Retail AI Assistants at Scale: Lowe’s, Home Depot, IKEA

🇹🇷 Türkçe: Bu yazının Türkçesini oku →

Here is the question an operations director should be asking this quarter, and almost nobody is. If an AI assistant on a store already triples conversion for shoppers who use it, how much is the delay costing stores that still treat the category as a pilot? Not the ones running proofs of concept. The ones watching three winter quarters go by while the budget sits in next year’s deck.

I keep landing in the same boardroom conversation. The language is identical to 2011 about responsive design and 1998 about putting the catalogue online at all. Someone senior calls it premature. Someone else calls the vendor a hype shop. A quieter voice asks what the competition is doing, and the room looks for a slide that was not prepared.

So I went and read the earnings calls, the Google Cloud announcements and two years of attributed reporting on IKEA’s bot. Three retailers you have heard of have moved an AI assistant past pilot and into production, publicly, with figures their investor relations teams had to sign off. Those figures are the subject here. The second half is what the pattern costs on a mid-market WooCommerce estate.

Lowe’s reports shoppers who engage with its Mylow assistant convert at three times the rate of those who do not. Home Depot’s Magic Apron fields millions of questions a month across more than two thousand U.S. stores. IKEA’s Billie now resolves 74% of inquiries, and the 8,500 call-centre staff whose routine work it took over were reassigned to a remote design channel that reported €1.25 billion in revenue last fiscal year. Three different operating models for three different problems. The shape they share is what the back half of this piece is really about.

Lowe’s: The Assistant That Triples Online Conversion

Lowe’s launched Mylow in March 2025 as a customer-facing virtual advisor. Marvin R. Ellison, chairman, president and CEO, put a number on it on the Q2 FY26 earnings call of 19 August 2026: customers who engage with Mylow “convert at 3 times greater than customers that don’t,” with “roughly 25 million questions since its inception.” Online sales grew 15.7% that quarter per the SEC 8-K, against flat comparable sales elsewhere.

The second half of the story is Mylow Companion, the in-store associate version. Associate adoption is where pilots usually fail: the tool is built, staff ignore it, the shelf ticks over at the same rate. Lowe’s rolled Companion across all 1,700-plus stores and reported customer satisfaction scores lifted by 200 basis points in stores where associates were using it, per reporting on the same earnings call. The dual deployment is the lesson. A customer-side bot alone would have hit the adoption curve any search box hits, which is a fraction of the catalogue’s demand curve. The associate layer is what makes the knowledge show up at the aisle, not just on the site.

• 3x higher online conversion for Mylow users versus non-users (Marvin Ellison, Lowe’s Q2 FY26 earnings call, 19 August 2026).
• Approximately 25 million customer and associate questions handled since launch in March 2025.
• +15.7% online sales growth in Q2 FY26 against flat comparable sales elsewhere.
• Mylow Companion rolled across all 1,700-plus Lowe’s stores.

The transferable lesson is that a lift of this size is not a chat-widget outcome. It is the knowledge of the staff who already answer the question, surfaced earlier in the funnel. The lift sits in the gap between what a shopper types into a search box and what they would have asked an experienced associate on aisle 11.

Home Depot: Scale as the Point, Not the Side Effect

Home Depot’s Magic Apron started the year as a customer-facing assistant on the website. Jordan Broggi, EVP of Interconnected Retail, framed it on 11 January 2026 in the Google Cloud joint press release: Home Depot was “putting ‘Orange Apron’ expertise in the pocket of every customer” wherever they were, “the home, the jobsite, or in the aisles of our stores.” From the first announcement, the roadmap was about taking the thing in-store, not keeping it on the storefront.

Eight months later, on 27 August 2026, Home Depot announced the nationwide rollout of in-store capabilities across more than 2,000 U.S. stores, so a shopper in an unfamiliar aisle gets store-specific guidance from the same assistant they used at the kitchen table. The same announcement confirmed the running volume: Magic Apron fields “millions of questions per month,” against a store base of 2,364 and more than 470,000 associates.

• Rollout to 2,000-plus U.S. stores with localized in-store guidance (Home Depot press announcement, 27 August 2026).
• Magic Apron fielding “millions of questions per month” at that scale.
• Partnership publicly structured with Google Cloud, launched 11 January 2026.
• Store base of 2,364 retail stores and 470,000-plus associates covered.

The lesson is colder than Lowe’s. Home Depot has not published a conversion number, and I would not read that as hesitation. The point of this build is not the sale on the site. It is the time an associate loses looking for a bolt the customer could not describe. Scale the assistant to the real volume of human questions on the floor, and the economics come out of labour utilisation, not a conversion delta.

IKEA: The Bot Paid for Itself by Redeploying the People It Replaced

IKEA’s Billie has the longest production life of the three, launched in 2021 by Ingka Group. Its capability curve is published and attributed. Over the first two years it assisted 47% of shoppers who opened it, and per CX Today’s 3 August 2026 reporting of Ingka figures, that figure now sits at 74%. Customer satisfaction lifted to 89% from 60% pre-rollout. Those are large deltas on a surface where the baseline is usually a mild disappointment nobody has budget to fix.

What makes Billie the useful case for a buyer is what happened to the call-centre team it replaced. Ingka did not downsize those 8,500 roles. It retrained them as remote interior design consultants, selling by phone and video. That channel reported €1.25 billion in revenue last fiscal year, up from €1.08 billion, with 15% to 20% annual growth across three years. Ingka has committed to 10% of total revenue through this channel by 2028. The AI math and the HR math sit in the same sentence, which is unusual.

