On this pageJump to a section
A headline can travel a long way on one number. This week, the number was 95%.
An MIT report on generative AI in business landed, Fortune covered it on August 18, and within days the takeaway at every dinner table was "AI pilots fail." Reasonable people nodded. Others used it to say AI is hype. Both groups skipped the actually useful part, which is why, because the reasons are mostly choices, not technology.
And choices are exactly what a small operator controls.
What did the MIT report actually say?
It is called The GenAI Divide: State of AI in Business 2025, from MIT's Project NANDA. Its executive summary says that despite $30 to $40 billion of enterprise investment in generative AI, 95% of organizations are getting zero return, and that just 5% of integrated AI pilots are extracting millions in value while the vast majority remain stuck with no measurable P&L impact.1
5%
How solid is it?
Be careful here. The report calls itself preliminary. It is based on a review of over 300 publicly disclosed AI initiatives, structured interviews with representatives from 52 organizations and survey responses from 153 senior leaders. It also notes that some comparisons reflect self-reported outcomes and may not account for confounding variables.1 News coverage described the sample in slightly different terms, so read the report itself if you plan to quote it.2
I would treat "95%" as a loud signal that something is off in how most organizations deploy AI, not as a measured law of nature.
What does the report say is going wrong?
Not the models, according to the authors. The primary factor they name is the learning gap: tools that don't learn, integrate poorly or fail to match workflows.1
Its executive summary adds that most enterprise-grade tools fail because of "brittle workflows, lack of contextual learning, and misalignment with day-to-day operations."1
This divide does not seem to be driven by model quality or regulation, but seems to be determined by approach.
That is good news. Approach is the part you control.
Why do pilots stall, in operator terms?
When I read the report with a small business in mind, three failure modes stand out. The first is closely tied to what the authors found about evaluating tools on business outcomes. The other two map to their findings on workflows and ownership.
Failure one: a vague goal
"Let's try AI for marketing" is not a goal. It cannot succeed or fail, so it does neither, and the pilot drifts into a permanent "we're exploring."
The report notes that buyers who succeed evaluate tools on business outcomes rather than software benchmarks.1 For a small operator, that means a goal you can state in hours or dollars: "cut the time to produce a weekly client report from three hours to one" or "answer new-lead inquiries inside 10 minutes."
Failure two: no change to the workflow
A tool dropped on top of an unchanged process mostly adds a step. The report describes systems that do not integrate with existing processes as a main reason pilots die, and says winners demand process-specific customization.1
If you add AI to a task, decide what else disappears, and write the process down first, as in Your First AI Hire Is a Writer. Which step is replaced? Which handoff is removed? If nothing is removed, you have added work.
Failure three: no owner
The report's discussion of organizational design is the part most worth reading. It says success depends less on resources and more on decentralizing authority while keeping clear ownership and accountability.1
In a large company, that means a line manager who is accountable for a specific result. In a small business it means a single name. If everyone is "kind of in charge" of the AI experiment, no one is.
Why can small operators win?
Because the things that slow big companies are things you do not have.
Short loops
You can decide on Monday, run the test on Tuesday and know by Friday. A large organization may need a committee, a procurement cycle and a security review before the first prompt.
One owner who is also the user
The person who feels the pain is the person deciding. The report praises decentralized implementation with clear accountability, and a solo operator is the extreme version of that.1
Buy, then customize, instead of building
In the report's sample, external partnerships with learning-capable, customized tools reached deployment about 67% of the time, compared with about 33% for tools built internally. The authors flag that these are self-reported figures.1
For a business of one to ten people, that is a strong reason to start with proven tools and adapt them to your process rather than building from scratch.
Look at the back office first
The report found that GenAI budgets lean toward visible sales and marketing uses while back-office automation often yields better ROI.1 That matches how I see small businesses work: the quiet jobs, such as document handling, scheduling, follow-up and data entry, are often where hours disappear (and clean inputs matter, per Get Your Data Ready for AI).
How do you run a pilot that can actually succeed?
- Pick one job. Narrow and repeated, such as a weekly report or lead follow-up.
- Time the current way. Write down how long it takes and what it costs today.
- Define success in one sentence, with a number and a date.
- Name the owner. One person, with authority to change the process.
- Change the workflow. Remove a step, not just add a tool.
- Add a human check on anything that leaves the building. Anthropic's Project Vend (I broke it down in Project Vend) showed how confidently an AI can give away margin when nobody is checking.4
- Review at 30 days. Compare real time, including checking, against the baseline. Keep it, change it or kill it.
Measure honestly
Feeling productive and being productive are different. METR's randomized trial of experienced open-source developers found tasks took 19% longer with AI allowed, while the developers believed they were faster.3 That study covered a specific setting, but the principle is universal: time it.
The Y Combinator partners made a related point about AI startups on their Lightcone podcast, in a video posted on May 30, 2025: "evals are the true crown jewel like data asset for all of these companies."5 An eval is a written test of what good looks like for one job. In a small business, that is your baseline and your one-sentence definition of success, kept where you can check it.
VideoY Combinator
State-Of-The-Art Prompting For AI Agents
The hosts walk through a real production prompt and what founders learned about evals. Technical in places, but the evals section is plain.
What is my takeaway?
A pilot fails when nobody can say what success means, nothing about the work changes and nobody is responsible. Fix those three and you are already ahead of most of the 95%.
Running my business without staff makes every tool answer one question: does it remove work I would otherwise do, or does it just add a dashboard? Ask the same question of every pilot you start. It is also the question behind the AI-assisted systems I build for clients.
AI doesn't fail pilots. Unowned goals fail pilots.
Keep building.Ekai
Questions people ask
Why do most AI pilots fail?
According to MIT's NANDA report, the core barrier is a learning gap: tools that do not retain feedback, adapt to context or fit existing workflows. Organizational design and ownership also separate the few that scale from the many that stall.
Is it true that 95% of AI projects fail?
The report says 95% of organizations are getting zero return and that just 5% of integrated pilots are extracting millions in value. It is preliminary, based on interviews, a survey and a review of public initiatives, and its outcomes are largely self-reported, so treat 95% as a signal, not a law.
Should a small business buy AI tools or build its own?
In the report's sample, externally partnered, customized tools reached deployment about 67% of the time versus about 33% for internal builds, though the authors note those figures are self-reported.
What is the first step in a successful AI pilot?
Pick one narrow job, measure how long it takes today, name one owner and define what success looks like in hours or dollars before you start.
How do I know whether an AI tool is actually saving time?
Time the task before and after, including checking and fixing the output. A randomized METR study found experienced developers took 19% longer with AI tools while believing they were faster.
Sources
- 1
- 2
- 3
- 4
- 5
