How Sotheby's builds product when it can't afford to break things

Today she leads the objects platform team at Sotheby's: the tools and services that deliver data about every artwork and luxury item the auction house sells to its applications. Six years in, she has watched product and development grow from a function that once rolled up into marketing to a team that builds a lot of its products in house.

We got into why a 270-year-old brand can't run experiments the way eBay does, how one shared data model lets Sotheby's track an object across decades of sales, where computer vision and LLMs help match artworks, and the analogy that finally got leadership to fund data infrastructure.

Watch the episode on YouTube

Key takeaways#

How do you get into product management without a tech background?#

Start by caring loudly about the product. Sam was doing marketing and data entry at a small music startup and kept emailing the tech lead about the app. Why is this button here? Why is the font like this? It isn't doing what I think it should. He got fed up, showed her Jira and told her that's how you write a ticket. She ended up deeply involved in a redesign of the app. Someone in the coworking space told her that made her a product manager. She didn't know the word yet, but that person ended up hiring her, and she took a bootcamp to fill in the gaps.

Try this: if you keep spotting problems in your company's product, write them up as tickets. It's the fastest way to show you think like a product manager.

Can you do product like a startup inside a very old company?#

Sam was sceptical. When she interviewed in 2020, she figured Sotheby's would be the oldest company she'd ever work for and expected a lot of red tape. Her future boss told her it was like working at a startup. That turned out to be more right than she knew. Product and development used to roll up into marketing, because for a long time the website was mostly there to promote what you'd then go and buy in person. Since then Sotheby's has moved to building a lot of its products in house. A company that does this many different things has a lot of problems to solve, and Sam was given the leeway to pick up some big ones and run with them.

Try this: when you join a legacy company, find the big problem that nobody owns yet. That's often where the startup-style freedom sits.

Why can't luxury brands move fast and break things?#

Because the client pool is small and every client counts. Sam compares Sotheby's with eBay, which has users in the tens of millions. Sotheby's might get millions of website visitors, but the people who transact are a much smaller group. eBay can absorb a buggy feature or the loss of something a few users loved. Sotheby's can't afford to burn clients that way. So the team tests rigorously and asks one blunt question before every release: how much are we willing to burn? The upside is a large client-facing staff who can walk clients through new experiences, something eBay doesn't have.

Try this: before your next release, put a number on how many unhappy customers you could live with. If the answer is close to zero, plan more testing.

How do you roll out new features to a high-touch client base?#

Prepare the people who talk to clients first. At Sotheby's, specialists hear the feedback directly: a client was confused by something in the app, so why did you build it that way? Sam's team needs answers ready, and the changes socialised well inside the company, so specialists are never on the back foot when they talk about the company. That internal loop is a source of feedback as much as a risk to manage.

Try this: brief your client-facing team before launch day, with the reasoning behind each change. They will get the questions before you do.

What makes digitising a physical luxury experience so hard?#

The mindset, more than the technology. Sam shares an anecdote: there was a time when phone bidding was a scandal. People were used to being in the room, and the idea of calling someone across the country to bid for you met real resistance. Today it's in the movies. Sotheby's was already running online auctions before 2020, and going fully remote put fuel on the fire. Some of the work is still persuading people that technology can do more than complete tasks. It can actively help, with data that informs how the company works with clients. The encouraging sign is people experimenting with new tools on their own.

Try this: find the people already using tools you didn't roll out, and ask what they use them for. That curiosity is the mindset shift starting.

How do you personalise the experience for luxury clients?#

Know what each client collects, and suggest what comes next. If someone is a watch collector, Sotheby's shows them watches, and similar pieces to ones they've looked at before. If they've been buying watches for years, maybe there's art that relates. Delivery is part of the experience too. Artwork comes in every shape and size, and many pieces need a custom crate built by specialists, which adds steps and unpredictable delays. A Rolex fits the same watch box every time, so those deliveries can be standardised with clear updates. For artwork, the team is looking at agents that could read vendor communication and keep the client informed, without someone writing every update by hand.

Try this: separate the parts of your service that can be standardised from the true one-offs. Automate the first and give the second more care.

Why does data quality start with the people creating it?#

Because cleaning it up later costs money and dev time. Garbage in, garbage out is a mindset Sam coaches people on. Sotheby's specialists write the content for every object: size, dimensions, how it was measured, colours, materials. If that isn't captured consistently, tools can extract it from the text, but that is expensive. So the message is simple. If you want to see results in this application, you need to make the data nice over in that one. There's also a judgement call on history. Is it worth going back 270-some years to digitise everything?

Try this: show the people entering data what it powers downstream. Consistency improves once they can see the payoff.

How do you track a unique object across its whole history?#

Give the whole company one definition of the object. Sotheby's used to build every product end to end, and each one had its own idea of what an object was. Shipping cared about weight and didn't care about the artist. Online sales cared about the artist and not the weight. Data was lost and recreated at each step. Sam's objects platform team built a single model that every application reads from, adding or ignoring what it doesn't need. That creates a through line as a piece moves through the business, and over time. Something sold in 1998 can come back years later, and the team wants to know it's the same piece. That history, the provenance, is central to authenticity and to the brand.

Try this: list every system that describes your core product and compare the fields. The gaps show where data gets lost.

How can AI help match artworks across decades of records?#

Combine computer vision and LLMs, and keep a human in the loop. Sotheby's is building timelines of artworks' histories, which means deciding whether a piece sold in 1998 is the same one up for sale now. The photo from 1998 might be worse, there's a lot of untitled work out there, and descriptions vary. Their tools compare images and text together. When a match scores somewhere between 75% and 85%, someone on the team goes in and confirms it. Below that, they don't bother, because a match is very unlikely. The reviews teach them things too. For editions like Andy Warhol prints the pictures look almost identical, so the text description matters more than the image.

Try this: set a confidence band where AI decisions go to a person. Use what those people decide to improve the model.

How do you get leadership to fund data infrastructure?#

Tell a story people can picture. Sam says a big part of a product person's job is selling what you're building, and with data products people only see the tip of the iceberg. The CEO needs to understand why the money goes there and why the team isn't building something else. Analogies did the most work for her. After a few that left people scratching their heads, one stuck: data is the infrastructure of a city. The power plant doesn't care whether you're a restaurant or a home, it just has to deliver electricity you can use. Before the platform, everyone at Sotheby's ran their own solar panels. That works for a while, until the city gets big and one outage leaves you in the dark.

Try this: test three analogies for your least visible project on someone outside tech. Keep the one they can repeat back to you.

What does a small internal tool change at scale?#

A lot, if it removes the task everyone hates. Sam's proudest delivery from her individual contributor days is a button called Get Images. Photography at Sotheby's lived in a legacy system, and someone had to download every image to their computer and upload it again to the CMS. In user interviews people called it the number one time suck. The legacy system had no functional staging environment, which made it hard to build on, but the team got it connected. Now you hit Get Images in the CMS and the files show up, in about 10 minutes for big ones. Their back-of-napkin estimate was a full year's worth of time spent on downloading and uploading. Now it takes hours. A couple of days after launch, people were writing in to say thank you.

Try this: ask your colleagues for their most tedious recurring task. Measure the hours. That number is your business case.

---

Listen & watch#

• Watch the episode on YouTube
• Listen on Spotify
• Listen on Apple Podcasts
• All episodes at infinum.com/delivered DELIVERED is a production of Infinum and Your Majesty.