Last summer, a blockbuster report out of MIT landed with a thud across the business world. It found that roughly 95 percent of investments in generative AI had yet to deliver meaningful returns, and it became a rallying point for a growing cohort of AI skeptics who had been questioning whether the technology’s promise was outpacing its practical value.
Then analysts began referring to the final quarter of the year as the “Great Decoupling,” a period when markets appeared to stop rewarding companies for AI ambition alone. Instead, investors began to penalize organizations that could not clearly demonstrate how AI investments were translating into revenue, efficiency, or margin improvement.
Last year, Gartner also moved generative AI into what it calls the “Trough of Disillusionment,” signaling that the early excitement phase had passed and that businesses were entering a more difficult, reality-driven chapter of adoption.
The research firm said its reasoning for the downgrade centered on a widening gap between expectations and actual performance. By mid-2025, fewer than one-third of AI leaders had reported that their CEOs were satisfied with the returns generated by AI initiatives. Companies were spending close to $2 million per project on average, often without seeing proportional gains in productivity. Integration challenges, particularly when layering AI onto legacy systems, proved far more difficult than many organizations had anticipated.
Gartner also found that more than half of the companies it surveyed said their underlying data environments were not prepared to support AI at scale, effectively preventing pilots from moving into broader deployment.
That readiness gap is a critical part of the story, especially for restaurants, says Carl Orsbourn, senior vice president at Invisible Technologies, an AI technology and solutions provider.
“The technology itself isn’t the thing that’s failing,” he says. “It comes back to things like change management. It comes back to things like integrations. And most importantly, it comes back to the challenge of data.”
Restaurants are particularly exposed. A typical multi-unit operator may rely on 15 to 25 different systems to run the business, from POS and loyalty platforms to scheduling, inventory, marketing, and delivery tools. Over time, those systems have created what Orsbourn describes as a maze of disconnected data sources.
“What we’ve done is create a series of data silos that aren’t talking to each other,” he says, adding that the fragmentation becomes fatal once AI enters the picture. “When good AI meets bad data in a fight, the data always wins. So, if you are going to embrace AI, you have to do it with a data-first mentality.”
That shift requires rethinking how AI is framed inside organizations, adds Jen Kern, CMO at Qu, a restaurant software development firm. Too often, she says, it is discussed as a standalone capability rather than as something deeply dependent on the systems and processes beneath it.
“There is this view that AI is a feature,” Kern says. “Many operators fall victim to shiny object syndrome, bolting new systems onto shaky foundations. When integrations and data aren’t prioritized, everything eventually breaks.”
She also points to a growing sense of AI fatigue across the industry. After years of constant announcements, new tools, and shifting expectations, many teams feel drained rather than energized. The pace of change has created stress around adoption, retraining, and long-term relevance, especially when early experiments fail to deliver immediate wins.
“While AI fatigue is real, it is here to stay and can drastically improve operations,” Kern says. “But it requires clean, unified data and should be built into your foundation, not tacked on later.”
The Right Foundation
Moving beyond shiny object syndrome often requires restaurants to fundamentally rethink how they approach technology. Rather than treating each system as a standalone solution, operators need to view their tech stack as a single, connected ecosystem. Before investing in AI, data must be centralized, standardized, and able to move freely across platforms.
One key step is adopting an API-first approach. Too many organizations attempt to layer new AI tools on top of legacy systems that were never designed to communicate effectively. When core platforms like POS, loyalty, kitchen displays, and inventory systems lack open APIs, data gets trapped, forcing teams to rely on manual exports or fragile workarounds.
Another critical piece is standardizing how data is defined and stored. AI systems struggle with inconsistency. If the same menu item is labeled differently across locations or channels, analytics break down and insights become unreliable. Creating a single source of truth for menus and auditing customer databases to eliminate duplicates and inconsistencies are foundational steps that enable more accurate predictions and personalization.
