The challenge
What was deployed
The results
Last Updated: September 25, 2026
Lowe's launched Mylow, an AI home-improvement assistant developed with OpenAI, in March 2025. On the company's 19 August 2026 earnings call, CEO Marvin Ellison told investors it had handled more than 25 million questions from customers and associates, and that shoppers who use it online convert at three times the rate of those who do not. The number that matters is conversion, not chat volume — but the comparison is between self-selected groups, and Lowe's has published no controlled test, no model details and no cost figures.
The problem: shoppers arrive with a problem, not a part number
Retail search was built for a shopper who already knows what they want. Home improvement customers frequently do not. Someone with a leaking faucet knows the symptom, not the cartridge type; someone contemplating a kitchen knows the ambition, not the order of operations. For two decades the retail answer was a keyword box and a filter tree, which assumes the customer can translate their problem into merchandising vocabulary.
Lowe's says the translation step is exactly where behaviour has shifted. A year after launching the assistant, SVP of Digital Commerce Joseph Cano described the change in how people phrase queries.
Search used to be three or four keywords you tried to really hone in on
— Joseph Cano, SVP Digital Commerce, Lowe's, CX Dive
Cano told CX Dive that customers now type full sentences describing a whole project rather than honing in on a product term. That is a retrieval problem before it is a generation problem: the index was built for nouns and the queries arrived as narratives.
The second half of the problem is physical. Lowe's says the lion's share of its business is still done in-store, so a well-answered online question only pays off if the context survives the drive to the store — and if the associate the customer finds in aisle 14 happens to know plumbing.
What Lowe's actually built
Mylow was announced on 5 March 2025. Lowe's own newsroom describes it as developed in collaboration with OpenAI, drawing on Lowe's own expert advice content, and launched first on desktop and mobile web for MyLowe's Rewards loyalty members, with voice and in-app availability flagged for later that year.
Functionally, the assistant takes an unstructured, project-shaped question and narrows it to specific, purchasable products plus the steps to use them. A customer asking how much mulch they need gets a quantity answer and a buyable recommendation. That is the commercial hinge: the output is a cart, not a page of search results the shopper still has to sort through.
How the Mylow loop is described in public sources
The customer describes a problem or project in a full sentence rather than a keyword string.
The assistant draws on Lowe's how-to content and catalogue to return guidance alongside specific items that can be purchased directly.
Project details can be stored in the customer's profile so the plan is not trapped in a chat session.
An associate pulls up that profile in-store, walks the customer through the physical products, and can offer to follow up in later weeks.
Mylow Companion lets associates ask voice or text questions about in-store items when the customer's question falls outside their department.
The associate-facing half
The part that gets least attention is arguably the more operationally interesting one. Cano described Mylow Companion as a way for a non-expert associate to handle an expert question — asking the agent what part a customer needs and being directed to the right aisle. In other words, the same retrieval layer serves two very different interfaces: an anonymous shopper on a phone and a trained employee under time pressure on a shop floor. Teams building internal knowledge agents usually treat those as separate programmes with separate budgets.
Designing for other people's agents
Lowe's also decided that the assistant is not the only AI that matters. Cano told CX Dive the company is making its sites crawlable by AI agents so that third-party assistants return accurate Lowe's information, which involves designing separate journeys for machines and for humans.
No matter where a customer is searching, we want to make sure that we show up in the right way
— Joseph Cano, SVP Digital Commerce, Lowe's, CX Dive
That is a distribution decision dressed as an SEO decision, and it is the part most retailers are still arguing about internally.
The measured outcome
The figures that made this deployment investor-grade came from the second-quarter earnings call on 19 August 2026.
Ellison put the conversion number to investors plainly:
Online customers who use Mylow are three times as likely to convert as customers who do not use the tool
— Marvin Ellison, Chairman and CEO, Lowe's, PYMNTS
He framed the result as evidence that a well-designed agentic AI experience can be a clear driver in the purchasing decision. Cano, describing the product logic to OpenAI, put it more bluntly: most people arrive with a problem rather than a product in mind, and the tool figures out what they mean, and helps them finish the job.
Read the 3x carefully
Treating a users-versus-non-users gap as a causal lift
Customers who open an AI assistant are, on average, further down a project and more committed than customers who bounce off a category page. Some portion of any conversion gap is that selection effect, not the assistant. This is the single most common way AI business cases overstate themselves — and it is unfalsifiable once the feature is live for everyone.
What the public record does not tell us
Being candid about the gaps is more useful than padding the win. On the evidence available as of September 2026:
| Question an integration team would ask | What the public record says |
|---|---|
| Which models power Mylow? | Lowe's says it was developed in collaboration with OpenAI. No model family, version or routing strategy has been disclosed. |
| What does a conversation cost? | Not published. No inference cost, latency target or concurrency figure appears in any source reviewed. |
| How accurate is it? | No hallucination rate, grounding evaluation, escalation rate or human-review process has been published. |
| Is the 3x causal? | Described as a comparison between users and non-users. No A/B test, holdout or incrementality study is referenced. |
| What did Mylow Companion do for stores? | Described qualitatively by Lowe's executives. No published handle-time, productivity or attach-rate figures. |
| How are recommended products ranked? | Not disclosed. Whether merchandising, margin or inventory position influences recommendations is unaddressed publicly. |
The 15.7% online growth figure deserves the same caution. Ellison credited it to tailored digital experiences, expanded visualisation tools, marketplace growth and new fulfilment options alongside Mylow's contribution — a multi-driver attribution that no single feature owns.
