Coca-Cola is building AI agents – and it does not mean automating the can. The real action is in the operations: pulling data from internal databases, synthesizing reports that once took a human analyst three days to produce, and making decisions that used to travel through a chain of middle managers.
Kevin Hague, the company’s senior vice president of emerging technologies, laid out the architecture at NTT’s Upgrade 2026 conference in San Jose, California.
For founders watching the AI wave from the sidelines – perhaps because their product is physical, their market is regulated, or their team is lean – what Hague stresses deserves a closer look. Forget the scale; for entrepreneurs, the structural logic underneath his message is what just might unlock unseen potential.
The intelligence gap sits in operations, not products
Most AI coverage gravitates toward companies finding ways to add AI into their product through smarter recommendations or generative features. Hague’s talk, however, went the other way.
Coca-Cola is deploying agents against the operational layer, describing a system designed to eliminate a familiar bottleneck: the person who owns the spreadsheet.
“They’re the owners of that spreadsheet,” Hague told the audience, “and you know to call them on the phone or via Teams to get that data.” Coca-Cola decided to stop making the call outright.
The company went from 60 AI use cases in 2024 to roughly 700 in 2025 by replacing that call with an agent. The throughline in the strategy was clear; put a highly-deterministic AI agent on top of a specific database, make it accurate to 99%, and let it handle the extraction and synthesis a human used to spend hours on manually.
If the product isn’t digital, most founders might dismiss AI as irrelevant. But Hague’s architecture suggests they have been looking in the wrong place: the real opportunity is in operations. A 15-person team that deploys agents against inefficient internal processes can operate like a company three times its size.
Vertical agents beat general-purpose models
Coca-Cola is not building a centralized AI orchestration system, but a hierarchy. At the base sits what the executive deemed “individual contributor” agents – deployed directly on top of specific databases, highly deterministic, operating at roughly 99% accuracy. Above them are “manager” agents, then “director” level agents, and at the top, a “VP level” that synthesises everything underneath.
“It’s a pyramid just like a human organisation. At the base of the pyramid is all the data, and it trickles up until you finally get why something is happening,” Hague explained.
The architecture has a name in the broader industry: vertical multi-agent orchestration. And Coca-Cola is far from alone in adopting it. In fact, Bain & Company’s April 2026 research argued that modern agentic AI demands exactly this kind of layered structure – featuring orchestration, observability, and governed data access – with agents deployed as separate yet connected versioned services that can be scaled, updated, and rolled back independently.
German chemicals giant BASF, meanwhile, is building something similar, describing its approach as “a supervisor of supervisors.”
What sharpens Hague’s framing is his distinction between data-driven and intelligence-driven companies: whereas a data-driven company collects information, organizes it and hands it to a human to interpret, an intelligence-driven one builds the interpretation within the system and, by the time a human executive engages, the question has already been framed, context assembled, options narrowed.
“If you’re a general or a commander, you’re going to ask, why is something happening? How is something happening? What should we do about it? That’s not really a data-driven enterprise; that’s an intelligence-driven enterprise.” Hague stressed.
The takeaway for founders should be that the defensible product here is not the general-purpose AI model, but rather agents that know one domain or workflow deeply enough to slot into this kind of layered architecture.
European founders building vertical AI for regulated industries – healthcare, energy, financial services – are particularly well-positioned, because domain depth and compliance knowledge are exactly what base-layer agents require.
The real product: change management
If there is a single number worth remembering from Hague’s talk is his ratio: 80% change management, 20% technology. “The technology stack is actually already here. Really what we’re looking at is a change management problem.”
Every AI project at Coca-Cola requires three roles: a project manager who understands the business context, a technical person who can write the spec, and a stakeholder who will use the output. “If you’re missing one of those three pieces, it’s going to fail,” Hague said.
And the evidence supports the claim. An MIT NANDA report analysing over 300 enterprise AI deployments found that 95% of pilots delivered no measurable P&L impact. Hague cited the study directly on stage: the technology was working, he argued. The problem was rollout, which brought him to the ratio that anchors his entire approach.
This framing opens a very specific commercial lane that founders can take advantage of. Enterprises across the region know they need AI; what they are lacking is the internal capacity to train, deploy, and integrate it into existing workflows without fracturing operations.
Those who sell the implementation alongside the technology may find shorter sales cycles and stickier contracts than those selling pure capability.
The unresolved question
Hague was candid about the timeline. Even under optimistic assumptions, reaching 1,000 active users of Coca-Cola’s intelligence-layer tools by the end of 2026 would make him “super happy.” Full transformation is further out – 2027, 2028, maybe longer.
“The chances of that happening is probably pretty slim if I’m direct. Because it just takes time,” the executive admitted.
That timeline gap is where the opportunity lives. For every Coca-Cola spending three years rewiring its data architecture, thousands of firms are still running the spreadsheet-owner model Hague described at the beginning of his talk.
Meanwhile, the companies building tools, services, or frameworks to help them move up the stack are selling into a structural shift that enterprises are only beginning to name.
The hierarchy is where enterprise AI is going. The question for founders is whether they are building the tools that help companies climb it, or still selling into the flat org chart that came before.
Disclosure: The reporter was invited to Upgrade 2026 by NTT and travel expenses were covered by the company.
