Brent Shiely

The problem

As the product manager for inventory, payments and fulfillment on Best Buy's digital channel, I tracked how inventory turnover flows through to earnings per share, and the numbers bothered me. Units sat in stores and warehouses, hidden from customers for good reasons: safety-stock floors, third-party fulfillment rules, clearance status. Each rule made sense on its own. Together they stranded inventory that online customers wanted to buy. Returned Open Box units shipped to one central warehouse and lost value every week they waited to be graded and resold.

What I saw

Customers can't buy what they can't see.

Every store was already a warehouse, paid for and close to customers. What was missing was the sourcing logic to make that inventory visible and sellable.

What I built

I led the Ship-from-Store pilot, starting with 50 underperforming stores, and then iterated on sourcing rules one SKU at a time:

• Third-party SKUs sitting as returns on store shelves: check on-hand store inventory first. Nearly all of it shipped within a week.
• Clearance SKUs: remove the safety-stock floors and sell down to the last unit. If one store can't pick it, the next store does.
• Items dropped from a single store's planogram: ship them out the back door instead of letting them sit.
• Open Box: enable it for Ship-from-Store and in-store pickup, so stores could sell returns directly.

The sourcing logic, which picks the fulfillment location by transit time and cost, is patented: U.S. Patent 10,636,079, Demand-Based Product Sourcing (2020).

The result

• Scaled from 400 to 1,417 stores in one year, all 1,050 Best Buy big-box stores plus 367 Best Buy Mobile stores, making the entire U.S. chain a fulfillment network.
• Best Buy credited ship-from-store, along with digital marketing, for a 29% rise in comparable online sales in the first quarter of fiscal 2015. In the third quarter, it credited the chain-wide rollout for a 21.6% rise.
• Store pickup grew from under 15% of online sales when I started to 42% by the 2019 holiday quarter, as reported on Best Buy's earnings call. In fiscal 2025, pickup was about 45% of online revenue.
• Open Box recovery went from under 40 cents on the dollar to over 90. In the same period, Best Buy named returns, replacements and damages among the drivers of its Renew Blue cost savings.

What I learned

The biggest opportunities live in the gaps between products and between teams. Stranded inventory wasn't a store problem or a digital problem; it sat between the two, waiting for logic.

The problem

In June 2014, a Best Buy shopper browsing laptops saw prices, specs and reviews, but not whether they could get the item or when. Availability hid behind a "Check Shipping & Availability" link. The lightbox it opened said only that an item "usually leaves our warehouse in 1 business day." The arrival date took another click, and store pickup took a third. Vendor-shipped items carried a flat 10-day promise, however fast the vendor actually shipped.

What I saw

A vague date costs conversion. Shoppers comparing options choose the date they can see and believe. The data to give every shopper a real date already existed: where the units were, where the customer was, and how long carriers took between the two.

What I built

• Zone-based shipping promise: carrier time-in-transit data from each origin, plus the customer's ZIP code, plus real-time available units, produced a personalized promise starting at 3 days.
• Vendor-aware promises: for vendor-shipped items, the promise used where each vendor's warehouses actually were, the units in each one, and that vendor's real processing time. Vendors shipped slower than Best Buy, but still faster than the flat 10-day promise implied, so shoppers got an honest, earlier date.
• A date for every visitor: when a shopper hadn't entered a ZIP code, we estimated their location and showed a personalized date anyway. A date beat no date.
• Proven at checkout, then moved up the funnel: we launched in checkout, where the conversion signal was clearest. Once the lift showed up, we expanded step by step to cart, the product detail page, the product listing page, and finally search, aligning two separate search teams. Each shopper saw their own promise, consistent at every step from search to checkout.

The result

• Shoppers shown a personalized promise converted 40%+ higher than those without one.
• Vendor-shipped items grew from about 10% to about 15% of ship-to-home units.

What I learned

A delivery date is a product, not a disclaimer. Prove value where the signal is strongest, then let the results earn the next step.

The problem

When I joined, Best Buy stored product availability in the item master: the system that describes what a product is, not where it sits or how fast it can reach someone. A product record can't know which stores have units, or how quickly a customer in a given ZIP code can get one. Order management built its own availability picture from each warehouse's feed, so the website and order management often gave different answers. Fulfillment and inventory were one undifferentiated domain.

What I saw

Every button state is personal. "Add to Cart," "Pickup Only," "Pre-order" and "Sold Out" should all depend on who is asking. If no store or warehouse can serve a customer for that product and condition, new or Open Box, it's out of stock for that person, even while it's in stock for someone 50 miles away. Getting there meant one definition of availability that every system trusted.

