Proprietary Data, AI Moats and Differentiation

In May 2019, two investors at Andreessen Horowitz, Martin Casado and Peter Lauten, published a curve that should sit in front of every board that hears the words “data moat”. It came from companies building automated assistants that answer customer requests. As the assistants were fed more and more real conversations, the share of requests they could handle rose quickly at first, then bent, and then flattened near about 40 percent. Past that point, more data bought almost nothing¹.

Their conclusion ran against the standard pitch, in which more data makes a better product that brings more customers, forever. On the curve, each new piece of data was worth less than the one before, while the unusual cases that would have helped were getting harder and more expensive to collect. The first firm to gather a large pile had a head start. A rival with a smaller pile would reach the same plateau a little later and then stand level with it.

Many organizations hold data that nobody else has: transaction histories, maintenance logs, customer records, years of operating decisions. Many of them describe it as their AI moat. Sometimes they are right. This chapter gives you the test that tells you which case you are in.

Unique data is a head start, not yet a moat

The previous chapter, AI and Competitive Advantage, argued that an advantage needs a gap in cost or customer value and a guard that stops rivals closing it, and it described the AI flywheel through which use improves the offer²³. Proprietary data is the guard most often claimed. It is also the one most often overstated. The economists Andrei Hagiu and Julian Wright, after studying how firms actually compete on customer data, concluded that most leaders overestimate how much protection it gives, because they treat data-driven learning as if it were as strong and lasting as a true network effect, and it usually is not⁴.

The distinction that matters is between an asset and a moat. An asset is valuable today. A moat keeps you ahead tomorrow, because using it makes it better in ways a competitor with money and a good model still cannot reproduce.

That makes the board-level question narrow and answerable. Not “how much data do we have?” but “does more of it keep making the product better for the people who use it, and could a rival get the same improvement another way?”

Scale effects are not network effects

Much of the confusion comes from one phrase, data network effect, which bundles three different mechanisms.

A classic network effect exists when a product becomes more valuable to each user because other users join. A telephone network or a marketplace works this way: a buyer gains directly from every extra seller. That value does not depend on anyone learning anything.

A data scale effect is weaker. More users generate more data, and the data may improve the product for everyone. But the users gain nothing from one another directly; they gain only through whatever the system learns, and learning runs into diminishing returns. Casado and Lauten argue that most of what gets called a data network effect is really this¹.

Within-user learning is weaker still. Hagiu and Wright use the smart thermostat as their example: it learns a household’s preferred temperatures within days. That makes the device better for that household, but it does almost nothing for the next customer, and it stops improving once the routine is learned⁴. It can still raise switching costs, which is a real benefit, but it is not a data moat.

The honest test: plateau or rising curve

The cleanest way to judge a data claim is to draw the curve: product quality, as customers experience it, against the amount of data behind it. Two shapes are common; the chart below sketches them with illustrative numbers.

On the first curve, the lead is real but temporary. A competitor that starts later with less data reaches the same plateau and then competes on price, service or distribution, as if the data did not exist. On the second curve, the leader keeps pulling away, because each round of new data still teaches the system something it did not know.

Rising curves do exist, and they have conditions. Studies of Amazon’s forecasting and Yahoo’s search engine point the same way: more data kept paying only where it covered a long tail of rare cases, a world that keeps changing, and learning that spreads from one user to the others⁵⁶. Where those conditions are missing, expect the flat shape.

A useful image is the difference between an ore seam and a spring. A seam is rich at first, and each further load holds less ore and more rock; a neighbor who starts digging later reaches the same rock. A spring keeps flowing because fresh rain keeps falling on ground that you hold. The question for every claimed data moat is which one it is.

Seven questions before you call it a moat

Hagiu and Wright turn the curve into seven questions, and they make a useful checklist for a board paper⁴. Each answer pushes a claimed moat toward the seam or toward the spring.

Three of the seven deserve extra attention, because they are the ones most often skipped. Shelf life cuts both ways: data that goes stale fast weakens the value of an old archive, but it rewards a firm that keeps collecting fresh data and acting on it quickly. Imitation is about the improvement, not the data: if a rival can see what your product now does well and copy it by other means, your data protected nothing. And reach is the difference between a data scale effect and within-user learning: if one customer’s data helps only that customer, the moat is a switching cost at best.

Two conditions sit underneath all seven: the data must be usable in practice, as Data Strategy for AI showed, and you must hold the right to learn from it, the subject of a later section.

What keeps a curve rising: outcomes captured in the work

When a data curve keeps rising, you will often find the same ingredient underneath: the system learns what happened after each decision. A forecast is followed by actual demand; a recommendation by whether it was accepted; a repair by whether it held. Without that outcome, more activity data only makes the pile bigger. With it, each decision becomes a lesson, and the lessons include exactly the unusual cases that a flat curve lacks.

