Picture yourself standing in the plumbing aisle of a hardware store, holding a corroded elbow joint you unscrewed from under your sink an hour ago. You don’t know the thread size. You don’t know if it’s PVC, ABS, or galvanized steel. There are maybe a hundred and twelve near-identical fittings between you and the checkout, and about twenty minutes before the store closes. Somewhere on that wall is exactly the part you need. Whether you find it depends entirely on the next person who talks to you: someone who actually knows the difference between a compression fitting and a slip joint, or a well-meaning employee three weeks into the job who points vaguely toward “plumbing” and wishes you luck.
That gap, between “the answer exists somewhere in this store” and “someone can actually get it into your hands,” isn’t really a hardware store problem. It’s an information problem. And it’s the exact one Lowe’s spent the better part of three years, and a partnership with OpenAI, trying to close with an AI assistant it calls Mylow.
What Problem Was Lowe’s Actually Trying to Solve?
Home improvement retail has a structural mismatch that most other retail categories don’t. A grocery shopper mostly knows what they came for. A Lowe’s shopper often doesn’t, not really. They know their kitchen faucet is dripping, or that they want to “redo the deck,” and from there they’re translating a household problem into a shopping list they don’t have the vocabulary for. Chandu Nair, Lowe’s SVP of data, AI, and innovation, has described this directly as “a large information gap” on the buyer’s side, and it’s the same gap the confused customer in the plumbing aisle is standing in.
Search boxes don’t close that gap, because search is a keyword match, not a diagnosis. And associates, however good, don’t scale: Lowe’s runs more than 1,700 stores, staffed in large part by people who rotate departments, work seasonally, or are new to the trade themselves. A customer’s odds of getting a genuinely knowledgeable answer shouldn’t depend on which associate happens to be nearby when they ask. That’s the problem Lowe’s set out to fix, first for customers, then for the associates themselves.
What Did Lowe’s Build, and How Does It Actually Work?
Lowe’s answer, announced in March 2025, was Mylow: a generative AI assistant built on OpenAI’s models (GPT-4o), reachable through the Lowe’s website and app by voice or chat. A customer can ask something as loose as “how do I remove carpet before installing tile,” and Mylow doesn’t just return a product list. It uses retrieval-augmented generation (RAG) to pull from Lowe’s own product catalog, project content, and real-time store inventory, then hands back a project plan: steps, links to how-to guides and videos, and the specific products needed at each stage, ideally in stock at the customer’s local store. NVIDIA NeMo Guardrails sit on top to keep responses on-topic and prevent the assistant from wandering into subjects that have nothing to do with home improvement.
The gap Lowe’s was closing was never really a search problem. It was a translation problem, between what a customer knows and what a plumber knows.
In May 2025, Lowe’s extended the same underlying model to a second product: Mylow Companion, built for associates themselves and rolled out to sales-floor devices across all 1,700-plus stores, which Lowe’s says makes it the first at-scale generative AI deployment of its kind in retail. Companion is a knowledge tool, not a customer-facing chatbot: an associate working outside their usual department, or new to the job entirely, can ask it a product or project question on the spot instead of guessing or paging a specialist.
Underneath both products, Lowe’s has also been building a parallel visual layer with NVIDIA: an interactive “digital twin” of physical stores, rendered in NVIDIA Omniverse and fed by DeepStream computer vision at the edge, currently piloted in two locations. It’s used less for chat and more for things like heat-mapping customer traffic, testing product placement changes virtually before moving a single shelf, and letting associates check what’s on a hard-to-reach top shelf via camera instead of a ladder. It’s a separate initiative from Mylow, but it’s built by the same team, on the same underlying premise: close the gap between what’s true in the store and what a person, customer or associate, can actually see or ask about.
| Before Mylow | With Mylow / Mylow Companion | |
|---|---|---|
| Finding a project answer | Keyword search or whichever associate is nearby | RAG-grounded assistant with real-time inventory + project guides |
| Associate product knowledge | Bounded by tenure and department | On-demand lookup across the full catalog, any department |
| Customer conversion (online) | Baseline site conversion | Roughly 2x higher among Mylow users, per Lowe’s |
| In-store CSAT | Baseline | +200 basis points where Companion is used, per Lowe’s |
| Monthly usage | N/A | 1M+ customer questions/month by early 2026 |
Did It Work, and Who’s Saying So?
Here’s where it’s worth slowing down, because most of these numbers come from Lowe’s itself, not an independent audit. By Lowe’s own reporting, Mylow was fielding over 1 million customer questions a month as of its Q1 2026 earnings update, and customers who engage with Mylow convert online at roughly double the rate of those who don’t. In stores, Lowe’s says Mylow Companion usage correlates with a 200-basis-point lift in customer satisfaction scores. Company-wide digital sales rose about 6% in Q1 2026, which Lowe’s attributes in part, though not exclusively, to AI-driven shopping tools alongside broader omnichannel investment; Lowe’s put roughly $521 million into capital expenditure that quarter, a meaningful share of it toward AI infrastructure and digital platforms.
None of that is independently verified by a third party the way, say, an SEC filing’s revenue line is. Conversion lift and CSAT delta are self-reported metrics from a company with every incentive to report Mylow favorably, and Lowe’s hasn’t published the underlying methodology (what counts as a “Mylow user” for the conversion comparison, for instance). What is independently checkable is the deployment itself: press coverage, the OpenAI partnership announcement, and trade outlets confirm Mylow and Companion are live at the scale Lowe’s describes, which is not nothing. A chatbot answering a million real questions a month, at 1,700 stores, for over a year, is a genuinely different claim than a pilot program in three test locations.