What I built

• Split the domain: separated Fulfillment and Inventory into two product domains with their own teams and roadmaps.
• A real-time inventory service: I owned the service that held every store's and warehouse's sellable units, net of pickup and ship-from-store thresholds, kept within 3 minutes of real time. Order management retired its per-warehouse feeds and consumed this service directly: one definition, used everywhere.
• A fulfillment service on top of it: using that inventory data, it returned each customer's real options by product, store and ZIP code in under 500 milliseconds. Store pickup became a decision based on stock near each customer, not a global per-product flag.
• Button State: a shared component every page used, so the add-to-cart button became a personalized decision for each customer, product and condition: Add to Cart, Pickup Only, Pre-order, home delivery, delivery plus installation, download, or Sold Out. Sold Out now meant no location could serve that customer, not that the item was gone nationwide.
• Large products: major appliances and TVs over 50 inches can't ship by parcel. Each follows its own path: a regional delivery warehouse, a local self-delivery store, or a transfer to the servicing location first.
• Delivery and installation dates online: the scheduling system behind delivery couldn't handle web traffic, so hundreds of times a week a customer reached checkout and was told to try again later. I sized the cost at about $21 per error, six figures a year, and built the business case for a web-scale service calendar. Another team built it; I integrated it so every customer saw real delivery and installation dates, and later could search and filter by them, built for the shopper whose refrigerator just died. Two U.S. patents cover this work: 10,776,848, Large Product Presourcing Search Engine, and 11,593,751, Pre-coordinating Delivery and Service Information for Item Searching and Filtering.
• Conversion features on the same data: "Ready in 1 hour," inventory held only at checkout, and countdown timers on the last available unit.

The result

• Store pickup grew from 15% to 20% of digital units in the first phase.
• 100% of units became sellable. Products stocked unevenly, in a handful of stores rather than everywhere, stayed listed online down to the last unit in the last store, as Pickup Only if that was the only option.
• Helped make Best Buy a credible online channel for major appliances and large-screen home theater, a major contributor as Best Buy rose to No. 3 among U.S. major appliance retailers in 2017.

What I learned

Before AI or analytics can help, the business needs one definition of the truth. Reconciling it across systems is unglamorous work, and it's what everything else stands on.

The problem

Every shopper saw the same deals in the same order. A gamer, a new homeowner and a My Best Buy Total member who had just bought a TV all got one list, whatever they owned, browsed or paid for.

What I saw

A deal only counts if you want the thing.

Best Buy knew a great deal about its members: what they bought, what they browsed, which membership they held. The deals page used none of it. The gap was between what Best Buy knew about a customer and what it showed them.

What I built

I led Personalization Services, the product team behind it.

• Top Deals for each member: on Google Cloud, we took in the deals, the members, their purchase history and their click behavior (screen views, add to carts, saves) and fed them to a machine learning model that ranked the deals each member was most likely to want, on the app and the web. The model retrained continuously, so every view, add to cart and save made the next ranking better.
• Deals in the app: a recommendation engine that put the most relevant deals in front of paid members.
• Better customer data first: personalization is only as good as the records behind it. Customer data enrichment filled the gaps, taking matched records from 81% to 89% of a database of more than 200 million.

The result

• Top Deals ran on the Best Buy home page. In August 2024, Best Buy's CEO told investors that testing showed customers receiving the personalized app homepage engaged with content, products and tools almost 70% more than customers who didn't.
• The app's deals engine lifted paid-member conversion 38% and app engagement more than 10%.
• Enrichment added about $20 million a year in gross profit; enriched customers generated about four times the revenue of the rest.

What I learned

Personalization isn't showing more. It's showing fewer, better things. Shoppers don't want a bigger catalog; they want one that knows them.

The problem

Most of the slow work in a product organization is translation: from one format to another, from a symptom to its root cause, from an idea to a story an engineer can build. It's done by hand, in meetings, and it's the reason good ideas take a quarter to ship.

What I saw

AI is very good at translation, as long as a human reviews the output and the parts that must be exact stay deterministic. One product leader with the right guardrails can now take an idea through scoping, design, build and testing in a day.

What I built

I use this method daily, in my day job and on my own projects. I own the design and the decisions; AI does a lot of the typing. The public examples, all on GitHub:

• Minneapolis From Above: an ambient, live top-down map of the Twin Cities. Real buses, light rail, planes, Amtrak, freeway traffic, satellites, river locks, weather and wildlife move in near real time across 20+ public data feeds, with terrain shaded from USGS lidar and a sky that follows the real sun. Every hover card says which layers are live and which are modeled. No accounts, no tracking, a strict security policy. (code)
• Yacht or Naught: a real-time multiplayer music trivia game for up to 10 players, on Cloudflare Workers and Durable Objects with the Spotify Web Playback SDK and a 2,468-song library.
• SwapCityClassic: a top-down driving game on real Minneapolis streets from OpenStreetMap, with building heights measured from USGS lidar, traffic and pedestrian simulation and an automated test suite, packaged as a single file that runs offline. Playable in the browser. (code)
• Shiely Capture: a Mac menu bar screenshot tool with Snagit-style scrolling capture and video.
• Clipbook: a macOS clipboard history in a grid of large tiles, stored locally in SQLite, that skips password-manager items.

I run the same method at home. With AI as my analyst, a line-by-line review of every bill, policy and fee in my household now saves about $10,000 a year, each line backed by a statement.

What I learned

English is now a programming language. The hard part isn't the AI; it's the guardrails: standards, review and deterministic outputs, so speed doesn't cost trust.