Salesforce cut 4,000 support roles using AI agents. Then admitted the AI had reliability problems significant enough to warrant a strategic pivot.

I have said this multiple times and received a lot of pushback. But this Salesforce story makes it clearer than anything I could write.

You cannot deploy AI in production workflows without infrastructure governing how it executes. Salesforce just figured that out. The hard way.

They deployed Agentforce across their own help site, handling over 1.5 million customer conversations. Cut 4,000 support roles in the process. Then their SVP of Product Marketing said: "All of us were more confident about large language models a year ago."

One customer found satisfaction surveys were randomly not being sent despite clear instructions. The fix was deterministic triggers. Another name for what should have been enforced from the start.

Human agents had to step in to correct AI-generated responses. That is the babysitting problem. The same one developers describe when they say half their time goes into debugging the agent's reasoning instead of the output.

They could have added LLM-as-judge. A verification protocol. Some other mitigation. But all of that is post-hoc. It satisfies the engineering checklist. It does not satisfy the user who already got a wrong answer and moved on. A frustrated customer does not give you a second chance to get it right.

They have now added Agent Script, a rule-based scripting layer that forces step-by-step logic so the AI behaves predictably. Their product head wrote publicly about AI drift, when agents lose focus on their primary objectives as context accumulates. Stock is down 34% from peak.

The model was not the problem. Agentforce runs on capable LLMs. What failed was the system around them. No enforcement before steps executed. No constraint persistence across turns. No verification that instructions were actually followed before the next action ran.

They are now building what should have been there before the 4,000 roles were cut. Deterministic logic for business-critical processes, LLMs for the conversational layer.

That is not a new architecture. That is the enforcement layer. Arrived at the hard way.