AI Won't Fix What You Haven't Fixed Yet

Why Most AI Projects Fail (And What to Fix Before You Build)

I was watching a video on the Stripe channel last week where Sundar Pichai sits down with John Collison to talk about Google's history with AI.

Every few minutes something lands that you need to sit with.

One moment made me rewind.

Collison asked Sundar why Google invented the transformer architecture but didn't ship the product that made it famous. Sundar pushed back. The framing, he said, was a bit misunderstood. Transformers weren't a research exercise sitting in a lab waiting to become a product. They were built to solve a specific problem inside Google's own products. Translation. Search quality. The architecture went straight into BERT and MUM, and some of the biggest improvements in search quality during that period came directly from those models running quietly underneath, doing real work on real problems.

ChatGPT came later and wrapped the same underlying idea in a chat window. That became the product most people encountered. And that became the mental model most companies carried into their own AI investments.

I've been thinking about how much damage that mental model has quietly done.

When "we should be doing AI" becomes the whole strategy

Most companies that started building with AI in the last two years began from the same place.

Someone in leadership said they needed to be doing something with AI. A team got assigned. Tools got evaluated, vendors got shortlisted, a pilot got scoped. And somewhere in all of that motion, nobody stopped to ask what question they were actually trying to answer.

Not in a vague strategy sense. In a specific, someone-could-check-the-answer sense.

AI doesn't supply that clarity. It runs on whatever intention you bring to it. And if the intention is "we should be using AI," that is exactly what it optimizes for. The appearance of progress. Dashboards that look good in a demo. Outputs that are technically correct and completely disconnected from any decision that actually matters.

S&P Global's 2025 research found that 42% of companies abandoned most of their AI initiatives that year, up from 17% the year before. The average organization scrapped nearly half its AI proofs of concept before reaching production. In most cases the model wasn't what failed. Nobody had fixed what was underneath before asking AI to build on top of it.

AI inherits whatever it sits on

This is the part that doesn't get said enough.

If the data is fragmented, AI works on fragmented data. If goals are unclear, it optimizes for unclear goals. If a process was already losing value before AI arrived, it will run on that process faithfully and at scale.

A 2024 survey of 500 enterprise data leaders found that 73% named data quality as their primary barrier to AI results. Ranked above model capability. Above computing costs. Above talent.

The tools are not what's failing. What's failing is what the tools are being asked to run on.

And the fix for that isn't a better model or a bigger budget. It's the slower, less exciting work of understanding what data you actually have, what it does and doesn't capture, and what question you are genuinely trying to answer with it.

Most organizations skip that part. Not because they don't know it matters. Because it doesn't show up in a demo.

The signal nobody acts on

Here is where it gets harder to talk about honestly.

Some companies do fix the foundation. They clean the data, define the workflows, align on what success looks like. AI starts producing something real. A churn pattern that wasn't visible before. A deal risk signal three weeks earlier than the team would have caught it. An honest read on where the product is losing people before they say anything.

And then the meeting happens the way it always happened.

The senior person goes with what they know. The team gravitates toward the interpretation that fits the story they already had. The output sits in a dashboard. The next decision gets made on instinct and experience and whoever speaks most confidently about what they already believe.

AI is only as useful as the organization's willingness to be wrong. And most organizations, even the ones that built something real, haven't prepared for that part at all.

This isn't stubbornness. It's what happens when an answer arrives fully formed and nobody worked through the reasoning to produce it. You don't own what you didn't think through. And you don't act on what you don't own.

Sundar described something in that conversation that shows the difference. He talked about how he uses Google's internal AI tool and asks it things like: what did people think about this thing we just launched, give me the five worst reactions. He wasn't using it to confirm what he already believed. He was using it specifically to find out what he didn't know and couldn't see from inside his own position.

That posture is rarer than most AI roadmaps account for.

Intention runs deeper than infrastructure.

McKinsey's 2025 AI survey found that organizations reporting real financial returns from AI were twice as likely to have redesigned their workflows before selecting any modeling techniques. Not after. Before.

The companies getting something real tend to be the ones who knew what they were for before AI showed up. Who had enough clarity about the problem they were solving that an honest signal had somewhere to land and someone willing to act on it.

There is something uncomfortable underneath that observation.

If the goal coming in was genuine, finding where customers were not getting value, understanding where the process was leaking, figuring out what was actually causing churn, AI has a real question to work with. It can surface things that would have taken months to see otherwise.

But if the goal was to cut costs, or generate more output from fewer people, the system will optimize for that too. Customers feel that signal before the metrics do. And when the metrics eventually catch up, the usual response is to build more on top of the same broken intention.

What Sundar was actually saying

Transformers were built to improve something specific. To answer a question that mattered for real users at real scale. That intention is what made BERT and MUM work inside Google's products years before most people had heard the words large language model.

That same clarity is what is missing in most of the AI projects failing quietly right now.

Not the model. Not the infrastructure. Not even the data, though that matters too.

The clarity about what question is actually being asked, and whether the people asking it are genuinely prepared to hear the answer.

The technology has been ready for that conversation for a while. The question worth sitting with is whether the organizations building on top of it are.