It's not just about having an AI. Having clean data, having governed data, and now having contextual data will be very much necessary to roll out any agentic systems.

Kartik P.Kartik P.Director- Data & AI, Pizza Hut

In most systems, a wrong number gets caught by a person. Someone notices it looks off on a report and corrects it before it travels far. Agentic AI removes that check, since the agents act on one another’s output with no one reviewing the figures in between. A flawed value doesn’t trigger an error or halt the process; it feeds the forecast that sets staffing levels, and every step downstream holds together on a figure that was wrong from the start. The problem becomes visible only later, in a decision that proves costly for reasons no one can readily trace.

That risk grows with the number of systems involved, and few operations run more of them than a global restaurant chain. Kartik Pillai is Global Head of Data & AI at Pizza Hut, overseeing the data infrastructure behind roughly 5,000 US restaurants and about 12,000 worldwide. He came to it from Yum! Brands, where he built the enterprise data strategy and governance foundations across its global markets, after leading the data platform and the menu and e-commerce migration behind Pizza Hut’s 5,500-plus stores. The pandemic accelerated mobile ordering, but third-party platforms and the multi-vendor ecosystem predated it. Every transaction now runs through a chain of systems for payments, inventory, labor, kitchen, and delivery, most built by different vendors and none designed to agree with the others. Coordinating that by hand is what makes the work a candidate for agents, and what makes the data beneath them so unforgiving.

The menu is a good measure of the scale. A single location can run a different menu for its own app than for third-party delivery, with another again for in-store dining, and across 12,000 stores the volume runs past anyone's ability to check it by hand. "It's physically not possible for everyone to review every menu published that goes through and trickles down to the system," Pillai says. Once that much published data moves without a person in front of it, the condition of each feed stops being a back-office concern. "Any data anomaly or missing data has multifold effects, almost an exponential way of exposing your gaps," he says, and at that scale a single wrong value doesn't stay where it started. It can travel out of one fragmented system into the forecast that sets a Friday staffing plan and surface as a real operational problem well downstream, long after anyone would think to connect the two.

When one number means four things

Take promise time, the window a customer is told to wait for an order. It looks like one of the simplest numbers in the business until you ask when the clock starts. "Does the clock start from when you placed an order, or when the product hit the make table, or, if it's an Uber Eats order, from when the driver picked it up?" Pillai asks. Each team along the chain can answer differently, and each can be certain it's measuring what everyone else is, which means the same KPI carries four definitions at once without anyone noticing. The same thing happens with a measure as common as same-store sales, where the headline looks shared and the calculation underneath it isn't.

His fix starts well before any dashboard, with the question of who gets to define the number in the first place. "It's a common misconception that the IT teams own the data. The reality is they don't," Pillai says. "They're the guardians of your data, and there has to be a business owner who certifies what a metric means and how it's calculated." When two functions read the same measure differently, he routes the disagreement to a data governance council with the standing to settle it across the enterprise, and that insistence on a single owner per metric is increasingly what decides whether governance holds once AI enters the picture.

The reason the ownership question isn't academic is what inconsistent definitions do to a model. Feed a machine learning system two versions of the same measure and the contradiction compounds into unreliable output, the kind of hallucination Pillai wants designed out before an agent ever acts on the number. Governance also runs into how evenly an organization can read its own data. "Not everyone is as data literate as we assume," he says, and while AI has made a chart easier to interpret by surfacing a next best action, knowing what a number says is still a step short of knowing what to do about it.

The gaps that don't show up on a screen

The work of keeping data clean is never finished, and consistent definitions keep a bad number from meaning different things in different places, but they don't stop a bad number from entering the pipeline. That job falls to observability, which reached software engineering long before it reached the data side. Pillai treats it as part of how pipelines get built rather than a layer added once they're running. "We have a robust observability layer and a data quality layer embedded into our pipelines. These are not afterthoughts," he says. With thousands of pipelines moving data in and out, building the checks alongside the engineering is the only version that holds at that volume.

What that rigor costs is set by what the data is worth, rather than applied evenly across everything. Some domains can absorb a 24-hour delay without the business noticing, while others can't survive more than two or three hours down. "Your data asset determines your SLAs, your rigor around observability and quality," Pillai says, with the level tied to the business use case behind each pipeline. The triage is the point, because the alternative is spending the same scrutiny on a feed that doesn't need it while under-protecting one that does.

That weighting matters most where agents depend on one another. Chain several together, each acting on the last one's output, and a single bad input never surfaces as an error. It moves through the system instead, shaping a forecast in one place and a reorder in another, until the result looks plausible and is wrong. Observability is what gives Pillai a chance to catch it while it's still one problem rather than a dozen.

Context the model doesn't ship with

Even accurate data can mislead a model that doesn't understand the market it came from. Pillai points to the working week, where Friday is part of the weekend in some countries and an ordinary business day in others, so a system trained only on Western demand patterns can misread the rhythm of an Asia Pacific or African market entirely. "Giving the system the right context greatly reduces the risk and cost of hallucinations that lead to bad decisions," he says, and the same holds inside a single brand, where a general model off the shelf doesn't know a company's policies or procedures and fills the gap with a guess.

The infrastructure under all of this is, in his view, largely solved, since the major cloud providers made the platforms reliable. The harder gap now is organizational. IT spent a decade getting faster, moving from waterfall to agile and on to ever-shorter delivery cycles, while the business it serves stayed closer to its old pace. "The IT and infrastructure teams have been agile, but the business is not," Pillai says, and the result is a widening lag between what the market is doing and how quickly a company can answer it, even when the data to answer is already there.

That lag shows up most clearly when a promising pilot meets a live restaurant. "A lot of POCs succeed, but they don't succeed in production, mainly because POCs are designed to succeed," Pillai says. A pilot runs in a controlled room with clean data and known inputs, where everything is accounted for in advance, and production offers none of those guarantees. It's part of why so many agentic projects stall before they reach broad deployment, and why Pillai keeps coming back to context as the thing that closes the gap. "It's not just about having an AI. Having clean data, having governed data, and now having contextual data will be very much necessary to roll out any agentic systems," he says.

None of this is unique to Pizza Hut, in his view. "As an industry, we went through data warehousing, data lakes, and all those developments, but we largely bypassed data quality and governance and jumped straight into agentic AI. Now a lot of mature organizations are going back and fixing their data foundations," Pillai says.

Which leaves the question the industry keeps putting off, the hope that AI can repair the foundations it's being built on top of. Pillai doesn't expect it to. "Will AI fix your bad data, your bad metric definitions, your data governance? Most likely not," he says. "It will just fail silently."

A loud failure gets caught. A silent one gets shipped.