The 50%
Problem.

For more than a century, advertising carried a famous problem: the buyer knew some portion of spend was working but could not identify which portion. The problem lasted because the market lacked a trustworthy meter.

Enterprise AI has arrived at the same place. Companies can count models, licenses, tokens, agents, tasks, and reported hours saved. None of those numbers answers the economic question: what changed in the business, and what would have happened anyway?

Activity is not the outcome.

A model-generated draft is an action. A resolved claim, collected invoice, retained customer, or avoided escalation is a business outcome. The two may be related, but they are not the same event.

AI value is unusually hard to measure because the contribution is distributed across a person, a system, and a span of work. A user saying that a tool saved an hour is useful evidence. It is not yet a verified economic result.

The same model keeps returning.

Each technology cycle introduces a new unit that is easy for the seller to count. The name changes; the allocation of risk does not.

The seller reports the proxy. The enterprise still has to defend the value—and carries the performance risk.

Four gaps keep the answer out of reach.

Advertising did not get smarter. Its accounting did.

Digital advertising changed when measurement moved closer to a business result. Clicks, conversions, holdouts, and incrementality tests did not make every ad effective. They made performance more countable, comparable, and contestable.

That changed the market. Buyers could direct capital toward measured performance. Sellers that produced results could accept more accountability. Risk began to move away from the buyer alone.

What a trustworthy meter would require.

1. 01
   An accountable outcome

   A business event close to the P&L—not another activity metric.

2. 02
   Instrumentation at the point of work

   A record of the AI action and the downstream event it may have changed.

3. 03
   A baseline and counterfactual

   A way to separate measured lift from coincidence, seasonality, and work already underway.

4. 04
   A ledger both parties can trust

   One chain from action to outcome that the enterprise and vendor can verify.

5. 05
   Risk aligned to the measured result

   Commercial terms that make performance a shared responsibility.

Research note. This article is adapted from OI's The 50% Problem research brief, which synthesizes 2024–2025 findings from MIT Project NANDA, McKinsey / QuantumBlack, BCG, S&P Global, Gartner, IBM, and Deloitte.