91 VIEWPOINT Jason Gladu is chief alliances officer at Convertr We see it all too often: an enterprise runs an AI pilot, the model is impressive and the demo works well, but the project still stalls. In fact, Gartner expects more than 40 per cent of agentic AI projects to be cancelled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls. Notice that nobody is cancelling because the model wasn’t clever enough. Yet, for the past three years, the assumption has been that advantage comes from having the better model. It doesn’t – your competitor can buy the same one. It’s the data sitting underneath the model that is your competitive advantage. It’s not just data quality that matters. 38 per cent of US executives trust AI agents with data analysis, according to PwC. But that drops to around 20 per cent when AI is expected to take real-world action. That’s not irrational – it’s a matter of irreversibility. An agent that drafts a recommendation can be corrected. One that emails a customer, issues a refund or sells an audience segment can’t. That’s why provenance is the real rate limiter. How far you can safely extend an agent’s autonomy comes down to how well you can evidence what it’s acting on. This flips the usual view of governance on its head. It’s traditionally been seen as the thing that slows you down. In an agentic operation, it’s the opposite: governance is what lets you increase autonomy. Companies that can prove their data is trustworthy can automate faster than those that can’t, and with fewer checks. The economics favour catching problems early. Checking a record when it comes in is deterministic – exact, repeatable and close to free. Asking a model to work out later whether a record can be trusted costs computing tokens every time you ask and only ever gives you a probability. Microsoft has put the potential reduction in AI token costs from better context as high as 80 per cent. In one enterprise deployment, Convertr blocked 811,000 bad requests at the door in a single year, and the organisation hasn’t had data subject access requests since go-live. It paid for itself in three weeks. My bet for the next three years is that the enterprises pulling ahead won’t be the ones with the best models. They’ll be the ones that can answer a simple question about any record in their estate: where did this come from and who gave us permission to use it? The constraint on agentic AI is no longer model capability, but whether a company can prove where its data came from. Most enterprises are optimising the wrong variable Data, not the model, limits enterprise AI Photo: iStock/gonin
RkJQdWJsaXNoZXIy NzQ1NTk=