.gif){kind=spacer}
Every retail leadership team investing in store AI ends up at the same question: what should the agent do first? Most answer it by picking the use case with the biggest payoff. The retailers getting adoption pick the one with the smallest downside, and let the payoff follow.
TL;DR
Rank in-store AI use cases on four things: the value of a right answer, how ready the data is, how much integration the agent needs, and what a wrong answer costs. On that basis, start with product location and stock checks, then new-hire procedure questions. Add price and promotion questions third. Leave returns decisions and BOPIS exceptions for last.
The last two carry the most value per case, which is why retailers are tempted to start there. They also depend on the messiest data and the most systems, and a wrong answer costs real money. The rule of thumb: let the agent read data before you let it change anything.
How do the five in-store AI use cases compare?
Most lists of in-store AI use cases treat every idea as equally ready. They aren't. Here is how the five compare on the four tests.
Risk depends on what the AI agent can do. Start with read-only guidance, then add order changes, refunds and price overrides with clear approval rules.
1. Why should product location and stock checks come first?
"Do we have it, and where is it?" is the question store associates hear most. When they don't know the answer, they leave the customer to walk to a terminal or find a colleague, and the customer waits or leaves.
This use case scores well on three of the four tests. The agent reads from two systems, inventory and the store layout, and can't change anything in either. A wrong aisle costs a few steps, not a refund. The weak spot is data. Without item-level RFID, store inventory records average about 63% accuracy, according to the Auburn University RFID Lab and GS1 US. Product locations are usually right, but stock counts often aren't.
The associate asks by voice, and the agent answers with the aisle, the bay and the stock count. When a count looks unreliable, for example when it shows stock but no recent sales, the agent says so, and the associate checks the shelf before promising the item.
Measure answer accuracy against the shelf. If accuracy holds for several weeks, your store data is ready for the use cases further down the list. That makes this use case a test of your data as much as a service to associates.
2. Why should new-hire questions come second?
A first-week associate doesn't know how to build an endcap, run a markdown or log a damaged item. They interrupt a manager, ask whoever is nearby, or guess. With high turnover, every store has a steady supply of first-week associates, so the same questions return month after month.
This use case is almost as safe as stock checks and needs even less integration. The agent reads the store's own procedures, so its main dependency is a document library, not a live system. Start with routine procedures. Route safety, cash-handling and policy exceptions to a trained colleague. The real risk is outdated content. If the SOPs say one thing and managers teach another, the agent will repeat the wrong version, so the SOPs need a review before launch.
The new associate asks on the floor, at the moment they need the answer, and the agent walks them through the procedure one step at a time. The value shows up in two places: fewer interruptions for managers, and new associates who handle more on their own in their first weeks.
Measure manager time spent on repeat questions, and time to productivity for each new cohort.
3. When should retailers automate price and promotion questions?
The shelf tag shows one price and the till rings up another. In the 2024 National Price Verification Survey, Massachusetts inspectors found that about 2.1% of 27,750 items checked across 327 stores rang up at a different price from the one displayed. Each mismatch usually needs a manager to resolve, while the queue waits.
The data here is the best of all five use cases, because the pricing system is the record of what an item costs. The difficulty is effort and risk. The agent needs live access to POS, the promotions calendar and loyalty pricing, and its answer becomes a commitment at the till. If it quotes the wrong price, the store either honours it or ends up with an unhappy customer.
Once lookups are working, the agent can confirm the live price, list the promotions that apply, and explain any gap, such as an expired sign or a loyalty-only price. If the shelf tag is wrong, it raises a task to replace it, so the next customer doesn't hit the same problem.
Measure the share of price questions resolved without a manager, and repeat mismatches on the same shelf.
4. Why should returns decisions wait until later?
A customer brings back an item without a receipt, or after the return window has closed. The associate isn't sure what the policy allows, so the answer depends on who is on shift. The stakes are large. US retailers expected 15.8% of 2025 sales to be returned, worth about $849.9 billion, and estimated that 9% of returns were fraudulent, according to the NRF.
