The spectre of SaaSpocalypse has continued to loom over the enterprise and business software markets. Each time a new company announces a cancelled contract, or a frontier AI provider announces a business-adjacent feature, stock prices of SaaS providers catch a cold. Even news of billion-dollar deals can’t reverse the fortunes of those stocks these days.
But companies don’t necessarily live and die on their stock prices. And prices are, after all, only partly based on data. The other part is a combination of educated guesses, emotions, and vibes. But as an onlooker and observer, I’ve seen a number of factors coalesce with Salesforce’s strategy and the broader marketplace, making it increasingly likely that Salesforce will find success in what has become the AI-enabled SaaS marketplace.
The Enterprise Difference
Early on in the AI boom, it was clear to me that business software would be the first to the table to successfully monetize AI. Anyone working in enterprise software for some time would have made the same observation. While the bust of the dot-com collapse hurt everyone and decimated retail companies, enterprise software companies were the last to be impacted and the earliest to rebuild from the ashes.
There is also the common sense that there is way more purchasing power in businesses than in individuals. Go back through history and look at the largest software companies by their valuation. The list of the most valuable software companies has always skewed heavily towards those who sold to businesses. So while ChatGPT woke consumers up to generative AI, triggering the current AI boom, more business-focused approaches always looked more likely to generate the kind of revenue you need to justify the spend on AI.
Distribution Over Model Ownership
It’s a well-worn premise that the first product doesn’t always win. In technology, we’ve seen this over and over again. Yet, OpenAI jumping first created an accepted wisdom that their head start and model quality had built them an insurmountable moat. This cemented a belief that models would be the engine for economic success in the AI marketplace.
On the other hand, in the early years of the boom, Apple and Google were pilloried for being caught flat-footed and getting left behind on their own AI strategy. Both also struggled with their initial forays into models (Bard) and incorporating the latest AI functionality (Siri). But what they both have is huge user bases.
More recently, a different story has developed. All Google needed to do was insert AI into www.google.com, and instantly they had a huge user base, normalizing use of their AI agent to nearly every web user. In conjunction with that, there are paid add-on features in G Suite and other tools.
Apple similarly has a very loyal user base. Whether you’re a serial Mac or iPhone user, Apple Intelligence is now here. Apple’s plans to drive both hardware and software services are now fully in play.
Last week’s earnings calls confirm the progress, with both posting strong revenue numbers. Google’s Gemini app reported passing 1 billion monthly active users, and Apple’s early feedback on their AI services has been positive.
Now look at enterprise software. Where is the distribution? Salesforce. Granted, we don’t know what “over 150,000 customers actually means. But there’s no denying that they’ve penetrated deep into the business market and (until the AI boom) have been considered the market leader for some time. Salesforce’s own statistics routinely cite that 90% of Fortune 500 are their customers.
Realizing the promise of their distribution will be different for Salesforce, but it’s a similar approach to Google. There is a small amount of AI that’s given away to all customers through the Foundations feature set. And as Salesforce continues to make tweaks to their pricing model, it seems likely they’ll be able to convert and win more customers for their AI suite of tools.
There’s still work to be done. There is still no clear killer use case that every business wants or needs. But this legion of existing customers is a strong foundation to springboard off of and grow a new AI-enabled SaaS business.
Pricing a Model for Fun and for Profit
AI feature pricing has been a thorn in the side for Salesforce since the first launch of what eventually became Agentforce. We’ve used up a fair amount of ink on this here at SF Ben. So let’s not rehash this history. Rather, let’s look at where things are, and where they may go.
Pay-per-resolution pricing has become the latest innovation among AI app providers. MeshMesh, whose team was recently hired away by Salesforce, was an in-ecosystem example of this. Salesforce may have also been looking over the shoulder of former Co-CEO Bret Taylor’s Sierra AI, who use the same pricing model.
There have been some stirrings about this that Salesforce gets to state the terms of “resolution”. That may be fair criticism, but it’s actually (uncharacteristically for modern AI) quite deterministic. This means buyers should be able to evaluate the terms of the meaning of “resolution” and decide whether they want to buy it or not. I think this is something that the market will decide on, and Salesforce will adjust if needed. And if they can roll out pay-per-outcome to more AI use cases, it seems a winning formula.
But what about model cost? For some time, it’s been an open secret that frontier AI providers are selling inference at a loss, especially for their most capable models. This has been subsidized by large funding rounds that they are now burning through. This can’t last indefinitely. And there are huge questions about whether the current economics can bring model pricing down.
Salesforce relies directly on both OpenAI and Anthropic models. We don’t know the pricing arrangement they have, but it seems improbable that Salesforce is paying the full inference cost. And recent moves to move away from loss-incurring flat-fee subscriptions to pay-per-usage models have been met with discontent, to say the least.
We don’t know how this is going to shake out, but it is highly unlikely that Salesforce isn’t thinking ahead to come up with a strategy that will allow them to incorporate state-of-the-art AI at a cost that will allow them to sell at a reasonable margin.
Small Models at Large
In 2016, Salesforce acquired an AI research company, Metamind. From this team, led by Richard Socher, Salesforce established an AI research organization that has consistently delivered the goods. Socher left in 2020, and AI research is now led by Silvio Savarese, although there are other notable members on this team, such as Shelby Heinecke, who often speaks publicly about Salesforce’s AI research work.
