Salesforce Pulls Back From Llms, Pivots Agentforce to Deterministic Automation

<worstpersonyouknowmakesareallygoodpoint.gif>

This reads like a list of obvious things that we have been saying for the past 3 years.

Presumably the 4000 sacked workers won’t be coming back though?

"While LLMs are amazing, they can't run your business by themselves... We ground AI in tight guardrails and deterministic frameworks, optimizing LLMs to deliver enterprise-grade reliability. Trusted. Reliable. Secure."

this sounds like it's copy and pasted straight from an LLM

Its unclear how they intend to fix these fundemantal problems tbh. Things like "Automate the 'must not fail' moments with rules, APIs, and triggers." And "Use policies, templates, function calling, and explicit do/don't constraints." Make sense but if you have deterministic workflows, what do the llms still add?

Using APIs makes sense but isnt the whole point of these things that they can automate away stuff, it feels like we're building really big complicated frameworks to put these things in. Does it still have any actual benefit for stuff like this?

Yeah, it's not their kingdom so it makes sense.

Eventually AI will simply be integrated in the OS and there will be no room for small players.

Recent and related...

Salesforce regrets firing 4000 experienced staff and replacing them with AI - https://news.ycombinator.com/item?id=46384781 - Dec 2025 (121 comments)

> when models were given more than eight instructions

I mean you can call a model 8 time, this seems they looking for excuses.

> If users asked unrelated questions

if you have enumerable business process why do you have chat interface, put down a set of buttons to start each business process

this seem some egregious misuse of the tech, which has it's problem, but hasn't had a fair chance here.

wonder if the whole project was just smoke and mirror to justify layoffs.

i believe this is via original at https://timesofindia.indiatimes.com/technology/tech-news/aft... (possibly an LLM redraft of the original)

You have to treat LLMs like humans under time pressure. If you give a person too many tasks in a short amount of time, they will probably also skip and/or rush some tasks. It's important to clearly identify which parts ought to be deterministic.

It's difficult to get both human-level intelligence and machine-level precision at the same time. You still need deterministic tools to do the precision work.

That said, I think LLMs definitely have a place and can provide a degree of flexibility to processes which was not possible before but it's going to be an iterative process to build those agents in a way which provides the right degree of flexibility and error-tolerance.

Then people in Saleforce will figure out that very little can be automated (Unless you want to sink 1 week of work into automating task which takes 5 minutes once a month) and then they will get back at employees again.

[delayed]

People went to the extent of letting agents to discover workflow steps dynamically. This is abuse of probabilistic logic to perform deterministic work. The hype has gone over the board, blurring the distinction.

"The message is clear: probabilistic models alone won't run mission-critical operations."

Sigh, its LLM output all the way down. All jokes aside, it seems pretty obvious right? There are things that should be more 'flexible'. This is where LLMs can shine. There are things that should be more rigid, which is where old fashioned if/then logic should work. Did execs just try to plop it all in without giving a chance to a more 'balanced' solution?

Salesforce... A company worth so much and able to layoff almost half its employees but does what...? Agentforce doesn't seem to be anything but another GPT wrapper with a lot more buzzwords?