» Unfair Advantage
Weekly Deep Dive

Most AI projects fail. The 5% that work share one habit

2026-06-15 · Unfair Advantage Editorial

Here is the number that should reframe how you spend on AI: 95% of corporate AI pilots return nothing measurable. That is the headline finding of MIT's "State of AI in Business 2025" report, which reviewed more than 300 initiatives, ran 52 organisational interviews, and surveyed 153 executives. Only about one in twenty pilots reaches production with value you can point to on the books. The reflex is to blame the technology, but the report is blunt: the gap is not model quality or regulation. It is approach.

The single biggest culprit MIT names is what it calls the "learning gap." Most failed tools do not plug into how work already happens. They sit in a separate window, demand that someone remember to use them, and never absorb the context of a specific job, so they get abandoned within weeks. The 5% that stick share the opposite trait: they are wired into one real, repeated workflow, and they improve as they are used. The lesson for a small team is almost a relief. You do not need a smarter model. You need a narrower job and a tool that lives where the work already lives.

Two more findings cut against the standard advice. First, where the money goes is usually wrong. Roughly half to two-thirds of AI budgets flow to sales and marketing because those pilots are easy to pitch in a meeting. But the report found back-office automation, the dull plumbing of finance, operations and admin, quietly delivers better returns. The flashy use case is rarely the profitable one. Second, build-versus-buy is not close. Buying a focused tool from a specialised vendor or partnering with one succeeds about 67% of the time; building your own in-house succeeds roughly half as often. For a five-person shop with no ML engineers, that is not a tie to agonise over. Buy the narrow tool, point it at the boring expensive chore, and resist the urge to roll your own.

The uncomfortable part is that the 95% were not run by fools. They were run by people who started with the technology and went looking for a problem, picked the exciting use case over the lucrative one, and measured activity instead of outcome. A pilot that "everyone is using" is not a win. A pilot that removed four hours of invoicing a week, and you can prove it, is. The teams that cross to the winning side do something unglamorous: they pick one painful, repeated task, set a number it has to hit, buy a focused tool, wire it into the existing process, and kill it on a deadline if it misses. None of that requires being early or technical. It requires being disciplined about the problem rather than dazzled by the tool.

So what does a small team do this week? Stop running open-ended experiments. Name one chore that costs you real hours or money, attach a target to it, and treat AI as an investment with a hurdle rate, not a subscription you feel obliged to have. The firms on the right side of this divide are not spending more. They are spending on purpose.

Why it matters

If you feel pressure to buy more AI because everyone else is, this is the data that says slow down. The teams that get returns are not spending more or moving faster; they pick one painful repeated task, buy a focused tool, wire it into existing work, and hold it to a number. That is a playbook a five-person shop can actually run.

Network impact

LatencyNo direct impact, though the report's core lesson reinforces it: a tool only gets used if it fits the fast workflow people already have. Bolt-on tools in a separate window get abandoned.
SecurityBuying focused tools from specialised vendors instead of building in-house shifts more responsibility onto the vendor's security posture. Check their data handling before you wire them into a core process, and never let an unsupervised tool touch payments or customer data.
ScalabilityPicking one narrow, integrated workflow over a sprawling general pilot is exactly how small teams scale AI without a graveyard of abandoned projects. Narrow and embedded scales; broad and bolted-on does not.

What to do

  1. Name one painful, repeated chore (invoicing, support triage, scheduling) before you look at any tool. Start with the problem, not the technology.
  2. Attach a target to it in one sentence: this must save X hours or add $Y by a set date.
  3. Look at the boring back office, not just sales and marketing — that is where the report found the better returns hide.
  4. Buy a focused tool from a specialised vendor rather than building your own. It wins roughly twice as often for teams without ML engineers.
  5. Wire the tool into the workflow people already use. If it sits in a separate window they have to remember, it joins the 95% that get abandoned.
  6. Set a kill date. If the pilot misses its number by the deadline, cancel it. Measure outcomes, not how many people are 'using it'.

Sources

« All articles