Salesforce Discovers Middleware

Marc Benioff posted something on Twitter about on-device agentic AI and small language models, which is a sentence that requires some unpacking if you know anything about what Salesforce actually is.

Salesforce is a cloud operation. Your data — all of it, the contacts, the pipeline, the emails, the deals, the notes your sales reps wrote at 11pm after their third glass of wine — lives on Salesforce's servers. This is the entire product. The whole value proposition for thirty years has been: give us your data and we will make it useful to you. So when Benioff announces that the future is on-device AI processing your data locally, you have to sit with that for a second.

The argument for running a small model on your laptop instead of a massive model in Salesforce's cloud is roughly: privacy. Your data never leaves the device. Which is a real and legitimate concern — except that your data already left the device. It's in the Salesforce cloud. Where it has been since 1999.

So who is this for.

Then I read the blog post, and there's a tell buried in there — they spend a significant amount of copy trashing large models for being energy-hungry and inefficient. Which is true. But it's the kind of truth you reach for when you're building a case that doesn't quite hold together on its merits, the kind of thing you say when the real answer is "we do not have the GPU capacity to compete at scale so here is why scale was bad all along."

Then I read the actual paper.

Benioff said it was a breakthrough in agentic AI. What xLAM actually is: a model fine-tuned to dynamically generate REST API calls. It reads an API schema and writes the request. That's the thing. A Python function that constructs an endpoint query, but the function was trained on a bunch of API documentation instead of written by a contractor in Bangalore.

The benchmark numbers are impressive if you compare a model specifically trained to call APIs against a general-purpose model that can also write poetry and explain the French Revolution. Of course the specialist wins. You could fine-tune a 1B model to beat GPT-4 at addition if you tried hard enough. The benchmark would be correct and the conclusion would be meaningless.

What Salesforce is actually building here — and this part is at least coherent — is an AI replacement for Mulesoft. Middleware. The connective tissue between enterprise systems, the part of your software stack that speaks fluent Salesforce to your SAP and fluent SAP to your legacy Oracle database. xLAM wants to be the layer that figures out which endpoint to call and how to call it, dynamically, without a human writing the integration.

That's a real problem worth solving. It's just not "a revolution in AI" — it's "we want AI to do the thing our other product does, but cheaper."

The honest version of this announcement would have been: we're researching whether a small fine-tuned model can handle API orchestration well enough to replace hand-coded integrations, and early results are promising in a narrow domain. That's interesting. That's worth a blog post.

Instead Benioff called it agentic AI for the enterprise and implied it runs on your phone and changes everything, and now here we are.

Nice that they're trying, though. Genuinely.