AI-powered startups are being built faster than ever, yet most will never make it past their first year.

Despite unprecedented access to powerful models, funding and development tools, around 95% of generative AI pilots at companies are failing. The reason is rarely the technology itself. More often, founders become so focused on what AI can do that they lose sight of the problem it is supposed to solve.

Many fall into the trap of viewing AI as a silver bullet, a technology that can turn a good idea into a scalable, profitable business almost overnight. But the reality is far more complex. Success does not depend on how advanced the model is or how many datasets it can process; it depends entirely on whether it solves a real, pressing problem for real people.

The fastest way to waste money on AI is to start with the technology rather than the problem, says Jordan Richards, founder and CEO of AI design-and-build studio &above .

“He says: Companies get excited about models, automation and novelty, then forget to ask the only question that really matters: ‘What measurable outcome are we improving?’ Most AI projects fail because they're built around what the technology can do rather than what the business needs to achieve.”

A common pitfall that Richards observes is what he calls the ‘human design gap’, a disconnect that appears when organizations prioritize technical specifications over how people actually work and make decisions.

“AI doesn't create value when it's switched on,” he says. “It only creates value when it changes behaviour, speeds up decisions, reduces friction, or improves customer outcomes. If it doesn’t do that, it’s just expensive theatre.”

One of the biggest myths in AI is that adoption will happen naturally if the technology is good enough. Intelligence alone is not a product strategy. “In a world where anyone can write thousands of lines of code, intentional design becomes more important to build the right systems that drive desired behaviours,” says Richards. “Only then can AI truly achieve high-value outcomes for the user.”

Building AI Before Validating The Problem

One critical mistake that startups make is investing time and capital into building AI before fully validating the problem that needs to be solved. The right question is not ‘what can AI do here?’ but rather, ‘is this challenge high value enough that people will actually change how they work to fix it?

When Dawn Barclay-Ross launched Fund Expo earlier this year, she didn’t set out to build a business that was 90% AI-powered. Her objective was to stop ‘watching good founders slowly bleed out while the funding industry pats itself on the back’. She realized that the only way to fix the problem at scale was to put AI in the engine room, and along the way, she made the classic mistake of building AI before fully validating the problem.

“On paper, it looked elegant,” she says. “Feed in a business profile, sector, and a few numbers, and get a neatly ranked list of funding options and a recommended path. In reality, founders are messy. Their businesses are tangled up with their personal finances, their families, and their psychology. They want to know how a decision affects the school run, the mortgage, and their marriage.”

The early prototype produced routes that were technically sound, but emotionally unworkable. “Raise this much equity, take this debt, here’s the IRR,” adds Barclay-Ross. “On calls, I could hear the pause. You could almost feel the thought: ‘That might make sense on a spreadsheet, but there is no way I’m doing that.’ So, we scrapped it.”

Instead of obsessing over models, her team started obsessing over conversations. They interviewed founders, funders, accountants, and lawyers, and mapped the real journeys; the zigzags, the near misses, the bad advice, the lucky breaks. Once they had the messy reality laid out, they started rebuilding the AI around it.

“The lesson we learned was that if your AI can’t cope with the human mess, it will be impressive demoware and a useless product,” she says. “The problem needs to be defined in human terms first, technical terms second.”

Treating AI As The Product, Not The Outcome

Barclay-Ross also admits she fell into a second, equally damaging trap: treating AI itself as the product, rather than as a tool to deliver a clear outcome. “Founders don’t wake up thinking, ‘I’d love some AI today,’” she says. “They wake up thinking, ‘How do I pay my team? How do I get this investor to take me seriously?’”

Once that perspective shifted, the company repositioned what it did. Fund Expo is not an AI company; it is a funding and growth ecosystem that happens to be 90% powered by AI behind the scenes.

“Instead of saying, ‘Our AI can process X datasets,’ we say, ‘In one afternoon, we can show you three realistic funding paths, the trade-offs on each, and what you need to do in the next 90 days’,” says Barclay-Ross. “The AI is essential to making that possible at scale, but it’s not the hero of the story. AI should be the invisible plumbing that lets you keep your promises to customers, not the promise itself.”

Removing Humans From The Loop Too Early

Startups often rush to automate entire processes from day one, believing that full automation equals efficiency and scalability. But trust is built gradually, usually when AI augments human decision-making first, then scales from there. Going straight to full automation before that trust exists is where implementations often break down.

Mental health support service Wobble was originally founded in 2024 as a traditional therapy marketplace, so the founder, Jack Murphy, had already spent time in this sector before launching the current product in May this year. He had previously created a product that delivered mental health support through AI, but had shut it down because he didn’t feel comfortable with it.

The decision to incorporate AI into Wobble came at the start of a pivot to a new format of mental health support: personalized human support delivered in minutes or hours rather than days or weeks, in a form accessible to people who would typically never seek help.

Murphy understood that speed and accessibility were key, but so was credibility and safety. Lacking the financial resources for a large team, he researched everything he could about where AI could stand in for one, running three rounds of consumer research before launching in May.

He says: “In the findings, 96% of people said the response coming from a real human rather than AI was essential or very important to them, crucially showing exactly where AI must never sit; anywhere near the therapeutic response itself."

However, that didn’t stop him from deploying AI aggressively everywhere else. He uses Claude Code to build and maintain the platform, while Claude runs the research analysis and most of the day-to-day admin. Beyond the founder, the team is one person: a clinical lead who oversees the clinical and ethical side.

“AI is effectively my engineering team, my data analyst and my back office,” says Murphy. “It just never speaks to a user. Advice doesn't carry weight purely on whether it's accurate; it carries weight because of who it comes from. Take the human away and the same information, right or wrong, lands with a fraction of the weight.”

Building AI Around Human Needs

The companies generating meaningful returns from AI are not necessarily the most technically advanced, but they are some of the most disciplined when it comes to solving real problems, building for the user, and growing team-wide capabilities.

Richards sums it up: “Long-term value in the AI era comes from building AI around enduring human needs to create measurable improvements for customers, employees and your bottom line. Technology changes quickly, but human needs don't. Building for the second will enable you to adapt to the first.”

For startups, the challenge is not building the most sophisticated AI. It is building something people genuinely need. Models will evolve, costs will fall, and new capabilities will emerge. The businesses that endure will be the ones that remember what many founders forget: that technology changes quickly, but human needs do not.