AI's Value Capture Problem

AI could be one of the most powerful value-creation technologies in history and still have a value-capture problem.

Alex Karp says companies buying AI risk leaking their IP into Anthropic and OpenAI. Satya Nadella calls the answer sovereignty: a firm keeping control of its own intelligence instead of renting it back one query at a time. They are saying similar things but communicating them a bit differently: the scarce asset is no longer only the model, but it's also the context and know-how the model learns from the aggregate patterns across your company and it's competitors.

Take insurance. Imagine State Farm, Progressive, Allstate, Travelers, Chubb, AIG, Liberty Mutual, and 100-plus smaller carriers all running claims through the same model. Every carrier feeds it the same stream of context: the accident description, photos, repair estimate, adjuster's note, borderline approval, fraud flag, override, payout, appeal, recovery outcome.

At first this is obviously useful. The model moves claims faster, flags suspicious cases, learns which repair estimates run inflated, which medical patterns look strange, and which overrides later become losses.

But if the same model learns from every carrier, is your claims judgment still your advantage? The underwriting exception that protected your loss ratio becomes a benchmark. The fraud pattern your team caught early becomes a feature sold back to the market.

Notice what you keep and what you lose. The insurer still owns the risk, the customer relationship, the regulator, and the loss ratio. The shared model increasingly owns the learning curve. Your mistakes, overrides, and hard-won claims intuition become training signal.

That intuition is part of your real intellectual property. Not the registered kind, the patents and the brand, but the operating kind: how your people price risk, catch fraud, read ambiguity, and use everything the firm knows. The model can dissolve one of your moats by making that scarce judgment reproducible.

This is why enterprises focused only on protecting data are thinking too narrowly. The deeper asset is institutional context and know-how: the judgment in people's heads about hard professional work.

The labs understand this. OpenAI and Anthropic are reportedly scaling data 10x year over year and spending billions mobilizing domain experts to create the tasks that train agents. A task iis expert work packaged into something a model can learn from: prompt, environment, action, rubric, verifier, score.

Now take life sciences. Anthropic's made it's direction clear: tools for researchers today, more autonomous discovery over time. Claude for Life Sciences and Claude Science put literature, agents, scientific artifacts, reproducibility, and compute into one "workbench". If thousands of biotechs use that system around targets, assays, safety, endpoints, and kill-or-continue decisions, the nightmare is not that Anthropic sees some biotech’s specific discovery; it is that Anthropic learns what serious drug discovery questions and judgment looks like across 1000s of companies while also entering that space.

First-party products are how to capture this “learning” at scale. In insurance, the model dissolves your edge into the industry's baseline. In pharma, it can do that and then compete with you using what many taught it. It also makes your actual moats more exposed (more on that later)

I think no one would argue with the fact that AI creates value by making private know-how usable at scale. But it also makes the "know-how" less scarce. If every insurer, bank, or biotech can access the same capability through the same model, what used to be your edge becomes the industry’s baseline. The value does not disappear; it gets split: customers get lower prices or better service, the model vendor gets the learning, and you get a front-loaded productivity gain that competition wears down.

Here is why so little of the durable value stays with you.

1. If everyone gets the same edge, customers keep it. Imagine an auto manufacturer using a model to negotiate semiconductors, resin, freight, contract manufacturing capacity, and substitute parts. The edge is buying better than the next manufacturer: knowing which supplier shortage is real, which quote embeds excess margin, and when preserving supply matters more than squeezing price. If every manufacturer runs procurement through the same model, the model does not just lower costs. It makes buying more "similar". The best buyer loses the spread between its process and everyone else's. Suppliers adapt too: once every buyer arrives with the same should-cost analysis, alternate-source map, and negotiation script, the playbook becomes priced in.
2. The model also captures what compounds. Imagine 1,000 resource-constrained biotechs using Claude for Life Sciences because they do not have the internal platform of massive pharma company. Each company owns its compound, lab cost, failed program, and regulatory trail. But the workbench can see the pattern across all of them: which tox signal killed the program, which assay gave false confidence, which endpoint was weak, and which patient subgroup wasn’t the right one. If it sits across enough biotechs and pharmas, it can see failure patterns no single company can see. While data advantage is in exclusivity, a shared workbench breaks exclusivity by aggregation. And because Anthropic intends to develop drugs of its own, the tool you adopt for efficiency is built by the entity whose endgame may be to do what you do, using what it learned by watching the field do it.
3. You contribute the unique and receive the average. You contribute differentiated judgment, data, context, and decisions: the fraud pattern your team alone caught, the supplier bluff your buyer ignored, the trade your PM killed before the market saw it. You get back the blend of everyone's. Citadel would never want every pod in the world trained on its best PM's kill criteria. For the best firm, that is the losing trade: you hand over above-average judgment and receive the mean
4. Data rights are not learning rights. Companies know how to negotiate retention, confidentiality, security, access controls, and training opt-outs. But the more important question is who owns the derived judgment: tasks, feedback loops, evals, workflow traces, corrections, failure modes, decision patterns, agent skills, and product insights. Once the model company knows the hard problem, it can acquire the job logic another way. It can source experts to create cases that test the same decisions: should the model raise rates, tighten underwriting, flag fraud, exclude a segment, or accept a worse loss ratio to keep a profitable customer? The reasoning becomes trainable.
5. The gain is front-loaded; the dependency compounds. The first adoption creates a real productivity jump. But once competitors run the same model, that jump becomes the baseline, and what remains is not your edge, it is your dependency on the next upgrade. Everyone will capture the first uplift but the vendor captures the recurring learning curve. Year one, the factory model reduces downtime, but then every rival has the same predictive maintenance workflow and the vendor owns the process intuition you now depend on.

None of this means zero capture. The first mover banks real profit in the window before rivals adopt. The only problem is that the durable value goes to whoever owns the learning, and by default that is not you. Which turns the whole thing into decisions made workflow by workflow, task by task. Where your work is generic, pool it and take the gain, because there you are protecting mediocrity. Where your people's judgment is the product, keep it off the shared model.

Here is the simplest way to see it. Think about TikTok, YouTube, and Google: you think you are the customer, but you are the raw material. Every video you finish teaches the algorithm what works, and that learning is the real product, sold to the next advertiser and used to hook the next user.

That is how CEOs should think about Anthropic and OpenAI: TikTok for enterprise data, except the feed is your work and the engagement signal is your judgment. The model providers are that machine pointed at the most expensive know-how in your company. Your experts show up for help with claims, trades, clauses, suppliers, trials, risk calls, and production problems. Every hesitation, override, escalation, approval, rejection, and second look teaches the model how your company thinks.

On TikTok, the creator at least gets paid. Here, you supply the data, context and the learning ("know how") from the data, and the platform can sell the finished product back to your entire industry, or eventually choose to compete with you in the case of pharma.

So the executive question is simple: do you want your own company's TikTok, or do you want to use the shared one? You probably need to route through both.

Before putting any high-value workflow into Anthropic, OpenAI, or another shared model, ask one question: if every competitor learned how we handle this decision, would we still be better than them?