Not all data carries equal weight. Building toward AI means prioritizing high-signal inputs, such as detailed transaction histories, real-time inventory and cost data, and labor patterns tied to speed of service. These streams form the backbone of more advanced applications.
Equally important is centralization. Sales data in the POS, guest feedback in third-party tools, and labor costs in spreadsheets limit visibility when kept separate. Bringing those streams together in a centralized repository allows AI systems to see the full picture rather than isolated snapshots.
This matters because more advanced forms of AI require broad context. Agentic systems, for example, need visibility into multiple dimensions of the business to make smart recommendations or take action. Without unified data, those systems simply cannot function as intended.
Kern believes agentic AI will be the most transformative development for restaurants in the coming years. Unlike reactive systems that respond to prompts, agentic AI is designed to pursue goals, plan actions, and execute tasks autonomously. Instead of producing a single output and stopping there, these systems can monitor conditions, interact with other software, and follow through until an objective is met.
Agentic AI differs from traditional generative tools in several ways. These systems can break down complex objectives into smaller tasks, monitor real-time conditions, use external tools like APIs or databases, and maintain focus on a goal over time. Rather than acting as assistants, they function more like digital team members.
In practice, that could mean rerouting a delayed delivery without human intervention, resolving a customer issue end to end, or analyzing labor levels against live order volume to recommend staffing adjustments in the moment.
From Business Intelligence to Decision Intelligence
So, is AI truly worth the investment, especially compared to traditional analytics teams? One way to answer that question is to think about the distinction between business intelligence (BI) and decision intelligence (DI), says Tammy Billings, director of business development at restaurant technology and decision science firm SignalFlare.ai.
“Business intelligence is looking at reports and data,” she says. “That’s what most of our analytics teams are doing now.”
Billings describes business intelligence as the essential foundation. It involves gathering data, organizing it, and presenting it through dashboards and reports. Decision intelligence builds on that base, using AI to forecast outcomes, recommend actions, or automate decisions altogether.
“Decision intelligence is that next layer up,” Billings says. “I like to say it’s a business intelligence tool with artificial intelligence and decision intelligence.”
Her advice is straightforward. Before diving into advanced tools, organizations must ensure their BI and data infrastructure are solid. Without that groundwork, even the most sophisticated AI systems will struggle to deliver value.
Orsbourn has seen what happens when analytics outpace decision-making. In a previous role running the finance team for a multi-unit restaurant chain, he watched as the company’s analytics group produced roughly 360 different reports, creating a flood of information that overwhelmed teams rather than helping them act.
He says dashboards alone don’t help unless they translate into timely, specific actions. That gap between insight and execution is what he calls the latency effect. The longer it takes to act on information, the less useful it becomes. Modern AI systems, he explains, can shrink that gap by accelerating or even automating decisions.
He illustrates the difference with an operational example: the contrast between learning after the fact that equipment might need service versus anticipating a failure and acting before it happens.
“That’s the type of decision we want to see in our restaurants,” Orsbourn says. “And that means we need to be out of the latency, and we need the decision-making speed for that kind of thing to happen.”
Trust remains one of the biggest hurdles, too. With traditional BI tools, operators can retrace the math and verify results themselves. That transparency builds confidence. AI systems, by contrast, often operate as black boxes. The reasoning behind a recommendation isn’t always obvious.
“With a BI tool, you can go back and add up the numbers and go, ‘OK, it’s correct,’” Billings says. “How do you do that with AI? It’s a lot of learning and it’s a lot of training those models.”
Billings says trust develops over time. Early adopters closely monitor outputs, provide feedback, and adjust models until results align with real-world experience. That process explains why successful brands spend extended periods testing and refining AI before expanding deployment.
In that sense, she says, early adopters are laying the groundwork for those that follow.
Ultimately, moving from data to decisions requires more than clean inputs or advanced tools. Real value emerges when operators trust the system enough to act on what it produces. That means investing not just in technology, but in learning, confidence, and organizational change.
“Now, when do you start? Only you can make that decision,” Billings says. “But a lot of people started five years ago.”