In the past few months, there has been more and more talk of specialized “small” LLMs. But do you know who was talking about this first? Salesforce. In fact, the architecture of the Atlas Reasoning Engine consists of a number of smaller specialized models that serve different roles in the reasoning loop. This creates huge efficiency.
Where some agents use the same umpteen trillion parameter model for every point of the agent loop, Atlas is using several smaller specialized models. Popular agent frameworks like LangChain/LangGraph and APIs like OpenAI are now doing similarly using the director/executor pattern. But Salesforce has used this structure since at least early 2025.
If this level of geekery is beyond where you ever want to go with AI agents, the long and the short of it is that Salesforce has the AI chops to build capable smaller models, with optimal inference costs. This could give them early access to profitable AI revenue (if they’re not profitable already).
The Army That Will Implement It
Salesforce’s recent relationship with its community has been dicey. Lack of care taken to attend to them in the way they were accustomed since 2006, missteps with product roll-outs, and the apparent turning of the back on values which attracted many to the “Ohana” have eroded what was once an enthusiastically warm relationship.
But there is still a vast number of technical professionals who have built their careers around Salesforce. Disappointment with the changes in the tenor of Salesforce’s relationship with the community has not eroded loyalty to the technology. This army of implementers is slowly learning and adopting Agentforce as a tool to solve customer problems. SF Ben’s most recent surveys (admin and architect) have shown that the number of people working on Agentforce projects is slowly edging up, pricing issues are being attended to, and Agentforce revenue has risen above the $1.3B mark as of the last earnings announcement.
If you’re not one of the ~30% currently working on Agentforce, that means you’re available to skill up and hop onto a project. Salesforce as an agent platform may seem weird to many who are implementing agents with CrewAI, LangGraph, or LlamaIndex. But you know who it won’t feel as weird to? Yep. A Salesforce Admin, Developer, or Architect. And when the deals start closing that need someone to make an agent to handle pre-lead qualification, there will be someone there ready to skill up and take it on.
In the past, Salesforce has made savvy investments in upskilling their army of practitioners. This work has been going on virtually, with several series looking to upskill remotely. But is that enough? In 2016, adoption of Lightning was languishing. A large part of the friction is that people were too used to building with what they knew: the “classic” UI with page layouts and Visualforce.
The Lightning Now Tour, a series of enablement events run locally, completely turned that around. If Salesforce were to make a relatively small investment to bring enablement to people where they work could accelerate and onboard Salesforce professionals to be truly competent with building agents.
The Vibe Code Threat
Hand wringing about the end of SaaS has been habitual this past year. Even this past month, there has been news of more people “vibe coding a Salesforce.” Can we just admit this isn’t new? Build versus buy has been the age-old question in the world of CRM. I can tell you about a certain project team from a major US bank in the early 2000s I had to train. They had built their own CRM, but now were about to implement one from the company I worked for at the time. Whether to build your own or implement someone else’s has been a question for as long as enterprise software has been sold to companies.
Is it easier to build it now? Maybe. But the challenge isn’t in building the features; it’s all the other bits of scalability, security, and infrastructure that Salesforce provides that companies will miss. Still, some will take this on and succeed. So will there be some additional attrition from traditional SaaS vendors? You bet there will. Will the overall SaaS market shrink? Maybe. Can Salesforce weather this attrition? I’m certain of it. In the end, Salesforce won’t want to, but it can lose a few customers and come out just fine.
We still don’t quite know what the killer AI use case is. But someone will find it. And when it arrives, Salesforce can be counted on to fast follow and bring that to their customers, wrapped up, easy to implement, with all the infrastructure support they’ve always provided.
How Headless Plays Into the Story
If you want to see a great example of hedging your bets, just watch Salesforce these days. Headless 360 is not new. Rather, it is the realization of the API-first platform Salesforce always promised, but perennially fell short of. The difference now is Salesforce can’t afford to do that anymore.
Agents are amazing at working around things, but they are most reliable and consistent when given consistent factors to deal with. Sure, a developer might have been able to hack their way around that little missing part of the metadata API, or that standard object that didn’t quite work like other standard objects. But we don’t want agents to do that.
In its marketing, Salesforce sells Headless 360 as the way to get Salesforce to every agentic surface. In other words, whether you’re interacting with an agent on Slack, Salesforce, Claude Cowork, or anywhere else, when you summon Salesforce, it will arrive like a genie out of a lamp with that little bit of data, or that little function that empowers you to get that work done and not break your flow.
But for every other Salesforce customer – not necessarily beer-bonging Agentforce – just like always, you’ll benefit, too. Headless 360 should mean fewer inconsistencies, and more regular old APIs. Yes, Salesforce is obliging the agentically oriented by making them all available as MCP endpoints or subagents. But that doesn’t happen without the underlying working APIs.
Final Thoughts: A Vision for Salesforce
The Salesforce of the future will look different from its first 27 years. What was once a simple sales automation app grew into a platform that supports vast amounts of the world’s business. But the innovation that Salesforce has always pushed, even if obscured by over-the-skis marketing motions, will persist. In that way, this is the same old Salesforce. This does not mean an easy path, nor is it a guarantee, but there is no denying that what has looked disorganized, or even chaotic on the part of Salesforce for the past three years, is beginning to show promise for the way forward.
How do you expect Salesforce to cope with the coming transformations brought about by AI agents? Is there some challenge that you think may make a difference to the above? Make sure to share in the comments below or on your social media feed.