Timeline
Lowe's announces Mylow, developed with OpenAI, on desktop and mobile web for MyLowe's Rewards members, with voice and app availability planned for later in the year.
Roughly a year in, Lowe's describes the behavioural shift to sentence-length queries, the profile-to-store handoff, the Mylow Companion associate app, and a deliberate effort to make its content readable by third-party AI agents.
On the Q2 earnings call, the CEO reports more than 25 million questions handled and a 3x conversion differential for online Mylow users; online sales grow 15.7% year over year.
What an integration team should take from this
Four transferable lessons, in rough order of how hard they are to copy.
1. Instrument against a number your CFO already reports. Mylow is discussed in terms of conversion, not containment, CSAT or deflection. That is why it appears in an earnings call rather than a slide deck. If your assistant cannot be joined to revenue, order rate or cost per resolved case at the record level, that join is the first engineering task — not the last.
2. Build the handoff, not just the answer. The project profile that an in-store associate can retrieve is the least glamorous component and probably the highest-leverage one. An assistant that ends its usefulness when the tab closes captures a fraction of the value of one whose output becomes durable state in the customer record. Most of the deployments worth studying share this property.
3. Serve the employee with the same retrieval layer. Mylow Companion reuses the knowledge substrate for a different audience with different latency and trust requirements. That is an architecture decision — one index, multiple surfaces — and it is far cheaper than running two programmes. It also creates a free evaluation channel: associates notice wrong answers faster than customers report them.
4. Assume other agents will shop on your behalf. Making a catalogue machine-readable for third-party assistants is a hedge against losing the customer relationship to whichever agent the shopper happens to use. Retailers who wait until agentic shopping is mainstream will be negotiating from a weaker position. Our ongoing research on agent-readable commerce surfaces suggests this is where the next round of differentiation lands.
The honest scoreboard
Lowe's has published a volume number, a directional commercial number and a set of credible executive descriptions of the architecture. It has not published an experiment, a cost structure or an accuracy measurement. That is more disclosure than most retailers offer and less than an engineering team needs to plan against. Treat the 3x as a reason to run your own test, not as a benchmark to promise your board.
If you are scoping something comparable — conversational discovery, an associate knowledge agent, or an agent-readable product surface — the sequencing matters more than the model choice. Decide the measurable outcome and the holdout before writing a prompt. We cover that sequencing in our engagement approach and our notes on model selection, and you can start a scoping conversation if you want a second opinion on the test design.
Frequently Asked Questions
What is Lowe's Mylow?
Mylow is Lowe's AI-powered virtual home improvement assistant, announced in March 2025 and developed in collaboration with OpenAI. It answers project and how-to questions in conversational language and returns specific, purchasable products alongside the guidance. It launched on desktop and mobile web for MyLowe's Rewards members, with voice and in-app availability signalled for later in the year, and is paired with an associate-facing app called Mylow Companion.
How many questions has Mylow answered?
More than 25 million questions from customers and store associates since launch, according to CEO Marvin Ellison on Lowe's second-quarter earnings call of 19 August 2026, as reported by PYMNTS. That figure combines the customer-facing assistant and the associate-facing Companion app, so it should not be read as 25 million shopping sessions.
Does Mylow actually increase sales?
Lowe's says online customers who use Mylow are three times as likely to convert as those who do not. That is a comparison between two self-selected groups, not a controlled experiment, and shoppers who open an assistant may already be closer to purchase. Lowe's has not published an A/B test or incrementality study, and the quarter's 15.7% online growth was credited to several initiatives, not to Mylow alone.
Which AI models does Mylow use?
Lowe's states only that Mylow was developed in collaboration with OpenAI and draws on the company's own expert advice content. No model family, version, routing approach, latency target or per-conversation cost has been made public. Anyone benchmarking against this deployment should treat the technical stack as undisclosed rather than inferring it from the vendor relationship.
What is Mylow Companion?
It is the associate-facing side of the same capability. Store employees use voice or text to ask questions about items in store, which helps a worker who is not the resident expert in, say, plumbing or electrical answer a customer question and point them to the right aisle. Lowe's has described it qualitatively; it has not published handle-time or productivity figures for it.
Are other retailers doing the same thing?
Yes. Home Depot's Magic Apron assistant, built with Google Cloud's Gemini models, follows a similar pattern of conversational, project-specific advice that recommends products, indicates where to find them in store and suggests what else a project needs. The competitive question is shifting from who has an assistant to whose assistant reliably ends in a completed, correctly specified cart.
Twarx analysis
Original interpretationThe interesting metric is not the question count — it is that Lowe's measured the assistant against conversion, a number the CFO already cares about, rather than against deflection or satisfaction scores. That choice is what got the deployment onto an earnings call. The weakness is the same choice: a users-versus-non-users comparison is a self-selection artefact until someone runs a holdout.
Read the numbers honestly
Analysis by
Rushil Shah · AI Systems Builder & Founder, Twarx
