SaaS vs. AI or SaaS + AI?

Is the future SaaS vs. AI, or is it SaaS + AI?? Will AI "entirely replace the business logic layer" in a SaaS application, as Satya Nadella predicts?? Or is Marc Benioff's view of SaaS and AI cooperating together a more accurate model?

In particular, highly regulated "systems of record" in areas like finance, HR, and healthcare would be very difficult to replace with AI (for at least the immediate future).? And perhaps the most helpful way to think about it is that SaaS applications will become "tools" for AI agents (as described in a Google whitepaper, more below).? Let's discuss…

Satya Nadella (CEO of 微软 ) offered his view on the BG2 pod (with Bill Gurley & Brad Gerstner ) in Dec 2024, saying that SaaS applications are nothing but business logic built on top of CRUD databases (Create, Read, Update, Delete), and all the business logic will soon be in the AI agents, and they'll update the database directly.?

Keep in mind that Satya has taken Microsoft from $1T to $4T in market value in the last 10 years, and they have a large investment in OpenAI (and the source code, by the way, just in case OpenAI goes under).? He's obviously a smart guy, he is well aware of what the technology is capable of, has an insider view of the future capabilities, and his perspective is that AI will replace SaaS.

And if that weren't enough, on the All In podcast 2025 predictions episode, they nominated enterprise applications and the software industrial complex to be the "worst performing asset" of 2025!

Deep breath… here's my hot take (no disrespect to these industry giants)… the problem that Satya and the All-In Besties are overlooking is the complexity of the combination of the business logic and the data in a "system of record" (SoR), and the difference between "probabilistic" (AI) and "deterministic" (static code/apps) computing.?

Domains like finance, human resources, and healthcare require testable, repeatable known rules and outcomes when it comes to creating, reading, updating, and deleting data.? There are also laws and regulations in different states and countries that all need to be followed… and audits that are required… and it needs to work every time, very quickly, 24x7.? This is a 100% deterministic problem.? AI excels at logic, reason, research, planning, and all sorts of highly complex things, but it doesn't do well at fast, precise transactions (at least not yet).

So where I think this is actually going is that SaaS products will become primarily a "system of record" with an API, and behind that API you'll have all the highly deterministic business rules and data.? And then AI will call the API to do things like create, read, update, and delete data, but AI won't be allowed to do it straight to the core database (sorry, Satya).? In other words, the SaaS product will become a "tool" for an AI agent to utilize, and SaaS will still retain its value as a product, albeit in a slightly different form in this new ecosystem.

This idea of agents and tools is described in a 谷歌 whitepaper called Agents published in Sept 2024 by authors Julia Wiesinger , Patrick Marlow , and Vladimir Vuskovic . Their distinction of these different macro components in future software architecture are very helpful as we sort out how everything will work together.

Let's consider a very simple example of an AI agent ordering a ride for you to the airport.? You expect that your AI agent will contact Uber, Lyft, or others, compare prices and evaluate options, then book your ride via the API to the chosen product.? You don't expect it to determine which stranger might be willing to drive you to the airport, call them, schedule a time, and negotiate a price.? You expect it to use Uber (or Lyft) as a tool to get the job done.

SaaS applications are no different than Uber in that example.? (BTW, I hate to ruin it, but Uber is technically a B2C SaaS application.)? Let's look at the most wildly successful SaaS business application, Salesforce…

Marc Benioff , CEO of Salesforce , has been on multiple podcasts in the last 6 or so months touting the new Salesforce AI agent called Agentforce.? His most recent conversation with Peter Diamondis on the Moonshots pod is a great example.? Salesforce has already used AI agents to dramatically reduce their own customer service interactions escalated to people, and they're helping their customers do the same.? The Salesforce AI agents have access to your company's Salesforce data, and those agents can call the Salesforce API or access your Salesforce data on your behalf, if you just briefly describe what you're trying to do.? His version of the future is SaaS + AI, and I think he's spot on.

There still might be traditional websites and apps, but those will start to be used less and less as AI and voice become the interface to everything, with the visual screen only used as an optional way to verify what you just said (before it's executed) or to view the results of what the AI accomplished.? As a SaaS vendor, you must be in the cloud (co-located with the AI hardware), have a great API, and be highly scalable to thrive in this new AI world.? Those that do will preserve their value as a "system of record" that is a useful tool for AI.?

I'm excited about the future of AI, and we all need to be learning more and pushing the boundaries of what's possible, while we responsibly manage the transition to this new era of "ambient intelligence and ubiquitous computing", as Satya put it so well 10 years ago when he took over as CEO of Microsoft -- even though, respectfully, I think he's a little early on predicting the replacement of SaaS with AI.