Personalize each store before you personalize each shopper.

01

Agentic personalization works at three levels: store, moment, person

Agentic personalization is retail personalization done by AI agents. An agent reads what is true about a store and its customers, decides what to change, makes the content or offer, puts it out across channels and checks what happened. A person approves anything that goes public, spends money, changes a price or messages a customer. Older tools suggest a segment and wait. Agents run the whole loop, store by store, every day.

For a chain, personalization happens at three levels. The store: its neighborhood, the people who live and work nearby, its calendar and the rival across the road. The moment: weather, local events, the hour of the day. The person: the regulars who gave you their number and agreed to hear from you. Each level needs different data, carries different risk and pays back on a different clock, so the order you build them in matters.

Take a 40-store pet supply chain. The store near the university sells small bags and cat litter to renters and stays busy until 9. The suburban store sells 40-pound sacks to families with big dogs and yards, and its Saturday grooming books out by Wednesday. Today both get the same national calendar, the same post and the same coupon. Agentic personalization starts by letting each of them become the store its street needs.

02

Segment personalization paid off online, then stalled at the store door

Classic personalization works. McKinsey finds it most often drives a 10 to 15 percent revenue lift, and that 71 percent of consumers expect it while 76 percent get frustrated when it is missing ¹. BCG reports that retail personalization leaders grow revenue 10 percentage points faster than laggards, and that personalized offers can return as much as three times what mass promotions do ². Yet retailers put less than 5 percent of their promotional spending into personalized offers ².

A segment is a compromise between relevance and the cost of making one more version.

The reason is cost. BCG's leaders spend $10 million to $40 million a year on the technology and talent behind it ². That buys data scientists, a content studio and a testing program, all built for the website, the app and the inbox. A segment is a compromise between relevance and the cost of making one more version. Ten segments means ten emails to write, approve and test. A thousand means a team a regional chain cannot afford.

So for most chains, personalization stops at the store door. The email knows a customer's last order. The store she walks into has the same window, the same offer and the same Google post as 39 others. Where hyper personalization does run, shoppers are not always grateful. Gartner found personalized messages and offers created negative experiences for 53 percent of customers, who were 3.2 times more likely to regret a purchase ³. Twilio found only 45 percent of consumers feel understood by the brands they use ⁴.

03

Start with the store, because most of retail still happens in one

E-commerce was 17.1 percent of US retail sales in the second quarter of 2026 ⁵. More than four of every five retail dollars are spent somewhere else, mostly at places with a street address. Yet the store is the least personalized thing a chain runs. Budgets go to the flagship or get averaged into regional and national plans. Every store gets the same calendar, whether it sits by a campus, a hospital or a ballpark.

The store is the least personalized thing a chain runs.

The evidence for tailoring the store is old and consistent. Bain found that fitting assortments to local tastes can lift same-store sales by 1 to 3 percentage points, and that only 10 to 15 percent of the inventory needs to change ⁶. Competitors are part of a store's context too. In a randomized field experiment, promotions sent to shoppers near a rival's location earned more as discounts deepened, while the same offers near the retailer's own location saturated and cannibalized sales it would have made anyway ⁷.

Store-level personalization is also the safest kind. It needs no personal data. Nobody feels watched when the Elm Street store stocks what Elm Street buys, keeps the hours its street keeps and describes itself in words its neighbors use. It is also what AI assistants can see. Ocally's founder, Shreyas Kulkarni, grew up in a family of local business owners and holds a simple belief: local businesses are the heart of the neighborhood, and they can't be reduced to optimized, cold, templatized places. Each location has to belong to its street and serve the people there.

04

The moment is free data that most chains ignore

Weather, local events and the hour of the day are public, cheap and predictive. A study of 673 brick-and-mortar stores found that daily weather moved sales by up to 23.1 percent depending on the store's location, and by up to 40.7 percent depending on what was being sold ⁸. Adding the weather forecast improved sales predictions up to seven days ahead ⁸. Two stores 30 miles apart can need opposite things on the same Saturday.

Weather also changes how people respond to an offer. In a field experiment with over six million mobile users in China, responses to promotions were 1.2 times higher and 73 percent faster in sunny weather than in cloudy weather, and lower and slower in rain ⁹. The wording that worked best changed with the weather as well ⁹. A single national message, written two weeks ahead and sent everywhere at 10 a.m., cannot know any of that.