Returns score high on value and poorly on everything else. Most return policies leave room for manager judgement, and that judgement is rarely written down, so the agent works from incomplete rules. It also needs order history and customer records, and a wrong answer means money leaves the store. That combination makes returns a poor first use case and a strong fourth one.
Start with the agent advising, not deciding. It looks up the order, applies the written policy, and tells the associate the likely outcome, along with the policy clause behind it. Anything outside the policy, such as a high-value item or a repeat return without a receipt, goes to a manager with the details attached.
Measure manager escalations per return, and how consistent outcomes are across shifts and stores. Let the agent decide routine returns only after its recommendations have matched manager decisions for a sustained period.
5. Why should BOPIS exceptions come last?
An associate picking a buy-online-pick-up-in-store order finds an empty shelf. Now they have to choose a substitute, reach the customer and update the order, often while also covering the floor. Researchers at Auburn University note that these BOPIS out-of-stocks happen because store inventory records are rarely fully accurate.
BOPIS exceptions may be the most valuable use case on this list, because each order is a sale already made. They are also the hardest to get right. The agent depends on real-time pick status and stock data, connects to order management, inventory and customer messaging, and contacts the customer directly. A poor substitute or a wrong ready-time reaches the customer straight away. This is why stock checks come first: they prove whether your inventory data can support BOPIS before the customer is involved.
When that groundwork is in place, the agent ranks problem orders by pickup time and suggests substitutes from live stock and the customer's past choices. It then asks the customer to approve by SMS or call, and updates the order.
Measure exception resolution time, substitution acceptance rate and orders ready on time. Our guide to AI for BOPIS follows the full order journey, from picking to handover and uncollected orders.
When should retailers change this order?
The ranking fits most retailers, but three situations justify a different order:
- BOPIS is a large share of sales, and order management already shows pick status in real time. Move BOPIS exceptions up. The data gap that holds them back is already closed.
- Your stores have item-level RFID or recent cycle counts. Stock checks become even safer, and price questions can run alongside them from the start.
- Turnover is your biggest cost. Start with new-hire questions. They need the least integration of the five and are often the fastest to launch.
How does NuPlay's agent support this order?
NuPlay's agent resolves 95% of store associate queries autonomously. It connects to POS, ERP, inventory, order management and HR systems. A retailer can start with read-only lookups and add returns and BOPIS later on the same agent, without a new integration project for each step.
For decisions such as refunds, the agent cites the policy clause it applied and hands anything outside policy to a manager, with the context attached. NuLoop, NuPlay's self-improvement layer, learns from every interaction, so accuracy improves as associates use the agent. For how to roll out AI across the store floor, see our complete guide to AI for in-store associates.
FAQ
Which in-store AI use case should retailers start with?
Most should start with product location and stock checks. The question comes up more than any other, the agent only reads data, and a wrong answer costs little. It also shows whether store inventory data is accurate enough for the use cases that follow.
Why not start with the use case that has the most value?
The highest-value use cases, returns and BOPIS exceptions, need the most systems and the cleanest data, and their mistakes reach the customer or cost money. Starting with them raises the chance of an early failure that makes associates stop trusting the agent.
When is a store ready to add the next use case?
When the current one answers accurately for several weeks in a row, and associates use the agent before they ask a colleague. Adding use cases before that point spreads a shaky foundation across more of the store.
How do you launch your first in-store AI use case?
- Pick your first use case using the ranking and the exceptions above.
- Pull the 50 questions associates ask most often for that use case, from managers, the store phone and team chats.
- Check the data behind the answers: inventory accuracy for stock, current SOPs for new hires, live prices for promotions, written policy for returns.
- Set the escalation rule: which answers the agent gives, and which go to a manager.
- Add the next use case only once the first is consistently accurate.
