the integration will stop supporting.

Make's November 20 update covers both: retired Gemini 2.5 models in its Google Vertex AI app, and the OpenAI Reusable prompt option in Generate a response. I would handle them as two migration queues and use the same review process before returning either to a live workflow.

Find the actual dependencies in each scenario

Start with active scenarios and the templates your team copies most often. For each affected module, record the scenario owner, last successful run, model or prompt reference, and the consumer of its output.

This is more useful than searching scenario titles. A workflow called “Triage inbound tickets” might contain an AI step on only one branch, and its owner may not remember which integration supplies it. Open the module and record the app and field that actually contain the dependency.

Separate already-failing work from the migration deadline. A retired model requires a supported replacement now. A prompt dependency that still runs deserves a tested replacement before Make's stated cutoff.

Replace the model in Google Vertex AI modules

Make says scenarios using gemini-2.5-pro, gemini-2.5-flash or gemini-2.5-flash-lite in its Google Vertex AI app now fail. Open the affected modules and check Model. Make cites gemini-3.5-flash as one possible replacement. The Google Cloud model lifecycle table can help you check retirement dates and recommended options.

A supported model is not necessarily a drop-in substitute. If a ticket router expects a specific category or a strict object, run the same representative inputs and inspect the fields the next module receives. Include a short message, a long one and a case that should go to human review. For modules that accept files or other media, test one of those inputs too.

Check your replacement in the real target project and location before go-live. A healthy test in a different project does not prove the production connection can use it.

Move OpenAI reusable prompt references into explicit inputs

According to Make's notice, migrate the deprecated Reusable prompt selection before November 30, 2026. This advice is for Make's Generate a response module, not a claim about every OpenAI prompt or endpoint.

Make recommends copying the prompt text into Text prompt or Advanced input. For Advanced input, add a Message with Text content under Input. The module reference lists the surrounding request fields, including Instructions and conversation options.

Before replacing the ID, write down which prompt version the scenario is meant to use and what inputs it fills in. Rebuild the mappings deliberately. Copying a template into a text box does not prove the ticket subject, customer status, or other dynamic values still reach the model. Inspect the assembled input on a synthetic run and check for literal placeholders and unintended blanks.

Keep one agreed version of the prompt if several scenarios used the same external definition. Once they hold copies, editing one Make module won't update the others.

Test the output contract before resuming the schedule

Make documents how to clone a scenario. Use a test copy with sample inputs and safe destinations. Don't let a copied support flow send replies or write customer records while you're inspecting its output.

If a scenario needs both changes, test one at a time. Compare the successful path and the fallback path, then watch the next eligible run after switching production. An AI module returning text is only one check; the downstream parser, routing conditions and destination modules must receive what they expect.

Keep the last known working prompt and mappings for reference. A rollback that restores a retired model or a deprecated prompt reference is not a recovery plan. Decide in advance where affected work should go for manual review while you correct a failed migration.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help AI and automation teams explain product changes and technical workflows through clear writing and practical storytelling.

Book a 30-minute call.