Agentic AI was the most common pitch at LEAP, and two studies at the same event found the returns lagging

No product category was described more often at LEAP 2026 than agentic AI. IBM and Ejada deployed an AI agent to accelerate enterprise modernisation. Horizontal Digital demonstrated an agentic concierge built with Salesforce. Link Development launched an agentic AI operating model. Unifonic unveiled an agentic customer experience strategy. Alteryx brought an AI agent with a two-way SAP link. ARIS pitched process context as the missing layer for enterprise agents. AgilePoint offered governed low-code automation. Salam and UnifyApps signed on agentic automation for Saudi enterprises. Cloudflare spotlighted agentic infrastructure and security. Liferay brought an AI hub with a governance argument. Two research findings presented at the same event describe where those products currently land, and TechScoop reported each of them separately as it appeared. Deloitte found that most Middle East AI projects remain in pilot mode. PwC found Saudi firms reporting gains in operating model ahead of financial returns. Set side by side against the pitch volume on the floor, they read differently than either does alone. Neither finding is surprising and neither is damning. Both describe a normal position for a technology roughly two years into enterprise deployment, and the gap between them and the sales pitch is the useful thing to examine. The distinction PwC's finding rests on is worth drawing carefully. Operating-model gains are real: a process runs with fewer handoffs, a team handles more volume, a task that took three steps takes one. Financial returns require that improvement to show up as reduced cost or increased revenue, which needs something further — either the freed capacity gets used for something else, or headcount changes, or throughput rises enough to matter at the top line. Most organisations achieve the first and stop short of the second, because the second requires management decisions that technology does not make. A team that processes claims 30% faster has a genuine operating gain, and it produces a financial return only if the organisation does something with the 30%. Deloitte's pilot finding points at a related bottleneck with a different cause. Agents stay in pilots for reasons that are mostly not about model quality. An agent that takes an action — issues a refund, updates a record, sends a payment instruction — needs permissions, an audit trail, an error-handling path and somebody accountable when it acts wrongly. Those are governance and integration problems, and they are harder than the model work. This is why the more credible products at LEAP were the ones addressing that layer rather than the agent itself. ARIS pitching process context is an argument that agents need to understand the process before they can act in it. AgilePoint's governed low-code framing and Liferay's governance argument are about the same thing. Cloudflare's agentic infrastructure and security positioning treats agents as a workload with a security perimeter. Those vendors have identified where deployments actually stall. The Saudi regulatory environment sharpens all of it. SDAIA's AI Adoption Framework requires a defined human intervention point and a named model owner for public sector deployments, and reaches suppliers through procurement. An agent operating autonomously inside a government workflow has to satisfy those requirements before it runs, not after. That is a constraint on speed and a considerable advantage for the eventual result, because organisations that solve accountability before deployment do not have to unpick it afterwards. The clearest counter-example from the week is worth naming, because it shows what the far side looks like. Saudia Cargo says AI is now running its freight operations rather than advising them. That is the distinction that matters — a system making operational decisions rather than producing recommendations a human then actions. One named enterprise saying it has crossed that line is more informative than a dozen product launches. What would move the whole category from pitch to evidence is more disclosures of that kind: a named Saudi organisation, a named process, a volume figure, and a before-and-after. The research studies establish that most projects have not got there. The product announcements establish that the tools now exist. The missing layer is customers willing to say publicly what happened when they used them, and at present that list is very short.