Two practical points follow. First, outcomes are captured in the workflow, not in the data team: the crew, clerk or salesperson closing the task records the result, so the workflow has to make that easy. Second, outcome capture can be undone. A change of contractor, system or priority can cut the loop without anyone noticing, and the story later in this chapter shows how fast the curve then falls.

The right to learn: use rights and trust

A data asset you are not allowed to learn from is not a usable moat. It helps to think of rights as layers. Holding data, because you need it to run a service, is the bottom layer. Using it to run and improve your own operations sits above that. Training AI systems on it often needs a further basis that nobody asked for when the data was collected. Sharing or selling it needs the most.

Rights can also move against you, and machine data is where they move most.

Data a manufacturer assumed was its own may now reach a rival through the customer. Trust works the same way. Customers share data with organizations they expect to use it fairly. Use it in ways they did not expect, and the flow that feeds the curve can dry up, which is a cost no balance sheet shows until it is paid.

Story: Flint’s lead pipes, two seasons, one city

There are no competitors in a city’s pipe-replacement program, which makes it an unusually clean test of what data is worth. Nothing changed between the two seasons in this story except whether the learning loop ran.

In 2016, after the water crisis, Flint, Michigan, had to find and replace the lead and galvanized steel service lines feeding thousands of homes. Its own records were incomplete, and every excavation cost money: a few hundred for a quick hydro-excavation, 2,500 to 5,000 for traditional digging⁷. Two researchers then at the University of Michigan, Eric Schwartz and Jacob Abernethy, built a model that estimated, for each home, the probability of a hazardous line from its age, value, location and old city records. Crucially, it was designed to learn from the result of every excavation, and the team later proposed choosing some digs partly to learn from them⁸⁹.

The loop worked. In the last months of 2016, about nine in ten of the city’s traditional excavations found a hazardous line⁸. By late 2017, the hit rate was still above 80 percent⁷.

In 2018 the city made a different choice. A new contractor took over, and residents told the mayor, in effect, that crews had checked a neighbor’s house and skipped theirs. Wanting to leave nobody behind, the city dug every house on selected blocks across all wards, set the model’s targeting aside, and stopped sending 2018 results to the researchers, so the model stopped learning in December 2017⁷⁸. The residents’ concern was legitimate; a program that seems to skip people loses public trust. But the cost of cutting the loop was large. By mid-August, the city’s hit rate was 19.7 percent, below the 31.4 percent that digging at random among eligible homes would have produced. In Ward 5, where the model expected the most remaining lead, 163 digs found it 95.7 percent of the time; in Ward 4, where it expected the least, 702 digs found it 2.4 percent of the time⁸.

In February 2019 the city agreed, in a settlement with residents’ groups, to return to the model and to prioritize up to 5,200 addresses it identified¹⁰. The contractor in 2018 had the same budget and even a map of the model’s predictions. What it did not have was the habit of recording what each dig found and feeding it back. That habit, not the archive and not the algorithm, was the asset no newcomer could buy. And the reason the loop was cut, public trust, is part of the same lesson: a learning system that the people it serves do not trust will not keep running.

What this means for leaders

The first discipline is vocabulary. When a business case says “data moat” or “data network effect”, ask which of the three mechanisms it means and ask to see the curve. If the honest answer is a data scale effect that will plateau, value it as a head start and plan to win on something else once rivals catch up. A correct label stops the organization overpaying for a guard it does not have.

The second is to invest in the spring rather than the seam. The money that wins is rarely the money spent on storing more history. It is the money spent on recording outcomes inside the work, shortening the time between a lesson and a better product, and reaching the unusual cases that keep the curve rising.

The third is to protect the loop as an asset. Name an owner, and treat any change of contractor, system or process that could cut outcome capture as a risk to the moat itself. And put the right to learn on the same list: what you may use data for, and what customers expect, are part of whether the moat exists.

Check yourself

1. The organization with the most data in a market will keep the lead.MythFact
2. Many claimed data network effects are really data scale effects.MythFact
3. A smart thermostat that learns one household’s habits builds a strong data moat.MythFact
4. A rising data curve needs conditions such as a long tail of rare cases and learning that spreads across users.MythFact
5. If we already hold customer data, we may train AI on it.MythFact
6. In Flint, the 2018 drop in hit rate came from losing the data archive.MythFact

Reflection: seam or spring?

What comes next

With advantage now tested, one chapter remains. Module 04 Synthesis — The Executive AI Strategy brings the module’s choices together into one strategy a leadership team can defend.