If you read enough AI coverage, you will see two stories at once. Tools are getting better every month, and most companies are not seeing money from them. Both are true. The gap between them is where owners lose time and budget, so it is worth understanding before you spend anything.
What the failure numbers actually say
The figure that got the most attention came from an MIT NANDA report in 2025. It found that about 5% of AI pilots reached rapid revenue acceleration, and the other 95% stalled with little measurable effect on the profit and loss statement.[1] The report drew on 150 leader interviews, a survey of 350 employees and a review of 300 public deployments.
Read that number carefully. "Failure" there means no measurable P&L impact, not a pilot that broke. A project that saves a bookkeeper four hours a week would not show up in that count. The report also came to us secondhand through press summaries, and companies are often reluctant to share failures. Treat 95% as a direction, not a measurement.
Other sources point the same way. McKinsey's 2026 survey of 1,719 respondents in 97 countries found that 37% attribute at least some EBIT impact to AI, roughly flat from the year before.[2] Gartner predicts that more than 40% of agentic AI projects will be canceled by the end of 2027 because of rising costs, unclear business value or weak risk controls.[3] That is a forecast, not a count, and it comes from an analyst firm that sells advice. Still, the three sources agree on the cause: the technology is rarely what fails.
What the winners do differently
McKinsey separates out a group it calls high performers, about 6% of respondents. Nearly three-quarters of them report fundamentally redesigning workflows around AI, compared with about a quarter of everyone else.[2] The MIT report points to a "learning gap": generic tools that do not fit how the business actually works stall once the novelty fades.[1] Put together, five habits show up.
1. They start with a workflow, not a tool
A weak project starts with "we should use ChatGPT for something." A strong one starts with "quotes take us three days and we lose jobs to faster competitors." The workflow comes first because it tells you what the tool must do, who touches the work and what a good result looks like.
2. They pick one number
Every workflow gets one measure: hours per week, days to send a quote, percent of invoices paid on time, calls answered. If you cannot say which number should move, you cannot tell whether it worked. Capture the baseline before you change anything.
3. They buy where a product exists
The MIT summary reported that tools bought from specialized vendors, and partnerships, succeeded about 67% of the time, while internal builds succeeded about a third as often.[1] The article does not explain the mechanism, and "success" was defined narrowly, so do not over-read it. The practical lesson is modest: building from scratch is slower and harder to maintain than most owners expect.
4. They put a person at the gate
Only 22% of small business owners in the Bluevine survey said they were completely confident AI could handle low-level tasks without supervision.[4] They are right to hesitate. Anything involving money, commitments or sensitive messages should stop at an approval step until the error rate is known.
5. They keep owning it after launch
A workflow that nobody reviews drifts. Inputs change, vendors update their models, staff find workarounds. Name one person responsible, and look at the numbers monthly for the first quarter.
Warning signs your project is heading the wrong way
- Nobody can state the problem in one sentence without naming a tool.
- The success measure is "adoption" or "usage" instead of a business number.
- The vendor demo used clean sample data, and nobody has tried your messy real data.
- There is no one assigned to review outputs.
- The plan depends on staff "figuring it out."
Before you start, answer four questions
- What specific workflow are we changing, and who does it today?
- Which one number should move, and what is it today?
- What is the worst thing that could happen if the AI gets it wrong, and where is the approval gate?
- Who owns this in 90 days?
If you cannot answer all four, you are not ready to buy anything. That is a useful finding, and it costs nothing. Working it out is the first part of an assessment.
Sources
- Fortune, summary of MIT NANDA "The GenAI Divide: State of AI in Business 2025"
- McKinsey, "The state of AI in 2026: On the road to ROI"
- Gartner press release, June 25, 2025
- Bluevine, "2026 Small Business AI Trends Report"
Figures are quoted from the sources above and were reviewed in October 2026. Where a source is a secondary summary of other research, we say so in the text. Survey results reflect each survey's own sample and definitions. Examples labeled as illustrations use made-up numbers to show the method. Nothing here is a forecast or a guarantee of results.
Want this worked out for your business?
Tell us what is wasting your time.