Events and hours work the same way. A graduation, a home game, a street fair or the first cold week each changes who walks past a store and what they need. The quiet Tuesday afternoon is a different store from the Saturday rush. A district manager with eight stores cannot read eight forecasts and eight event calendars every morning and rewrite the plan. An agent can, for 40 stores, before the doors open.

05

The person: fewer messages, sent to people who asked for them

Person-level personalization is where most software budgets go and where most of the risk sits. It only works if you know who is buying. BCG notes that Sephora ties 95 percent of its transactions, online and in its hundreds of stores, to loyalty accounts ². Most regional chains are far from that. Their first-party data sits in the POS, a loyalty app and an email tool, full of duplicates, with no clear record of who agreed to what.

Use what a customer told you, for the reason they told you.

How you use the data matters as much as having it. Researchers found shoppers accept personalization built on information they gave directly, on the same site. They object when it is inferred, or gathered somewhere else, and revealing those flows makes ads less effective ¹⁰. On a trusted loyalty program site, the same openness made ads work better ¹⁰. The rule for a chain is short: use what a customer told you, for the reason they told you.

Shoppers are wary. Two thirds of Americans (67 percent) say they understand little or nothing about what companies do with their data, and 70 percent have little or no trust in companies to use AI responsibly ¹¹. So keep the person level narrow and useful: the size that came back in stock, the reward about to expire, the regular who has not been in for 90 days. Each is a message a customer is glad to get.

06

Agents make a new version almost free, so knowledge becomes the bottleneck

The cost of producing a variant is collapsing, and that changes the economics of AI personalization. Stanford's AI Index reports that the cost of running a model at GPT-3.5's level fell more than 280-fold between November 2022 and October 2024 ¹². Epoch AI finds the price of a given level of performance falling between 9 and 900 times a year, depending on the task ¹³. Testing is being automated too. Braze agreed to pay $325 million for OfferFit, whose reinforcement-learning agents experiment on their own to replace manual A/B tests ¹⁴.

Do the arithmetic for one chain. Forty stores, 52 weeks and four channels (store page, map profile, social post, in-store sign) make 8,320 pieces a year, before a single message to a single customer. No team writes that by hand, which is why chains send one version. An agent can write all 8,320. The real question is whether they are right.

An agent that knows nothing about Elm Street will write 40 polished versions of the same generic post. So the work that matters moves upstream, into knowledge: what each store sells, who lives nearby, what is on the local calendar, what the rival charges, what the brand never says. Every fact needs a source and a date, or the agent will fill the gap with a guess. Downstream, the work moves into judgment: approvals, guardrails and measuring each store against its own starting point.

That shifts labor. The brand team writes rules and examples, and stops writing every post. District managers become editors who approve a week of local work in minutes. Store managers and staff supply what no model knows, like the school fair or the regular who asked about a breed. Testing stops being a quarterly project, because every store becomes its own experiment.

07

Consent is the license, and the rules tighten every year

Twenty-four states now have comprehensive consumer privacy laws, after Alabama, Louisiana, Oklahoma and Vermont joined in 2026 ¹⁵. Most give consumers the right to opt out of targeted advertising and the sale of their data, and most require affirmative permission before a business uses sensitive data ¹⁵. California's rules on automated decision-making took effect in January 2026, covering significant decisions such as credit and jobs ¹⁶. Offers are not on that list yet.

Texts carry their own risk. Federal telemarketing rules require prior express written consent before a business uses an autodialer to send advertising messages to a mobile number ¹⁷, and the law lets each recipient recover $500 per violation, up to three times that if the violation was willful or knowing ¹⁸. Multiply by a list of 20,000 regulars and one careless send becomes a problem for the whole company, which is why consent should be checked again right before every message goes out.

Prices are the sharpest edge. The FTC's study of surveillance pricing found intermediaries working with at least 250 clients, including grocery, apparel, health and beauty and convenience stores. It described tools that could show a shopper profiled as a new parent higher-priced baby thermometers based on zip code and time of purchase ¹⁹. Since November 2025, New York requires a notice reading THIS PRICE WAS SET BY AN ALGORITHM USING YOUR PERSONAL DATA, with penalties up to $1,000 per violation ²⁰.

