If you run a business, you have probably seen three stories that seem to say the same thing. An MIT study found that 95% of company AI pilots deliver no measurable return. Klarna, the payments company that said its AI assistant was doing the work of 700 customer service agents, started hiring people again. And McKinsey, the firm large companies pay to advise them on AI, is reportedly planning to cut a few thousand of its own back-office roles.
Put those together and it is easy to land on one of two conclusions. Either AI is overhyped and you can safely ignore it, or it is moving so fast that you are already behind. Both are wrong, and the reasons are sitting in the details of those same stories.
What Did the MIT Study Actually Find?
MIT's NANDA initiative found that only about 5% of corporate generative AI pilots produced a measurable impact on profit. The report, The GenAI Divide: State of AI in Business 2025, drew on interviews, surveys and a review of about 300 public AI deployments, and was first reported by Fortune in August 2025. The 95% figure has been repeated in boardrooms and on LinkedIn ever since.
Two details rarely travel with it.
The bar was high. A pilot only counted as a success if it moved past the trial stage and showed measurable results within roughly six months. Plenty of worthwhile projects take longer than that, and some benefits, like a team that answers customers faster, never appear as a tidy line on a profit report.
It is a research report, not a verdict on AI. It is an industry study of how companies run AI pilots, not a peer-reviewed measurement of whether the technology works. That does not make it wrong. It makes it a strong signal about how companies are going about AI.
The most useful findings sit further down the page. Companies that bought AI tools from specialist providers succeeded about 67% of the time, while internal builds succeeded only about one-third as often. And although more than half of AI budgets went to sales and marketing tools, the biggest returns showed up in back-office work: cutting outsourcing, reducing agency costs and streamlining operations.
In other words, the 95% is less a story about AI failing and more a story about where companies chose to point it.
Is AI Failing, or Are Companies Failing at AI?
The report's lead author, Aditya Challapally, put the problem down to a "learning gap". The technology largely works. What does not work is dropping a general-purpose tool into a business and expecting it to fit how that business actually runs. A chatbot is great at helping one person draft an email. It struggles when the job involves your pricing rules, your customer history and the three exceptions everyone in the office just knows about.
McKinsey's 2025 State of AI survey points the same way. It found 88% of organisations now use AI in at least one part of the business, yet only about 6% qualify as "high performers" that attribute more than 5% of their earnings to AI. What sets that small group apart is not a cleverer tool. They are far more likely to have redesigned how the work gets done around AI, instead of bolting AI onto the old way of working.
If you have read our piece on operational debt, this will sound familiar. Technology makes a clear process faster and a messy one more expensive. AI simply does both at greater speed.
What Happened at Klarna, and Why It Matters
Klarna became the headline example of AI replacing people after announcing in 2024 that its AI assistant was doing the work of 700 customer service agents. By May 2025 its chief executive, Sebastian Siemiatkowski, was telling Bloomberg the company was recruiting people again. His explanation was unusually candid: "As cost unfortunately seems to have been a too predominant evaluation factor when organizing this, what you end up having is lower quality."
He added that it was "so critical" customers know "there will be always a human if you want."
The lesson is not that AI customer service cannot work. It is that Klarna judged the project on one number, cost, and lost ground on the thing it was not watching: how customers felt. Every business planning an AI project should ask the question that got skipped. Besides savings, what are we going to watch?
Why Is McKinsey Cutting Jobs?
In December 2025, Bloomberg reported that McKinsey leaders were planning to reduce headcount by about 10% in some non-client-facing teams, potentially a few thousand roles, phased in over 18 to 24 months. The firm said it wanted to make its support functions more efficient, reflecting the same changes it advises clients to make.
Set the headline aside and look at how the move is shaped. It starts in the back office, not with the people clients see. It is gradual, spread over two years rather than announced overnight. And it is aimed at specific functions, not a sweeping promise that AI will run the firm.
That is almost exactly where the MIT research says returns show up first.
Are Small Businesses Behind on AI?
Not as far as the headlines suggest. According to the US Census Bureau, about one in five US businesses (19.8%) were using AI as of early May 2026.
Use climbs with size: 37% of firms with 250 or more employees used AI, compared with fewer than 20% of firms with four or fewer employees, and use among businesses with fewer than 20 employees did not change significantly between December 2025 and May 2026. It varies by industry too, from 39.7% in information businesses to around 14% in retail.
McKinsey's survey shows a similar gap among companies that already use AI: nearly half of those with more than $5 billion in revenue have begun scaling it across the business, compared with 29% of those under $100 million.
If that makes you feel late, look at it from the other side. The biggest companies are further along, and they are also the ones funding many of the pilots behind MIT's 95%. The next wave looks similar. Gartner expects more than 40% of "agentic" AI projects, the systems that take actions on their own rather than just answering questions, to be cancelled by the end of 2027 because of rising costs, unclear business value or weak risk controls. Its analyst Anushree Verma described most current projects as "early stage experiments or proof of concepts that are mostly driven by hype and are often misapplied."
A business that has not spent heavily yet gets to skip that expensive lesson. Going second is an advantage, as long as you actually learn from the companies that went first.