• Billie now resolves 74% of inquiries from shoppers who open the tool, up from 47% in the first two years.
• Customer satisfaction lifted to 89%, from 60% pre-rollout.
• 8,500 former call-centre staff retrained as remote design consultants.
• Remote design channel: €1.25 billion in revenue last fiscal year, up from €1.08 billion.
• Growth rate: 15% to 20% annually over the past three years.

The lesson is the structural one. Billie did not pay for itself because deflection is a line on a cost sheet. It paid for itself because the people it replaced went on to generate more revenue than the cost it took out. If your business case for an assistant is pure headcount reduction, you have modelled the smaller of two numbers. The larger one is what the newly freed people do next, and if your plan has no answer there, the project is in the wrong org.

What the Three Share, When You Line Them Up

Three retailers, three different motions. One is selling the conversion lift (Lowe’s), one is building the operational layer (Home Depot), one is reassigning the labour (IKEA). What they share, read side by side, is that none of them deployed the assistant in isolation. Each built the bot and the thing the bot was for in parallel. Mylow has Mylow Companion. Magic Apron has an in-store localisation layer with 2,000-store coverage. Billie has 8,500 retrained consultants with a revenue line of their own.

The pilots that stay pilots usually try to prove the bot works on its own first, and only then plan staffing, catalogue and floor layout around it. These three did the opposite. They built the operating model the bot slots into, treated the bot as connective tissue, and published the results of the whole thing rather than the model. A buyer evaluating a vendor should be asking what the equivalent connective tissue is in their own estate before they ask which model to use.

The second thing they share matters. All three are reporting these numbers themselves, from investor relations, press offices and leadership interviews. Each figure in this piece sits inside a filing or a statement with a name, a title and a date. That is the sourcing bar a buyer should hold their own agency to. If the figures you are shown are not attributed to a named executive at a named company with a date, they are not the figures.

What It Takes to Ship This on a Mid-Market WooCommerce Estate

If you are running a 2,000 to 20,000 SKU WooCommerce catalogue, here is the shape of work that produces a result in the family above rather than the family of pilots that quietly end. For a store with a serious catalogue, a trained merchandising team and sign-off already in hand: 10 to 14 weeks for the first production surface, 1 to 2 days per quarter to keep it honest.

The first layer is the content the assistant retrieves from. Your product data model has to carry the fields a shopper actually asks about, not the fields the import plugin happened to ship. On WooCommerce that is global attributes, custom product fields through ACF, and a disciplined taxonomy for installation guides, specification sheets and compatibility rules. If this layer is thin, the assistant will hallucinate confidently. If it is right, the assistant answers from your catalogue and you get the first lift the Lowe’s figure points at. Count 3 to 5 weeks of modelling and clean-up for a mid-market catalogue, longer if the attributes have never been systematised.

The second layer connects the catalogue to the LLM. We ship this as a dedicated WordPress plugin that reads a vector index built from your product data, help-centre articles and returns policy, pipes a question through a prompt template per intent, and returns a response grounded in links to your own pages. The engineering decision worth caring about is where the index lives. Self-hosted on managed Postgres with pgvector keeps data and cost auditable; a third-party embeddings API gets you going faster and binds you to someone else’s billing. We start on the first and open the second only for peak traffic. Count 2 to 3 weeks of implementation plus a week on evaluation harnesses.

The third layer is the escalation path to a human, and it is what decides whether this is a toy or a business tool. The Billie 74% resolution figure is impressive precisely because the other 26% land somewhere the shopper can actually get helped. On WooCommerce that means a live handoff from the chat surface into your help-desk tool, with session context attached so the human agent does not start from scratch. The attach point is boring plumbing and it is where most pilots quietly fail. Count 2 weeks, and expect to redo one integration when the help-desk vendor’s API changes mid-build.

The fourth layer is measurement. Ship without instrumenting and you will be arguing about whether it worked for a year. The minimum useful metrics for a quarterly board review are deflection rate, containment rate, assisted-conversion lift against a holdout cohort, and the one most teams skip, grounding fidelity: the share of answers citing a page in your own site rather than the model’s prior knowledge. On WooCommerce we land this as a custom GA4 event stream plus a weekly export into a BI surface the merchandising team already uses. Count 1 week of instrumentation, 2 days per quarter of review.

The fifth layer is the organisational one, and almost no agency talks about it because there is nothing to sell with it. Who owns catalogue clean-up when the assistant surfaces a wrong answer? Who decides when the escalation routes to sales rather than support? Who signs off on the retrain cadence? For a mid-market store the honest answer is one named person inside the business, half their time for the first quarter and a day a week after. If nobody has that time, delay the project a quarter and free them, rather than shipping anyway.

Add it up and you have a 10 to 14 week first production surface, roughly a mid-market replatforming engagement, with the bulk of the money in content modelling, retrieval infrastructure and escalation plumbing, not the model. That ratio is the most important number in this piece because it is the one that gets misread. The model is cheap. The business around it is not.

If you want a sharper read on what this looks like on your catalogue, the honest ask is not a demo of our stack. It is a one-week audit of the five layers above, with a written answer on which you can ship against this quarter and which need straightening out first. That is the conversation worth having before the next deck. Our contact page is here, and we return briefs within two business days.