Every retail business already has an intelligence layer. For most of the last two decades it looked like a dashboard: sales by day, stock by SKU, campaign performance, a weekly email from the analytics team. The intelligence was real. It just stopped at the screen. A person had to read it, decide, and then go and do the thing in three other systems.
That is the part that is changing. The new intelligence layer does not stop at the screen. It reads the same data, proposes an action, and increasingly carries the action out, with the merchant approving the decisions that matter and letting the rest run. This post is about what that shift looks like from the merchant's chair, without the vendor gloss.
The numbers say it has started
The buyer side is easiest to measure. Adobe, which tracks more than a trillion visits to US retail sites, reported that traffic referred by AI assistants grew 393% year over year in the first quarter of 2026, and that by March those visitors converted 42% better than other channels. A year earlier AI referrals converted worse than average. The shopper who arrives through an assistant now arrives ready to buy.
The merchant side is earlier, and honest surveys show it. Salesforce's State of Commerce found that 28% of commerce organisations use agentic AI today, with another 44% planning to within six months. Gartner's forecast is blunter: over 40% of agentic AI projects will be cancelled by the end of 2027, mostly because the cost and the value were never clear. Both things are true at once. The shift is real, and most of the first attempts will be thrown away.
Four levels of a store's intelligence
It helps to think in levels, the way people talk about self-driving cars. A retail business sits somewhere on this ladder, and different functions can sit at different rungs.
Level one is reporting. The system shows what happened. Sales are down in the north region. Stock of a fast seller is at four days. A human notices, or does not.
Level two is recommendation. The system says what to do. Reorder 400 units of this SKU. Move this campaign's budget to the channel that is converting. The human decides and executes elsewhere.
Level three is delegated action with approval. The system prepares the whole action, the purchase order filled in, the price change drafted, the campaign built, and waits for a yes. On approval it executes. The human's job shrinks to judgement.
Level four is autonomous within a policy. The system acts without asking for a class of decisions the merchant has pre-approved: restock below a threshold, pause an ad set that breaches a cost cap, answer a delivery question. Anything outside the policy escalates to level three.
Most retailers today are at level one, with pockets of level two. The interesting work in 2026 is moving specific, well-bounded functions to level three and a few to level four. Nobody serious is trying to run a whole business at level four, and the Gartner cancellation number is largely the story of teams that tried.
What Walmart's numbers tell you
The clearest public example of a retailer moving up the ladder is Walmart. Its shopping agent, Sparky, is customer facing, but the reported effects are operational. On its first quarter 2026 earnings call the company said customers using Sparky have an average order value about 35% higher than non-users, units bought through it more than quadrupled quarter on quarter, and weekly active users doubled. Behind the customer agent, Walmart described using AI to decide how inventory is positioned and which node fulfils an order in real time. That is a level three or four decision that used to be a planner's weekly job.
You do not need Walmart's scale to do the equivalent. You need a bounded decision, clean data for it, and a place for a person to say yes.
What changes for the merchant
The job changes from operating to supervising. A merchant at level one spends the morning in dashboards and the afternoon in admin panels. A merchant at level three spends the morning reviewing a short list of proposed actions and the afternoon on the decisions no system should make: which brands to carry, what the store stands for, whether to open a second warehouse.
The data has to be good, because now something acts on it. A wrong stock count on a dashboard is an annoyance. A wrong stock count that drives an automatic reorder is a cost. Gartner's retail analysts put it plainly: autonomous decisions need clean data pipelines, real-time inventory signals and integrated systems. Most of the work of moving up a level is data work.
Policies become explicit. Level four only works if the merchant has written down, somewhere a machine can read, what it is allowed to do. Reorder up to this value. Discount up to this percentage. Never email a customer twice in a week. Most businesses have these rules in someone's head. Delegation forces them onto paper.
What does not change
Judgement stays human. An agent can prepare a Diwali campaign; it cannot decide whether Diwali is the moment to reposition the brand. It can flag that a supplier's lead times are slipping; it cannot decide whether to fire the supplier.
Accountability stays with the business. If an agent over-orders, the merchant owns the stock. Every serious deployment I have seen keeps a person on the hook for every irreversible action, and that is a feature, not a limitation.
The customer still wants to be treated like a person. McKinsey's consumer research found 68% of shoppers had used AI tools in the past three months and 62% had used them to compare prices or brands. They are happy for a machine to find the product. They are not happy for a machine to fob them off when the delivery is late.
Where to start
Pick one decision that is frequent, bounded and reversible. Restocking a fast-moving category. Pausing under-performing ad sets. Answering order-status questions. Move that one decision to level three: the system prepares the action, a person approves it. Measure how often the person changes the proposal. When the answer is "almost never" for a month, write the policy and move it to level four.
Then pick the next one. Retail businesses will not become intelligent in a single project. They will move up the ladder one decision at a time, and the ones that treat it that way will still be running their agents in 2028 while the big-bang projects are being cancelled.