What the Successful 5% Do Differently
Across the MIT, McKinsey and Klarna stories, the same few habits keep appearing. None of them needs a technical team.
- They start with a problem, not a tool. "We want to use AI" is not a goal. "Quotes take three days and we lose deals while customers wait" is.
- They look at the back office first. Invoicing, scheduling, data entry, reporting and paperwork are less exciting than a chatbot on your website, and that is exactly where MIT found the biggest returns.
- They buy before they build. Tools from specialist providers succeeded far more often than homegrown ones in the MIT research. Unless AI is your product, building your own is rarely the place to begin.
- They put the people closest to the work in charge. MIT found adoption went better when line managers, not only a central technology team, drove it. The person who does the task every day knows where the exceptions hide.
- They decide what success looks like before they start. Name the number you expect to improve and the number you refuse to let slip. Klarna measured the first and not the second.
- They keep a human where the customer can feel it. Automate behind the scenes freely. Be deliberate about anything a customer hears, reads or waits for. It is the same "assist first, act later" line we draw in our guide to choosing a CRM with AI agents.
A 30-Minute AI Reality Check for Your Business
Before you sign up for anything, sit down with your leadership team and answer these honestly.
- Which task costs us the most time each week, and who actually does it?
- If AI handled half of that task, how would we know? Which number would change?
- What must not get worse while we try: customer experience, accuracy, compliance?
- Who will own this, the person doing the work or someone who only manages it?
- Does a tool already exist for this, before we consider building anything?
- What will we do if nothing has improved after 90 days?
If you cannot answer the second and third questions, you are not ready to buy a tool yet. You are ready to get clearer on the problem, which is cheaper and more useful anyway.
What to Do in the Next 90 Days
Keep it small and specific.
- Pick one back-office task that runs often, follows mostly predictable rules and does not face customers directly.
- Write down today's baseline: how long it takes, how often it runs and how often it goes wrong.
- Trial one existing tool for 90 days, with the person who does the work in charge of the trial.
- Review honestly at the end. Keep it, adjust it or drop it, and write down why.
If you want a practical head start on choosing that first task, our free field guide to getting real work out of AI walks through it step by step. One well-run 90-day trial will teach your business more about AI than a year of reading headlines, this one included.
The companies in the news are not proof that AI does not work. They are proof that AI rewards the same discipline as any other business investment: a clear problem, a measured result and someone accountable for both. Most teams can run that first trial themselves. When the question gets bigger, such as where AI fits across the whole business or which vendor to trust, that is the work of our AI strategy practice and exactly the kind of decision management consulting is for.
FAQ
Is it true that 95% of AI projects fail?
Not quite. MIT's NANDA initiative reported in 2025 that only about 5% of corporate generative AI pilots showed a measurable impact on profit, using a demanding definition of success: the pilot had to move beyond the trial stage and show measurable results within roughly six months. The research drew on interviews, surveys and a review of about 300 public AI deployments. It shows most companies are struggling to turn AI pilots into results. It does not show that AI itself does not work.
Why do most AI projects fail to deliver results?
Mostly for business reasons rather than technical ones. The MIT research pointed to a learning gap: generic tools do not adapt to how a specific business works. Companies also tended to spend on sales and marketing tools when the biggest returns were in back-office work, to build their own tools when bought ones succeeded far more often, and to measure projects on cost alone. McKinsey's research found the small group getting real value were much more likely to have redesigned how the work gets done around AI.
Should a small business still invest in AI?
Yes, but small and specific beats big and vague. According to the US Census Bureau, about one in five US businesses (19.8%) were using AI as of May 2026, so most smaller firms are not behind a crowd. Start with one frequent back-office task, record how long it takes today, trial an existing tool for 90 days with the person who does the work in charge, and keep, adjust or drop it based on the result.
Why did Klarna bring back human customer service agents?
Klarna had said in 2024 that its AI assistant was doing the work of 700 customer service agents. In May 2025 its chief executive, Sebastian Siemiatkowski, told Bloomberg that cost had been too predominant a factor and the result was lower quality, and that customers should always be able to reach a human if they want. The lesson is to measure customer experience alongside savings, not that AI customer service cannot work.
Where should a small business start with AI?
Start with a problem, not a tool. Pick one back-office task that runs often, follows mostly predictable rules and does not face customers, such as invoicing, scheduling, data entry or reporting. Decide in advance which number should improve and which must not get worse, give ownership to the person closest to the work, and prefer an existing product over building your own.
Sources
- MIT NANDA, The GenAI Divide: State of AI in Business 2025, as reported by Fortune, August 18, 2025
- McKinsey, The State of AI in 2025: Agents, Innovation, and Transformation
- US Census Bureau, "Large Firms With at Least 20 Employees Biggest AI Users", May 26, 2026
- Gartner, "Over 40% of Agentic AI Projects Will Be Canceled by End of 2027", June 25, 2025
- Bloomberg, "McKinsey Executives Plot Job Cuts in Slowdown for Consulting Industry", December 15, 2025
- CX Dive, reporting Klarna CEO interview with Bloomberg, May 9, 2025
*Published by EncubIQ Consulting | Last Updated: September 2026*