Gartner ran the numbers on AI investments last year. One in fifty paid off in any meaningful way.

Last quarter, the world spent $242 billion on AI. That’s four times what it spent the quarter before. The money is going in. The returns mostly aren’t coming back.

MIT puts the failure rate at 85%. Projects that never made it to production, or did and showed nothing. And 61% of senior leaders say they’re under more scrutiny to prove ROI than a year ago.

Nobody is panicking about the technology. The technology works.

The problem is everything that happens before anyone touches it.

The Same Pattern.

A team spots an opportunity. Someone builds a demo. The demo looks good in the meeting. There’s a budget, a build and a launch. Six months later someone asks: is this working? Nobody has an answer.

Why doesn’t anyone have an answer? Because nobody wrote one down before they started.

The 98% who get nothing from AI tend to make the same mistakes.

They picked a tool before they had a problem. The conversation started with ‘we should be using AI’ before anyone asked what for, or whether that was even the right approach.

They built for the demo. A demo is optimised for expected inputs, a controlled setting and someone paying attention. A production system has to work at 2am on a Tuesday with no one watching. These are not the same thing.

They skipped the failure mode question. Specifically: when this breaks, who finds out first? And how bad is it? That question belongs at the start. Not in the post-mortem.

They called launch the finish line. The system went live. Everyone moved on. No one had written down what success looked like, so no one could check.

They automated something broken. Messy processes don’t become clean when you automate them. They just fail faster.

Briefly Speaking.

The 2% who get something back share one habit. They write a brief before they build anything.

Not a proposal. Not a requirements doc. A brief. Four questions, answered in writing before any code gets touched.

What problem are we solving? What does a good outcome look like in numbers? What breaks when this fails? Who finds out first?

At Bynry Foundry, this is how every engagement starts. Before design, before tooling, before any vendor conversation. Sometimes the brief leads somewhere uncomfortable: AI isn’t the right answer here. There’s a simpler thing that costs less and breaks less. We say that.

The Evaluate phase at the end isn’t optional. It’s how we close the loop on the criteria we agreed on at the start. Launch is when the work gets measured, not when it ends.

What Be Done.

If you’re already six months into a build with no success criteria written down, it’s not too late. Write them now, before you go live.

And if you’re evaluating a vendor right now, ask one thing: what did your last three post-launch reviews look like? If they can’t show you, the projects didn’t have any.

The 1-in-50 problem is not inevitable. But solving it starts with the right question, not the right tool.

What does your post-launch success metric look like? Drop it in the comments and we’ll give you a straight read.