This month Clarivate sold its entire life sciences data business for $600 million (Clarivate, 2026). Six years ago it paid $950 million for one company inside that business, Decision Resources Group (Clarivate, 2020). The whole segment went out the door for less than two-thirds of what a single piece of it cost.
I’ve spent the better part of ten years in this world. I built a healthcare data company, Carevoyance — commercial intelligence for medical device teams — and Definitive Healthcare eventually bought it. I know these companies. I know how hard the work is, and I think most of them are genuinely good at it. So I don’t say the next part lightly: the ground under all of them has moved.
Look at Definitive. It went public in 2021 worth about $4 billion. Today it’s worth under $100 million, less than a single year of its own revenue (market data). That’s not one bad quarter or one bad deal. It’s the market repricing an entire way of selling data, and I’m not writing this from the stands. I built in exactly the model getting repriced.
What actually got repriced
It’s easy to read this as AI killing data businesses. I think that’s exactly wrong. The data is still good. These companies spent decades building it, and it doesn’t get less valuable because a model showed up. What broke is the container they sold it in: the portal, the seat, the login.
Every one of these businesses ran on the same assumption, a person on the other end. A researcher who logs in to run a search. An analyst who pulls a report. A team that renews its seats every year. For twenty years they sold the data and made the customer drive to the pump to get it.
Agents won’t make the trip. When the thing running the query is an agent, the interface stops being an asset and becomes friction. Agents don’t log into portals, sit through onboarding, or renew seats. They call an API and move on.
And this isn’t a healthcare data story, it’s a SaaS story. The 2026 selloff got its own name, the SaaSpocalypse, when software multiples fell below the S&P 500 for the first time on record (Oliver Wyman, 2026). The same question is hollowing out the whole category: if an agent can do the work without your interface, what is the seat for?
The answers that don’t work
The first instinct is to bolt a chatbot onto the old product. “Chat with our database.” Everyone’s shipping one. It’s appealing because it reprices nothing and forces no hard board conversation, but it’s a nicer pump, not a different business. It competes with the frontier labs on the exact axis where an incumbent can’t win, defending an interface people are already leaving.
The second answer is to stop selling to people and sell to the models, piping your data to the frontier labs at inference time, priced per call instead of per seat. Smarter, still a trap. You’d be selling into a market of maybe five buyers, every one richer than you and able to build around you the day your margin gets interesting. That’s how you get commoditized slowly instead of quickly.
The third answer is to become a data connector, handing your data over structured and clean, ready for anyone’s agent to call. This one partly works. Not every company can structure its own data, and somebody has to. But two things sink it as a business. Validation is the hard part and the thankless one, because proving the data is right and keeping it right is most of the job. And “we’re the most accurate” has never been a moat, it’s the line every competitor uses. So the category collapses to price. We’ve watched this movie: ZoomInfo defined B2B contact data, and Apollo undercut it until the whole category raced to the floor. Connectors don’t spare you that. They get you there faster.
But doesn’t MCP fix this?
This is where a technical reader pushes back. The Model Context Protocol lets an agent call any data source directly, so expose your database as an MCP server, let the agent call it, and the “agents don’t log in” problem goes away. New pipe, same charging.
It’s the right question, and the answer runs the other way. MCP doesn’t rescue the silo. It finishes the repricing.
Your interface was your lock-in. People came to you because they’d learned your screens, your query language, your reports, and that habit was the moat, not the data under it. MCP dissolves the habit. When every source is one identical call away, the agent doesn’t care where the data lives; it calls whoever’s cheapest or best on the data itself. Hand your data over a standard connector and you’ve made it easier to route around you, not harder, the same price collapse the connector answer walks into, now with a standard plug.
And the orchestration you spent millions on, deciding what to pull, how to combine it, what to report, moves up to the agent. A small team rebuilds your workflow for their one use case in an afternoon. You go from being the product to being one call inside someone else’s product.
For a vertical workflow, the same protocol cuts the other way. When you own the workflow, you make proprietary, real-time data just by doing the work: the case that just closed, the field inventory right now, the price that actually applies to this account. That data isn’t in anyone’s archive and can’t be scraped, because it only exists because you ran the process. MCP lets you pull context from every surrounding system and be the source other agents call for operational truth. The silo uses MCP to expose an archive and gets commoditized. The workflow uses it to expose live truth only it produces, and compounds.
The fuel needs an engine
So the data isn’t the problem. Its home is. Fuel in a tank doesn’t move anything. It needs an engine, and the engine is a system of action: the workflow where the real decisions get made. What to stock. Who to target. How to price. What to bill, and when.
Put the data inside that workflow and it stops being an archive you visit and starts driving the decision. Here’s the part I’m actually excited about: the decision throws off new data, what sold, to whom, at what price, what moved in the field, and that sharpens the next decision, which throws off cleaner data still. The fuel runs the engine, and the engine refines more fuel. That’s a flywheel of continuous commercial improvement, not a database that starts aging the day you buy it.
Which means this data is worth more in that future, not less. It just has to live in the work instead of a silo people have stopped visiting. That’s what “vertical” actually means, not a database that serves every industry an inch deep, but one industry’s real work, all the way down.
We’re the engine for commercial workflows
Let me put a stake in the ground: Deviceflow is the engine for commercial workflows. We started in medical devices, and this is the whole reason we exist.
When GPT-3 landed in 2020, I was close enough to the data business to see the trade coming. “We have the data, you come query it” was headed to zero, because the querying was about to get automated. So I didn’t build another place to go get data. I built the thing that does the commercial work and produces the data as a byproduct.
In medical devices, the data that matters never lived in a database you could license anyway. It’s in the charge sheet a rep photographs after surgery, the purchase order a hospital emails as a PDF, the inventory transfer that happens over text. It’s unstructured, it’s moving, and it doesn’t reach a system of record until someone reads it and types it in. Deviceflow reads those emails and texts, pulls the data out, and pushes it into the ERP. We don’t sell teams data about their operations, we run the operations, and the structured data falls out the other side. A system of action, not a system of record.
And the market data these companies spent decades building? That’s exactly what makes the engine run harder: the reference prices, the provider and facility intelligence, the procedure volumes that tell a commercial team where to point. Fused to the workflow, that data is worth more than it ever was in a portal. That’s the part I’d rather build with them than around them.
Who puts them together
So here’s my bet on the AI race. The winners won’t be the data silos, and they won’t be the seat-based apps. They’ll be the vertical workflows that own the work and the data it throws off, the ones that turn a static archive into a flywheel.
The fuel and the engine aren’t rivals. One is decades of hard-won data. The other is the workflow where it finally gets burned. The only question left is who puts them together.