Why most AI initiatives die between the deck and production

The board approves the AI strategy in March. By August there's a pilot that works beautifully in demos. By December the pilot is quietly shelved and everyone has moved on to next year's planning. I've watched this cycle play out at enough companies that I can usually tell by week three whether an initiative will survive it.

The deck is almost never the problem. Strategy decks are genuinely good these days. The use cases are sensible, the market numbers are right, the risks slide even mentions hallucination. The problem is everything the deck doesn't have to deal with.

The pilot trap

A pilot's job is to prove the idea works, so teams build it to prove the idea works. Real data access gets mocked. Awkward edge cases get filtered out of the demo script. Security review gets deferred because it's just a pilot. Then someone asks to put it in front of actual customers, and you discover the whole thing needs rebuilding. Except now the budget is spent and the sponsor has mentally moved on.

The demo was the easy 20 percent. Production is the other 80: authentication against systems built in 2011, a data pipeline that breaks on public holidays, a legal team that has never reviewed an LLM product before and wants three months to think about it.

What actually kills it

In my experience it's rarely the model. The model is fine. What kills the initiative is one of these, and usually two at once:

• Nobody owns it. The data science team built it, the product team didn't ask for it, and the operations team is measured on numbers it doesn't move.
• There are no evals. Nobody can say whether this week's version is better than last week's, so nobody dares ship a change.
• Data access took four months. By the time the pipes were connected, the champion who sold the project internally had left.
• The workflow never changed. The tool exists, but the people it was built for were never given a reason, or permission, to work differently.

What the survivors do differently

The AI products I've seen make it into production share a boring pattern. They start with one workflow, not a platform. They ship to five named users in the first month, not a business unit in a quarter. They write evals before writing prompts, so improvement is measurable instead of vibes. And they put one person on the hook for a business number, not an accuracy score.

The plumbing matters more than the intelligence. I know that's an unglamorous conclusion for a technology this interesting, but the companies getting real value from AI right now are mostly winning on data access, deployment speed and workflow design. The model is the part you can buy.

A test I use: if the team can't name the five users who will touch the thing in the next 30 days, it's a deck, not an initiative. Fix that first.