· AI
MIT Says 95 Percent of AI Pilots Fail. The Other Five Percent Tell You What to Automate.
The failure rate everyone quotes to argue against automating is also a precise description of which automation to buy.
Someone sends me the 95 percent stat every few weeks, usually to justify waiting another year before touching AI. The number is real. The conclusion people draw from it is not. The same research that produced the scary headline also describes, in detail, what the automation that works looks like, and a small business owner can use that description as a test before spending anything.
My position: run any task you are considering automating through a three-part test drawn from why the failures failed, not from how annoying the task feels. Clear the test and buy the narrow tool built for that job. Fail it and the statistic is describing your project, and you should wait.
The honest case for waiting
State the other side fully, because it holds up. MIT's GenAI Divide report found that about 95 percent of custom enterprise AI pilots show no measurable effect on profit or revenue. Gartner goes further on a narrower category: it forecasts that more than 40 percent of agentic AI projects will be scrapped by the end of 2027, citing escalating costs, unclear business value, and weak risk controls. Gartner analyst Anushree Verma called most current projects "early stage experiments or proofs of concepts that are mostly driven by hype," and said the typical agentic AI pitch still lacks a meaningful return to show for it. Of the thousands of vendors now calling themselves agentic, Gartner counts roughly 130 as the real thing; the rest is what it calls agent washing, old chatbots and RPA tools with a new label. A business owner with no AI team is precisely the buyer least equipped to tell the 130 from the relabeled rest. Sitting this out until the category sorts itself out is not cowardice. For a lot of businesses this year, it is the correct call.
What the failures have in common
But MIT's own data does not say automation failed. It says a specific kind of automation failed, and the pattern repeats across both reports. Partnerships with specialized vendors reached production about 67 percent of the time, versus roughly a third as often for tools built in-house, an outcome VentureBeat traced to what the report calls a learning gap: generic tools don't adapt to a specific workflow, so they stall. Where it did work was unglamorous: back-office automation, cutting outsourcing and agency spending, reportedly saving the companies that got it right between two and ten million dollars a year, per that same reporting. More than half of generative AI budgets went to sales and marketing tools instead, the function both reports associate with the weakest returns.
One more data point worth separating out. VentureBeat's read of the underlying survey found that employees at over 90 percent of the companies studied already use personal AI tools at work on their own, while only 40 percent of those companies had bought an official subscription. The adoption question is already answered, for free, before you sign a contract. What is left to decide is not whether your team wants this. It is whether to formalize the one task they are already leaning on it for.
The actual test
Three questions, asked about one task at a time, not about "AI" in general:
- Is it narrow and repeatable, the same shape of input and output every time, not a judgment call that changes with the customer?
- Can you buy a tool built specifically for that job, instead of hoping a general chatbot adapts to it?
- Is a wrong output cheap and fast to catch, before it reaches a customer or a filing?
A task that clears all three sits in MIT's winning 67 percent. Invoice matching, scheduling confirmations, first-pass document sorting. A task that fails any one of them, something that needs judgment, something built from scratch in-house, something where a wrong answer surfaces weeks later in a customer complaint or a tax return, is exactly the profile both reports describe failing. That is not an argument against ever automating it. It is an argument against automating it yet, or against building it yourself when a narrow vendor tool already exists for that exact job.
Running that test task by task, honestly, before any money changes hands, is the actual content of a real AI readiness assessment, not a maturity score on a slide. If you want a second opinion on which of your tasks would actually clear it, that conversation is worth having before you buy the subscription or the custom build.
Sources
References used in this article. Links also appear alongside the relevant claims.
Let's make it happen.
You don't need a polished brief. A couple of lines about where you want your company to go is plenty, and we'll come back with what we'd tackle first.
We scope the work around your goals and timeline before quoting anything, so you know exactly what you're getting.