Our view on personalized pricing is simple. Price by store, for reasons a shopper would accept: local costs, local competition, a slow afternoon. Keep prices the same for every person who walks into that store. The legal exposure of pricing by person is growing, and the trust cost lands on every store in the chain, including the ones that did nothing wrong.

08

Discounts, drift and sameness are the quieter ways it fails

The easiest thing for a personalization engine to learn is that discounts work. They do, until customers learn too. A study of more than eight years of shopper data found consumers became more price and promotion sensitive over time as promotions increased ²¹. J.C. Penney showed how hard the habit is to unwind. In 2012 it discontinued its former promotional strategy, its comparable store sales fell 25.2 percent, and it re-introduced some promotions ²². An agent rewarded on response rate will find the coupon every time.

An agent rewarded on response rate will find the coupon every time.

Drift is the second failure. An agent writing for 40 stores will eventually say something the brand never would, or promise something a store cannot honor. Tribunals have little patience for blaming the software. When Air Canada argued its chatbot was responsible for its own words, a Canadian tribunal held the airline responsible, noting that it makes no difference whether information comes from a static page or a chatbot ²³. Whatever an agent publishes in a store's name, the chain owns.

Sameness is the failure nobody plans for. In a controlled experiment, writers given AI ideas produced stories rated more creative, but the stories were more similar to each other ²⁴. Researchers have also shown that when many firms rely on the same algorithm, decisions can get worse overall, even when that algorithm is more accurate ²⁵. If every chain runs the same models on the same prompts, every Saturday post converges. The defense is what no rival's model has: your stores' facts, photos and staff.

09

Shoppers now bring their own agents, and those agents personalize for them

The next shift comes from the other side of the counter. Google has said AI Mode will draw on past searches and, if a shopper connects it, Gmail for personal context ²⁶. Since November 2025, its Let Google Call feature phones nearby stores to ask whether they have an item, what it costs and whether there are promotions ²⁷. In January 2026 Google announced Direct Offers, a way to present a discount inside AI Mode to a shopper ready to buy ²⁸. In March its Universal Commerce Protocol added identity linking, so shoppers can keep loyalty and member pricing when an agent buys for them ²⁹.

Usage is real and growing. Adobe reported AI traffic to US retail sites up 393 percent year over year in the first quarter of 2026, converting 42 percent better than other traffic by March ³⁰. Checkout is slower to move. OpenAI scaled back its in-chat Instant Checkout in March 2026 to focus on product discovery ³¹. Discovery is where personalization lives, and it is moving into the shopper's own AI shopping assistant.

This changes what a store must do. A shopper's agent may already know her size, her budget and her dog's breed, so your guess at her segment adds little. What the agent cannot know is whether your Elm Street store has the 30-pound bag today, opens at 8 and runs puppy classes on Thursdays. Personalization shifts from persuading a person to being the right answer for her agent. Current facts per store, and a local identity an agent can describe in words, win that choice.

Two second-order effects follow, in our view. Offers become data that agents compare side by side, so a private discount sent to one segment is likely to surface for everyone, and margin leaks with it. Loyalty has to work through agents, so a regular keeps her member price when her assistant orders for her. Chains that sort out store facts and loyalty identity now will be easier for agents to recommend.

10

By 2030 the store is the unit. Start with five stores this year

Our outlook, and it is opinion. Within two years, chains with good store data will run a weekly plan per store, written by agents and approved by district managers, and the national calendar will become a template each store adapts. Within three to five years, shopper agents will handle a large share of local discovery, asking stores questions where shoppers used to click. Person-level offers will shrink to consented service messages as more states follow New York on pricing ²⁰.

We expect the winners to be the chains whose stores are easiest to describe accurately: what is on the shelf, when it is open, who it serves, what is happening there this week. Customer data volume will matter less than store truth. That is good news for a 40-store regional chain. Its stores are already different from each other. The job is to write that difference down and let agents act on it, with people checking the work.

Give the work an owner in operations, alongside the marketing team, because most of it happens in stores: shelves, hours, signs and staff. Budget for the people who review an agent's work as well as for the software. And stop one thing now: the national coupon blast that teaches your best regulars to wait for the next one.