Every Week There's a New AI Launch.

The real reason most AI projects fail has nothing to do with the tools.

OpenAI launched ChatGPT Work last week. It can plan, browse, write documents, build spreadsheets, and produce finished deliverables - all from a single prompt. Before that, it was Claude Fable 5. Before that, GPT-5.6. Before that, whatever Google shipped.

If you’re a business owner trying to make a clear-headed decision about AI, the launch cadence alone is enough to make you feel like you’re already behind.

You’re not. And the news cycle is actually one of the least important things you could focus on right now.

Subscribe now

Here’s the number that should be getting more attention than any product launch: 88%.

That’s the share of AI proofs of concept that, according to IDC, never make it to production. Not 88% of bad ideas - 88% of all of them. The ones that looked promising in the demo. The ones the vendor said were a slam dunk. The ones your team spent three months testing.

MIT’s Project NANDA put an even sharper point on it: 95% of generative AI pilots produced no measurable return on the P&L statement. Gartner predicted that 60% of AI projects lacking AI-ready data would be abandoned by end of 2026. One S&P Global survey found 42% of U.S. companies had already scrapped most of their AI initiatives - up from 17% the year before.

These are not startup experiments. These are companies with budget, leadership attention, and a genuine desire to make AI work.

So what’s going wrong?

The Problem Isn’t the Tool

The pilot purgatory problem - and that’s the phrase people in the industry are using - is not a technology problem. The tools are good. The models are fast and cheap. A flagship AI API that cost $30 per million tokens in 2023 runs about $2.50 today. The capability is there.

What’s not there, in most cases, is the organizational infrastructure to make anything stick.

The most consistent failure pattern looks like this: a business hears about a new AI capability, gets excited, runs a test, sees promising results, and then... nothing happens. The pilot sits in a shared folder. The vendor follows up twice. The internal champion moves on to the next initiative. Nobody owns the outcome. No process changed. No workflow integrated.

Researchers at McKinsey describe the core issue as “pilot without a production path.” You can get AI to produce impressive output in a controlled test. The hard part is answering: who owns this when it goes live, what does the process look like when it breaks, what data does it actually need to do its job at scale, and what does success look like in six months?

When those questions don’t have answers before the pilot starts, the pilot usually doesn’t end - it just fades.

The Data Problem Is Worse Than You Think

The single most common technical killer of AI projects is data quality - but it’s worth understanding why, because it’s not a technical problem at its root.

Most businesses have been generating data for years. They have CRMs, email archives, spreadsheets, accounting systems, scheduling tools. What they don’t have is data that’s clean, connected, and structured in a way that an AI system can actually use. It exists in silos. It has inconsistencies. Fields that mean one thing in one system mean something different in another.

A demo doesn’t need clean data. A demo uses curated examples. Production needs your actual data - the messy, real-world version of it.

Gartner’s forecast about data-unready projects being abandoned is pointing at exactly this: companies start a pilot on a clean test dataset, the pilot works, they try to go live, and the actual data breaks everything. That’s not a vendor failure. That’s an organizational readiness gap that takes real work to close.

The businesses that move from pilot to production consistently - and some do, to great effect - tend to make data readiness a first-order priority before they even pick a tool. They know what they’re working with. They’ve done the work to get it organized. The AI implementation is almost the easy part.

Who Owns It?

There’s a softer version of the problem that’s just as deadly, and it’s this: AI projects fail when nobody is accountable for the outcome.

A successful AI deployment has a business owner - someone who is responsible for whether the thing works, who can authorize workflow changes, and who has skin in the game if it doesn’t deliver. Not a vendor. Not the IT team. A person with a job title and a KPI that the AI is supposed to move.

In most failed pilots, that person doesn’t exist. The initiative lives in a cross-functional team that doesn’t have authority to change anything. It gets reported on in steering committee meetings and then quietly dropped when something more urgent shows up.

The companies that get AI into production share a pattern: they treat it like a business transformation project, not a technology experiment. They have an owner. They have a specific problem. They have a definition of success before the pilot starts. And they have a decision point: at the end of the test period, we either integrate this into the process or we stop.

A parts distributor in one documented case study freed $340,000 in working capital and cut stock by 35% - with a four-month implementation that started with a specific, measurable inventory problem. Not an AI transformation. One problem, one process, one owner, clear metrics.

What This Week’s Launches Mean - and Don’t Mean

When OpenAI ships ChatGPT Work and bills it as a workplace agent that can handle ambitious, multi-step tasks, that’s real. The capability is genuinely new. The question is whether your business is positioned to use it.

If you don’t have a specific workflow in mind, a clear owner for the outcome, and data clean enough to actually feed the system - then adding a new agent tool to the pile is just going to add another pilot to your purgatory stack.

The right response to a new AI launch is not enthusiasm or dismissal. It’s one question: does this solve a specific bottleneck in a process I already own, and can I measure whether it improves that process enough to justify adoption? If yes, run a tight pilot with an exit date. If no, note it and move on.

The companies getting real value from AI right now - not in the press releases, in the P&L - are the ones who stopped chasing launches and started doing the organizational work that makes any tool actually land.

If you’re sitting on a pilot that never shipped, or looking at a new AI capability and not sure how to evaluate it, that’s exactly what Black&Tan Labs works on. The Discovery engagement is a full day in-person that ends with a specific deployment roadmap - not a list of tools to try, a plan for what to actually build and how to make it stick. If that sounds useful, reach out.