← Blog

The Medallia Post-Mortem: When Owning Data Is Not the Same as Being AI-Ready

There is a version of the AI story that every PE-backed company wants to tell. We sit on a mountain of data. AI turns that data into an edge. The edge drives the next leg of growth. The exit prices in the upside.

It is a clean story. It is also the story that just cratered one of the more data-rich companies in the market.

Medallia, a customer-experience software company, was taken private in a deal reported at roughly $6.4B. The business sat on enormous volumes of customer-experience data, the exact raw material an AI thesis is supposed to feed on. The company was reportedly handed to its lenders after the value collapsed, with a reported figure of roughly $5.1B in equity wiped out. By reported accounts the company held large volumes of customer data but could not make that data AI-ready while the debt obligations kept running.

I want to be careful with the specifics. The take-private and equity-loss figures here are reported, not audited from the inside, and you should treat them as reported. But the shape of the failure is the part worth studying, because it is the shape I see forming inside portfolio companies right now.

Owning data is not the same as being able to use it. And under a leveraged structure, the gap between the two is where the equity goes.

The asset that was not an asset

The instinct in a deal like this is to count the data as an asset on day one. Look at the volume. Look at the customer relationships behind it. Look at the AI roadmap the data is supposed to power. The thesis writes itself.

The problem is that data volume and data usability are different things, and they are not even closely related.

A company can hold a decade of customer-experience records and still not be able to answer a basic question across them. The data lives in different schemas from different product generations. Customer identifiers do not reconcile across systems. Definitions drifted as the product changed. Half the records are structured for a use case that no longer exists. The volume is real. The usability is not.

This is the part the day-one model does not capture. It values the data as if it were inventory sitting in a warehouse, ready to ship. In practice it is closer to raw ore. Until you build the thing that refines it, you do not have an asset. You have a project.

I have made this argument before in the general case. AI readiness starts with data readiness. What the Medallia situation shows is what happens when you skip that step and put the unrefined version on the balance sheet anyway.

The clock the data project ignored

Here is the part that turns a hard data problem into a fatal one.

Getting fragmented customer data into a state where AI can actually run on it is a multi-year program. Master data so a customer is one customer across every system. A quality baseline so you know which fields you can trust. An integration layer so the data moves without a person dragging a CSV around. Governance so the definitions hold still long enough for a model to learn from them. This is not a quarter of work. Done properly it is years.

A leveraged structure does not grant years for free.

The debt does not wait for the data project. Interest accrues on a schedule that has nothing to do with how the integration backlog is going. Covenants test on dates that arrive whether or not master data management is finished. The cash that would fund the multi-year refinement is the same cash that services the debt. Every quarter the foundation is not ready is a quarter the company pays for the gap and gets nothing back from the asset that was supposed to justify the price.

That is the trap. The data thesis needed time. The capital structure was built for speed. When the two collide, the structure wins, because the structure has payment dates and the data project only has milestones.

This is a sharper version of a pattern I see across the portfolio world, where the portfolio company AI initiative stalls not because the idea was wrong but because the foundation was never there to stand on. In a lightly levered business a stalled AI initiative is a disappointment. In a heavily levered one it is a countdown.

Readiness is the gating item, not a workstream

The lesson is not “be careful with debt.” Plenty of leveraged deals work. The lesson is about sequencing, and it is specific.

If your thesis depends on AI, then data readiness is the gating item. It is the thing that has to be true before the rest of the plan can run. It is not a workstream that proceeds in parallel with the value creation that depends on it. It comes first, or the value creation it is supposed to enable does not happen.

Most plans get this backwards. The AI use cases go on the year-one slide. The data foundation goes in the “enablement” section underneath, treated as plumbing that the technical team will sort out while the real value gets created above it. Then month three arrives, the team discovers the data cannot support the use case, and the foundation work that was supposed to be invisible becomes the whole project. Now you are doing the multi-year program you did not budget time for, and the clock has already been running since close.

The companies that get this right invert the order. They treat readiness as the first thing the thesis is tested against, not the last thing it relies on. Before the model commits to AI-driven upside, somebody answers a hard question. Can this data actually support what we are about to promise, and if not, how long and how much to get it there? If the honest answer is three years and the hold is built on a structure that needs results sooner, that is not a detail to manage later. That is the deal.

What to do before the thesis depends on it

If you are running or backing a company with an AI or data story attached to the value creation plan, here is the work to do now rather than after.

Separate volume from usability. Stop describing the data asset by how much there is. Describe it by what you can do with it today, without a project in between. If the honest answer is “not much yet,” the asset is a project, and projects have costs and timelines that belong in the model.

Put a real number on the foundation. How long to get master data, quality, integration, and governance to the point where AI can run reliably? Months or years, with a cost attached. Vague is the enemy here. Vague is how a three-year program gets sold as a year-one enablement line.

Test the foundation timeline against the capital structure. Lay the data readiness timeline next to the debt schedule. If the foundation finishes after the covenants start biting, you do not have an AI thesis. You have a financing problem wearing an AI costume. Better to know that before close than to discover it in year two.

Sequence readiness ahead of the use cases that need it. The AI use cases that depend on clean, integrated data do not start until the data is clean and integrated. Pilots on sample data prove nothing about whether the foundation can carry production load. If a use case cannot name the data it needs and confirm that data is ready, it is not ready to fund.

Build the foundation while it is cheap. Foundation work in the early hold compounds. It guides the plan, it de-risks the AI bet, and it is already in place at exit. The same work done late, under covenant pressure, with the equity already thin, is a fire drill. And as I have written elsewhere, unaddressed data problems cost roughly half a turn of multiple at exit, which is the quiet version of the same failure that played out loudly here.

If you want a fast read on where a company actually sits, the AI readiness assessment walks through the prerequisites and scores the gap. It will not finish the foundation for you. It will tell you, in an hour, whether the foundation is a quarter of work or a multi-year program, which is the single most important input to the question of whether the thesis is real.

The bottom line

A company can sit on one of the largest customer data sets in its category and still fail, because the data on its own does nothing. It has to be made usable first, and making it usable takes time the capital structure may not have agreed to give.

The reported numbers around Medallia will get refined as more comes out. The lesson will not change. Owning data and being AI-ready are two different states, separated by a multi-year program, and a leveraged deal that mistakes the first for the second is paying full price for an asset it cannot yet use.

If the thesis depends on AI, readiness is the gating item. Build it before the thesis depends on it, not after. After is when you find out the clock was never on your side.

For a weekly brief on data strategy, AI readiness, and the failure patterns that quietly erode PE valuations, subscribe to Inside the Data Room. One constraint, one framework, one practical tool. Every week.