AI is as monumental of a shift as the Cloud, because it allows us to interact with software in a fundamentally different way than before. Why is it that five years into the AI revolution we still haven’t seen AI startups subsume their “old world” predecessors and lead to new business models for software? After all, cloud-based software replaced on-prem deployments in short order because their ease of deployment and staying up-to-date, as well as the birth of the software-as-a-service (SaaS) business model, lowered the barrier for customers to invest in the latest software solutions. When is the AI Era’s “Salesforce moment” coming?
We are at the dawn of the next wave of tech-driven transformation, but we won’t see AI software replacing the incumbent software vendors like Salesforce or ServiceNow that serve as systems of record. To understand where we should instead be looking, we must be clear on the actual role of an AI model in a value chain and the key differences between these two business strategies to see why only one has the more tractable path to dominating their chosen market.
AI models output a prediction based on the given inputs and become more accurate over time if they can collect feedback on the predictions. That by itself is impressive, and we should be careful to not conflate it with anything more. That prediction is only information, and information is onIy as useful as what you do with it-- for example, the most inept sales organization won’t be saved by an AI-powered CRM that tells them which leads to call. We conflate AI and automation, but it’s important to note that AI models on their own do not take action -- the action occurs if the AI model is paired with an automation that is triggered by the prediction, or how the user decides to act on that prediction. That data generated by that action, which affirms the model’s prediction, corrects it, or ignores it altogether, and the result of that action, are the data that make the AI model exponentially more powerful over time -- if the model can access it.
Therein lies the challenge of building AI product startups: when you’re only a software vendor to your customer, you cannot control whether the outputs of the AI model are being used at all. After all, if the sales team ignores your recommendations on how to contact a lead, your lead scoring model won’t be able to generate sales lift. If your customer is not incentivized or capable of using the model’s outputs in a way that realizes its fullest potential, you’re going to struggle to deliver the ROI necessary to justify the cost of your product. AI is incredibly expensive to deploy because you need to integrate to your customers’ systems of record to access the data that will serve as inputs to the model. If you are allowed to add automation to implement the model’s outputs, that requires an additional set of integrations into your customer’s systems of control, further raising the denominator for ROI.
This is where incumbent software vendors such as Salesforce and ServiceNow have an enormous advantage over AI product startups. These vendors’ customers already pay for the platform as a system of record, so the vendor has ready access to the requisite data without needing to incur the cost of integration. These platforms have also built app ecosystems offering tools that their customers use in their workflows or as systems of control, obviating another set of integrations. These platform vendors can push out AI models as part of new feature releases. The AI product startup, on the other hand, faces a tough choice between having to build custom integrations for each customer or building their own system of record and convincing customers to rip out their existing systems-- neither paths scale. Even if the incumbents’ models are underutilized by the end customers, the lower cost of entry means the ROI multiple of any AI model released by the incumbent vendor will be much higher than one from an AI product startup.
There’s one clear place for AI product startups to succeed, and that’s in industries where systems of record haven’t seen meaningful adoption, such as construction or Food and Ag. It’s a steeper hill to climb because you have to convince the customer to invest in instrumentation to collect the data and a system of record to manage it. There’s likely a very good reason why the customer hasn’t yet made those investments-- in these use cases, simply having the data at hand did not generate enough ROI to justify the investment, either because the data was too costly to capture, too voluminous for humans to make sense of, too sparse for humans to detect signal from, or the signal too subtle for humans to pick up on. These are all types of problems where an AI model can outperform humans and extract enough value from the data to justify the investments in collecting and managing it.
The lack of AI product startups succeeding as challengers to incumbent software vendors should not be taken as a sign that the promise of AI is overblown. The impact of AI already deployed by incumbent software companies points to the transformative power AI models, provided they are positioned in the right part of the value chain to deliver ROI. It just means that the breakout businesses built around AI won’t look like the software startups that came before them.