So, what should we take from his statements? After all, the words of the CEO of one of the top 3 valuable companies on this Earth carry some weight.
Let me start straight.
I call BS!
SaaS, first of all, is a delivery model of logic that also had some implications on vendors‘ business models and their approaches to pricing. For a variety of good and not so good reasons this delivery model succeeded vs. the prevalent model of on-premises software. Some of the more important reasons have been “no lock in by vendors”, “only pay for what you use”, “reduction of own infrastructure cost”. Of course, there are more. All of them being true – or not so much. One thing is for sure, SaaS led to a considerable centralization of compute resources. Hyperscalers emerged. Vendors took over the management of the application stack for their clients. It is very hard to envision that this gets reverted any time soon, even in a world with increasing trust issues and a good argument for edge computing.
What SaaS is not, or only marginally, is a way of creating business logic. Of course, software architecture changed to support multi tenancy vs. single tenancy, but this is rather an internal reason to make this delivery model more efficient for vendors. Clients do not care about single- or multi-tenancy, as long as their requirements are met. Just to be clear, multi tenancy, and with it, SaaS, first of all is a means to make software delivery and maintenance more efficient for the vendors. The benefits of this efficiency gain may or may not be forwarded to customers.
What Nadella suggests is a totally different way to build application logic. Instead of formulating it by coding and/or rule systems that automate a process, he envisions a network of specialized artificial agents that collaborate flexibly to achieve the desired outcome. Sounds cool and modern, doesn’t it?
Just that it ain’t work.
For several reasons.
Chief of them are accuracy, reliability and predictability. While coding and rule-based systems deliver predictable and repeatable – i.e., auditable – outcomes, the current breed of artificial agents does not. Instead, they are probabilistic and very likely deliver good results, which is often not good enough. And this is not fully mitigated by technologies like RAG. And in the vast majority of situations, businesses and their managers prefer, even need predictability. How well agents based on different foundation models fare in “consequential real-world tasks” is impressively evaluated in a research study titled “The AgentCompany”. The researchers found that even the best model (Claude 3.5 Sonnet) manages to complete only 24 per cent of the total of 175 tasks. I’d say that this is quite sobering. Research conducted by the Australian National Science Agency on the usefulness of Microsoft’s Copilot vs. expectations shows mixed results, too. Now, one can say that it doesn’t really matter whether a human employee or an AI agent makes the mistake as the result is the same. I, myself, use this argument a lot. But in this situation, we are talking about trading a predictable outcome for a probabilistic one.
Then, there is cost. Running large language models, let alone train them, is incredibly costly. Why would a CFO or his/her CEO in their sane mind replace their expensive systems with even more expensive systems? These systems being AI based cannot be the justification, cool or not. OpenAI is losing $5 bn in 2024 on revenues of $3.7 bn. Microsoft is planning to invest $80 bn into AI-ready data centers in its fiscal year 2025 alone. It would be naïve to assume that these are done out of charity and that there is no intention to create revenues out of these investments. So, the only reasons for doing so, are the agent-based system delivering far superior capabilities or saving even more money somewhere else.
But Nadella is right, too!
SaaS does not live up to many of its promises. We still do see considerable lock-in, we see pricing that is geared towards vendors and, in my opinion, worst of all, we see considerable “siloisation” of business logic, which led to disjointed data and, in consequence, broken processes. In addition, codified business logic has a high cost of change. Additionally, SaaS software traditionally offers less scope for customization than on premises software did.
In this sense, SaaS, or business software in general, needs an overhaul, although one can argue that few of these problems are SaaS specific and that it is in no way sure that agentic systems will resolve them.
Agentic systems require a data harmonization to work. Whether they access a multitude of different databases or just one, directly or via an abstraction layer doesn’t matter. The bottom line is that there is a considerable drive towards quality data.
Architecting systems based on autonomous agents that interact with each other has the potential of making applications truly smart. Theoretically, they can adapt their behavior automatically based upon changing customer/user requirements. Easier adaptation of systems is the holy grail for some time. Microservices and composable architectures are good examples for this force.
The necessary technologies are there and “only” need to be combined.
Staying close to architecture, agentic systems are platform systems, i.e., they strongly prefer a single (family of) vendor(s). For connected agents to work properly, they need a common language that all of them understand – aka an interaction protocol. This is entirely possible, but mandatory to avoid rising integration cost.
Pricing is different, non-technical, topic. Yours truly and other people have written about this topic. Here a good one by Phil Wainewright. Agentic software offers the possibility to achieve outcome orientated pricing and to align vendors’ and their customers’ interests. However, as Phil writes succinctly “Throughout the history of Saas, people have been talking about outcomes-based pricing, but while it sounds attractive on paper, it's really hard to deliver in practice.” Still, pricing models will change as there is no coupling between agents and users anymore in agentic systems.
Let me add another point. The interaction with software is traditionally built not around how humans interact but around the limitations of how computers – machines – can interact. I do not say that GUIs and point-and-click are all bad, but still, they are a band aid to enable the effective and efficient interaction with software. With the advent of NLP, NLU, NLG and LLMs, these limitations do not really exist anymore. Instead, humans can now interact with computers in human ways and are no more confined to the application window. This is game changing.
So, what now? BS or not?
In summary, one needs to read Nadella through a Microsoft lens. Agentic systems offer the possibility of drastically simplifying the (business) software stack by migrating a vast variety of different tasks into different types of software agents that work in orchestration. Doing so, is certainly changing the SaaS world into something that looks very different from now. Microsoft as a technology vendor supports nearly all of these tasks to an extent. Apart from his click-bait type of terminology he has a point or two.
But is SaaS dead? By no means.
Can a world as Nadella sketches it evolve? Sure as!
It only needs three ingredients to make this new world of SaaS happen:
· A bold vendor that offers considerable business functionalities via working agentic software
· At attractive pricing
· That convinces enterprising businesses that it is well worth the plunge
Exciting times ahead!
Comments
Post a Comment