Salesforce announced that the whole platform is going headless. The company is exposing every capability as an API, MCP tool, or CLI command, so AI agents can operate the entire system without ever opening a browser. The framing is that work no longer happens through a graphical interface.
It's a big bet. I'm not saying it's the wrong one. But are they actually done with the AR follow-ups, the inbound triage, the back-office plumbing? Because the headless CRM doesn't fix any of that. It just gives agents a faster way to touch it.
Here's what I think most people are getting wrong about AI right now. They're treating it as one thing. It's three different bets with three different risk profiles, three different timelines, and three different ways to be wrong. The order you take them in matters more than the bets themselves.
AI on the business
Think unsexy thoughts. Operations, back office, finance, support: the functions that run the company day to day. Take a workflow, break it into atomic pieces, chain them back together.
This is the bread and butter of what we do at Aisle, and I'd argue it should be table stakes for any company of meaningful size by the end of this year. If you can't describe how your AR follow-up, your inbound triage, your weekly board prep, or your content production runs as a set of composable pieces, you're behind. Not behind on AI. Behind on running a company.
The reason this is the first bet is because it's the cheapest to be wrong about. The unit of work is small and the blast radius is contained. You can ship a prompt that handles one piece of one workflow, see if it cuts the cycle time you expected, and either expand it or throw it away. Nobody outside the company ever sees the failure.
It's also the bet with the most obvious ROI. You get hours back, cycle time drops, quality improves, and the math is measurable within a quarter. You don't need a strategy deck to justify it.
Salesforce's version of this bet is the boring half of the announcement. The internal agent work that doesn't make a keynote: deal-desk prep, pipeline hygiene, the unglamorous loop of a company running itself. That's #1 work, and it's where the highest-ROI compounding actually happens.
AI in the product
You take assets you already have, usually data or knowledge sitting in the business doing nothing, and make them commercially useful through AI. Complex data operations that once took an army are now economically viable for most teams. If Data is oil (or the new moat), then AI capabilities are like fracking.
Or you take an existing aspect of your product and make it better. Better content surfacing. Better automation. Faster, more accurate, higher quality. The customer doesn't know they're interacting with AI. They just have something that they didn't have before. They notice that the search returns the right answer on the first try, the onboarding feels less stupid. Everything feels snappier, and more intuitive.
This is the bet most companies are underinvesting in relative to its return. It's harder than #1 because it touches the product, which means it touches customers, which means the failure modes are public. But the upside is real. You're not just saving cost. You're creating new commercial surface area from things you already own.
The risk profile is medium. You can ship it iteratively, A/B test it, pull it back. But it requires real product judgment, and not all teams have it.
Salesforce's version of this is the part of the announcement that's incredibly defensible: twenty-plus years of CRM data, support tickets, and configuration knowledge that competitors can't replicate, now feeding the AI layer underneath every existing workflow. That's #2. That's the moat. Most of the value of the headless announcement is downstream of that data being usable, not the headless part itself.
AI as the product
This is the loud one. The agentic, "the chatbot is the interface" version. Your customers explicitly know they're using AI. They can do things they couldn't do before. The product is the AI.
This is also the most volatile bet on the board.
The world has not consolidated. Salesforce's Agentforce Vibes 2.0 supports both the Anthropic agent SDK and the OpenAI agents SDK. Anthropic has Skills, there's MCP, and three months ago none of those existed as serious options. Every framework that looks like the obvious winner today has a real chance of being a footnote by next summer. If you're building your entire product surface on an interface standard that doesn't exist yet, you're going to throw a lot of work away.
Which doesn't mean don't do it. It means invest accordingly. Keep the bets small enough that you can rip them out when the framework shifts. Build the things that would be worth doing even if you had to replace the underlying framework entirely.
Don't bet the company on a posture.
The right #3 bet is one where the AI capability is the product, but the underlying business would still exist if you stripped the agent layer out. A legal research tool where the agent is the interface but the value is twenty years of case annotations. An internal analytics product where the agent answers questions on top of a data warehouse you already trusted.
The Salesforce announcement is interesting precisely because it's a #3 bet at a scale almost nobody else can afford to make. They made the decision two and a half years ago to rebuild Salesforce for agents. That's a long runway and a lot of conviction, and credit to them for committing early. But the bet only works because of the #1 and #2 work happening underneath. Strip out the data layer and the internal automation, and the headless platform is just a faster way to make the same mistakes.
That's the part I'd flag for anyone reading this and thinking they should be doing the same thing. If your back office is still held together with quarterly heroics, going AI native isn't going to save you. It's going to expose you.
Pick in order
Start with #1. The factory has to run before anything else compounds. Make composability the default in how your team thinks about work.
Move to #2 next. Look at what you already have and ask where AI makes the existing experience or the existing data meaningfully more valuable to the customer.
Take #3 selectively. Real bets, small ones, with the assumption that the ground will shift under you at least twice before it settles.
The mistake I see most often isn't picking the wrong bet. It's picking #3 first because it's the one that gets you on a panel. Picking #3 first gets you a keynote slot. Doing #1 and #2 first is what compounds.