AI Agent Action in workflows: Token Usage Optimization

The AI Agent action now uses significantly fewer tokens per execution, with no impact on output quality. Since usage is billed per token, this directly translates to lower costs and more room for scaling.

Cleaner Context — Duplicate and internal data that was previously sent to the model on every turn has been removed. Raw database internals, redundant contact details, and unnecessary workflow metadata are no longer included.

Smarter Tool Responses — Tool outputs used to contain dozens of irrelevant fields per record (for example, 30+ permission fields per user). Now, only pertinent fields such as name, email, and phone are passed to the model.

Conversation Memory Management — Long-running agents now automatically summarise older conversation steps while retaining recent ones in full detail, rather than sending the entire history on every turn.

Structured Output Optimisation — The final extraction step no longer duplicates the entire context, saving thousands of tokens per execution.

Tested on the same workflow and contact with zero loss in quality:

• First LLM call: 36% token reduction
• Total execution: 20% token reduction

Every workflow execution is now more efficient. The output quality remains the same, yet fewer tokens are used, leading to lower costs — meaning you can run more automations without increasing expenditure.