Most Companies Do Not Have an AI Strategy. They Have a Collection of Experiments.

Ask an executive team to describe their AI strategy and you will usually hear a list of projects. A chatbot in customer service. A coding assistant in engineering. A document summarizer in legal. Someone in marketing is testing image generation. Each project has a sponsor, a vendor, and a demo that impressed a leadership meeting.

That is not a strategy. That is a collection of experiments. I have spent more than twenty five years running technology organizations at Apple, AWS, PayPal, AppDynamics, and Fannie Mae, and I have watched this pattern before, with cloud, with big data, with digital transformation. The names change. The failure mode does not. Activity gets mistaken for direction, and eighteen months later the board asks what all of it added up to.

The uncomfortable answer is usually: a set of demos, a pile of vendor contracts, and no durable capability.

Experiments are easy to start and easy to celebrate. They require no hard choices. Nobody has to say no to anything, nobody has to fix the data layer, and nobody has to change how a team actually works. Every pilot produces a slide, and slides accumulate into the appearance of momentum.

The problem is that pilots optimize for demonstration, not for operation. A proof of concept runs on clean sample data, with an enthusiastic team, outside the security review, and with no expectation of uptime. Production runs on messy data, with skeptical users, inside compliance boundaries, and with an on-call rotation. The distance between those two states is where most AI initiatives die, and no volume of additional pilots closes it.

At AppDynamics and later at Fannie Mae, where my teams supported platform and reliability work across more than six hundred applications, the lesson was consistent: capability comes from what you operationalize, not from what you demonstrate. The same is true for AI.

A real AI strategy is a document, not a portfolio review. It answers four questions in writing, with names attached.

First, where does AI create business value for us specifically. Not in the abstract, and not by industry analogy. Which workflows, which cost lines, which revenue motions. If the answer is everywhere, the answer is nowhere.

Second, what foundations does that value require. Data that is accessible and governed. Platforms that let teams ship a model-backed feature without a six month integration project. Guardrails for security, privacy, and model risk that are defined before deployment, not after an incident. Talent, which usually means retraining the people you have, not just hiring the people you want.

Third, what will we not pursue. This is the part most strategies skip, and it is the part that makes a strategy real. Saying no to plausible use cases is what concentrates investment enough to matter. A strategy with no exclusions is a wish list.

The last thing a real strategy defines is how progress will be measured, in business terms, on a schedule someone is accountable to. Adoption counts and demo counts are not measures of value. Cycle time, cost per transaction, error rates, revenue per employee: these are.

At Fannie Mae, my organization drove more than two hundred million dollars in infrastructure savings. None of that came from experiments. It came from committed targets, an operating model that assigned ownership, and a cadence of measurement that made backsliding visible. AI value works the same way. If nobody can tell you the number that should move, and by when, you are funding curiosity, not strategy.

There is a simple test. Ask three questions of your current AI effort.

Can anyone produce the document that says where AI creates value here, what we are deliberately not doing, and what number moves by when? If the answer is a deck of project updates, you have experiments.

Has anything gone from pilot to production with real users, real data, real controls, and a team that owns it after launch? If everything is perpetually in pilot, you have experiments.

Did any pilot change how you fund and build, or did each one end where it started? Experiments that leave no operating model behind are entertainment with a budget line.

None of this means stop experimenting. Experiments are how you learn, and I want teams running them. It means experiments should be downstream of a strategy, chosen because they test a bet the company has actually made, and judged by whether they graduate into operations.

The companies that will get durable value from AI are not the ones with the most pilots. They are the ones willing to write down where the value is, build the boring foundations underneath it, say no in public, and measure the result. That work is less fun than a demo. It is also the only version of this that compounds.