7-Eleven AI Leader On Building Decision Systems Instead Of Another Layer Of Dashboards

Most enterprise AI still ships as a reporting layer, generating dashboards and scheduled outputs that rarely change an operational decision. At 7-Eleven's India business, the analytics function runs on the reverse premise, where a model earns its place only when it moves a number someone already answers for, whether that's labor cost, sales, or forecast accuracy. The result is a working line between AI that informs and AI that decides, and it starts by refusing to build anything that can't be tied to a result.

Aroon Grover is GM of AI, Analytics, and Strategy at 7-Eleven, leading data science across the retailer's India network under Reliance Retail, with production forecasting, pricing, and workforce models running on more than 10 million rows of operational data. Across more than two decades in Asia-Pacific, he has built decision systems for retail and consumer goods, founded an AI advisory practice, and led large-scale computer-vision deployments. For Grover, the fastest way to lose a room is to lead with the model instead of the problem it solves.

"We're not looking at descriptive analytics or dashboards, because that gives you a rear view, and rear views aren't great when you're driving a car," Grover said. Roughly 90% of the team's work is predictive, aimed at what happens next rather than what already happened. Each project is pinned to a measurable return, which is how AI stops being a visibility exercise and starts operating as a tool.

• Outcome before model: The clearest proof sits in workforce planning. By forecasting store traffic and buying behavior, Grover's team reduced workforce cost-to-revenue by approximately 48%, one of the larger efficiency gains across the retail operation and the kind of cost-versus-growth tradeoff most IT leaders now face. Demand forecasting at the store, SKU, and day level lifted accuracy by 25% and reduced inventory waste by 18%. "If I'm talking about cost optimization, did I really save cost? There is a measurable, quantifiable return on investment on what AI is producing, rather than it being a dashboard or a fad in the company," Grover said.

• Adoption in capsules: Getting people to act on model output proved harder than building the models. An early attempt to push advanced systems broadly stalled, and Grover scaled back to sequencing by risk. Low-stakes forecasts came first, like packaged snacks with no real shelf life, so a supply chain lead could trust that an over-order wouldn't spoil. Shelf-life-sensitive items like fresh sandwiches waited until confidence was established. "Everything we do has to be bite-sized. If you do everything in understandable little capsules, people adopt it," Grover said. "Don't give someone an advanced model that changes every SKU in every store."

• Questions become features: Grover's models improved once he stopped designing them around what the math could do and started encoding the questions his CEO actually asked. Whether last Wednesday behaved like the one before, whether the week ran unusually hot or rainy, whether this Christmas tracked the last, which stores counted as affluent, each became a feature. Affluence got operationalized through the stores selling the most expensive chocolate. Translating the business's own language into inputs is how he orchestrates the models, and it doubles as evidence a leader was heard. "Instead of just looking at a machine learning feature inside a model, I started building in elements of what the conversations were like. That's how it gained trust, and it also improved the model," Grover said.

Adoption didn't only come from the top. Generative AI entered the company through employees using ChatGPT on their own and carrying it back to their work, a bottom-up pattern now common among workers. "When you talk about embedding technology, it's not only driven from the top, it's driven from the bottom," Grover said.

• Knowing when not to build: Discipline also means declining the model. Not every request has a data answer, so the team first asks whether the data is adequate, whether a model is required at all, and whether human experience would make the better call. When there's no precedent, like a promotion structure that's never been run, they surface the caveats up front and may offer only a rough surrogate. He keeps human review in the loop while warning against treating a model as something to override on instinct, which is why owning the kill switch sits at the center of the design. "Every problem that you have does not have a solution in data. Sometimes you have to go out there and experiment."

• ROI under simulation: A clean ROI figure invites the obvious question of what happens when conditions change, and Grover answers it with simulation. Every solution ships with its assumptions exposed, so teams can watch the return shift as inputs move, sometimes turning negative. A promotion reaching 20% of its audience returns one figure and 50% returns another, giving teams a range to test rather than a single number to trust on faith. "Assume you go to war and you think this many armies wins. Now assume the enemy shows up with five more that day. What happens?" Grover said.

For all the modeling, Grover expects the decisive skill to be a human one: knowing what to ask. Language models return an answer to whatever they're given, which moves the advantage to whoever frames the question well, a decision-rights problem now surfacing on more leadership agendas.

"Asking the right question is going to be your next big skill, and it always has been. CEOs are really great at asking good questions," Grover said. "You cannot assume that when you present a problem, it will give a solution. You have to know what outcome you're drawing out." For an operation that measures AI by the numbers it moves, that clarity is what turns a model into a decision.