You might be reasonably skeptical here: isn’t this just a well-marketed chatbot with a home-improvement skin? Worth naming that directly, because it’s the obvious pushback. The honest answer is: partly. The RAG-plus-guardrails architecture isn’t novel; plenty of retailers have shipped something structurally similar. What’s less common is pairing it with a second, associate-facing version running the same underlying knowledge base, and treating the store floor and the website as one continuous information problem rather than two separate projects. That pairing, not the chatbot itself, is the more interesting part of what Lowe’s built.
What Went Wrong Along the Way?
Lowe’s has been unusually candid about this, at least by retail standards. Early in the Companion rollout, engineers found associates struggling with the voice-to-text feature specifically, garbled input on a noisy sales floor is a real failure mode, and it had to be reworked. More structurally, Nair has talked about “several hard learnings” and now frames most of Lowe’s generative AI work as roughly 70% change management and 30% technology; the model is the easy part, getting a workflow to actually change around it is the hard part. (In my experience covering enterprise AI rollouts, that ratio is almost always closer to the truth than whatever number the vendor pitch deck used.)
There’s also an internal governance problem Lowe’s has talked about candidly: what Nair calls “AI sprawl,” any team spinning up its own narrow bot until you’ve got a pile of half-working, unmaintained agents nobody owns. Lowe’s response was a formal governance framework scoring new AI proposals on four criteria: real ROI, capital investment required, risk, and whether the team is actually willing to redesign the underlying business process, not just bolt AI onto the existing one. That last criterion is doing a lot of work; it’s an admission that most failed AI pilots don’t fail because the model is bad, they fail because nobody changed the process the model was supposed to improve. Lowe’s also deliberately avoided betting on a single AI vendor, working with OpenAI, Google Cloud, and NVIDIA across different parts of the stack rather than one exclusive platform, which limits lock-in but also means more integration surface to maintain.
AI Suite and Open-Source Implementation
Lowe’s actual stack, as far as it’s been made public, layers OpenAI’s GPT-4o as the reasoning engine, retrieval-augmented generation pulling live from Lowe’s product catalog and inventory systems, and NVIDIA NeMo Guardrails as a topic and safety filter sitting between the model and the customer. Separately, the digital twin work runs on NVIDIA Omniverse for the 3D store rendering and DeepStream for real-time computer vision at the edge, with Google Cloud handling broader analytics work outside the assistant products specifically. None of these pieces are exotic; what makes the deployment notable is that Lowe’s connected inventory, catalog, and project content into one retrieval layer that both the customer-facing and associate-facing tools draw from, rather than building two disconnected products.
An enterprise without Lowe’s budget can replicate the shape of this, if not the scale, almost entirely with open-source tooling. The RAG layer maps cleanly onto LlamaIndex or LangChain for orchestration, paired with an open vector store like Qdrant, Weaviate, or Milvus to index product catalogs and support docs. NeMo Guardrails itself is open source (NVIDIA released it as an Apache-licensed project), so the same topic-boundary approach Lowe’s uses is directly available without a NVIDIA contract. For the model layer, an open-weight model like Llama 3 or Mistral, fine-tuned or prompted against the same retrieval index, gets a smaller company most of the way to what GPT-4o is doing here, at meaningfully lower cost, with the tradeoff being more in-house tuning work to hit the same answer quality.
A practical three-step path: first, before touching a model at all, get your product and support content into one clean, queryable source of truth, RAG is only as good as what it’s retrieving from, and most companies underestimate how messy their internal content actually is. Second, build the retrieval and guardrail layer with the open tools above and test it internally, against your own staff, before any customer sees it; this is exactly the phase where Lowe’s caught its voice-to-text problem. Third, decide explicitly whether you’re building a customer tool, a staff tool, or both off the same index, because Lowe’s evidence suggests the staff-facing version is where the operational payoff (fewer escalations, faster onboarding) is easiest to measure honestly. Whatever you build, deploy it in shadow mode first, running silently alongside existing workflows so you can compare its answers against real outcomes, and validate before it acts on anything customer-facing. Open-source tooling can replicate Lowe’s architecture; it can’t replicate the years of catalog and inventory data hygiene, or the governance discipline, that make the architecture trustworthy at 1,700 stores.
FAQ
Q: Is Mylow built entirely by OpenAI, or by Lowe’s?
A: It’s a partnership. OpenAI’s GPT-4o models provide the underlying reasoning, but Lowe’s built the retrieval layer connecting the model to its own product catalog, inventory, and project content, along with the guardrails and integration into its website, app, and associate devices.
Q: Does Mylow replace Lowe’s store associates?
A: No, and Lowe’s has been explicit it isn’t trying to. Mylow Companion is framed as a knowledge tool that helps associates answer questions outside their usual department or experience level faster, not a replacement for them. Lowe’s leadership has repeatedly emphasized that human judgment still matters most for complex or in-person project advice.
Q: What’s the biggest limitation of Lowe’s AI rollout so far?
A: The clearest one is that Lowe’s own headline metrics (conversion lift, satisfaction score gains) are self-reported without published methodology, so they should be read as directional, not independently audited. The company has also been candid that early associate-facing issues, like voice-to-text reliability, and internal “AI sprawl” from ungoverned tool proliferation, were real problems it had to actively manage.
Further reading: