Here is the number that should reframe how every CEO thinks about AI. In 2025, MIT's Project NANDA studied 300 public AI deployments, surveyed 350 employees, and interviewed 150 leaders. The finding, reported by Fortune, was blunt: 95% of organizations got zero measurable return from generative AI. Only 5% saw real impact on the bottom line.
The instinctive reaction is to blame the technology. That instinct is wrong. The MIT researchers were explicit — the failure is almost never the model. It's data readiness, workflow integration, and the absence of a defined outcome before anyone starts building. In other words, the models work. The way companies aim them does not.
I want to make a specific, practical argument about that aiming problem — one I rarely see stated plainly. Most companies are putting AI in the wrong half of the business. They start with internal functions. The money, and frankly the easier wins, are in the services. Let me show you why, and why flipping that order is the single highest-leverage decision a non-technical CEO can make this year.
Two halves of every company
Look at any company through the lens of management theory and its work splits into two unequal halves.
The first half is the work that creates value for the customer — delivering services and producing goods someone outside the company pays for. This is where roughly 80% of a company's effort and resources should sit, because it is what brings in money.
The second half is the work that keeps the company itself running — logistics, marketing, accounting, legal, IT, right down to catering for staff. The customer never sees it, but the company can't function without it.
This isn't a new idea. It's the spine of Michael Porter's value chain from Competitive Advantage (1985): primary activities that create and deliver value to the buyer, and support activities that enable the rest. For the length of this article, let me use shorter words. I'll call the first half services and the second half functions.
The sharpest way to tell them apart is one question: is there an external paying customer at the other end? A service is always activity on a customer's request — there is an outside client whose money and satisfaction are on the line. A function is internal housekeeping: analytics, monitoring, logistics, bookkeeping. Useful, necessary, but the company is its own customer.
It's a simplification. It's also more than enough to explain the trap most companies are walking into with AI.
Where AI actually goes today — and why
Watch where AI is landing in real companies and a clear pattern appears: it goes into the functions.
The data backs this up. McKinsey's State of AI 2025 found that 88% of organizations now use AI in at least one business function, up from 78% a year earlier — and the most common homes for it are IT operations, marketing and sales, customer service operations, knowledge management, and product development. Content factories for marketing. Lead generators. Operations bots. Chatbots for the executive team. AI copilots for logistics. The internal back office is where the tools cluster.
I think this happens for two understandable reasons. Many CEOs simply don't want to risk testing AI on their actual customers — the back office feels safer. And the market is flooded with excellent function-shaped tools from AI vendors, so that's what gets bought. Let me be clear: I'm not criticizing anyone, and these solutions are usually technically sound and genuinely impressive.
But here is the catch the adoption numbers hide. McKinsey also found that only about a third of organizations have begun to scale AI, and no single function exceeds roughly 10% fully-scaled deployment. High adoption, low transformation. Everyone is piloting in the back office; almost no one is getting it to pay off. That is the 95% problem wearing a different outfit.
Why starting with functions is a deceptive kind of easy
Starting with internal functions looks like the safe, simple path. I'd argue it's deceptively hard, for five reasons — and the recent data turns each one from opinion into evidence.
1. Functions need heavy fine-tuning. Off-the-shelf AI is generic; every company's internal processes are idiosyncratic. The work isn't installing the tool, it's bending it to a workflow no vendor has seen. That's exactly the "workflow integration" gap MIT named as a top cause of failure.
2. The feedback loop is brutally long. You cannot tell whether a content factory is working in a week. You need a month, often a quarter, before the numbers mean anything. That delay is fatal to momentum — Gartner predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025. As Gartner's Rita Sallam put it, "executives are impatient to see returns on GenAI investments, yet organizations are struggling to prove and realize value."
3. It's a black box to the people deciding. Many CEOs don't have the technical background to judge whether a proposed internal AI solution is actually any good. You're asked to approve something you can't fully evaluate, aimed at a process whose results won't be clear for months. MIT's phrasing — no defined outcome before the build starts — is the polite version of this.
4. There's no interested customer at the bottom. A function tool is imposed from the top by management, onto staff who didn't ask for it. The result is predictable, and the numbers are sobering: a 2025 survey found 31% of employees — and 41% of Gen Z — admit to quietly sabotaging their company's AI strategy, by refusing to use the tools or ignoring their output (Workplace Insight). Nearly half of CEOs report their employees are resistant or even hostile to AI. When roughly 70% of change initiatives already fail on pushback, top-down function rollouts are walking into the headwind.
5. There are no standards for functions. Internal processes are rarely documented to the level AI needs, so what happens in practice is backwards: the company rebuilds its process around the tool instead of teaching the tool its process. That reorganization is expensive and demoralizing — and it's a big part of why S&P Global found the share of companies abandoning most of their AI initiatives before production jumped from 17% to 42% in a single year, scrapping an average of 46% of proofs of concept.
Why services should be step number one
Now flip it. Start AI with a service the company actually sells, and every one of those five problems softens or disappears.
1. The spec already exists — in someone's head. Even if a company has no written standard for a service, the person who delivers it to customers every day knows it cold: every stage, every checkpoint, every decision fork, every expected result. That tacit expertise is the specification you've been missing. You don't have to invent the process; you have to extract it.
2. You get instant expert feedback. That same employee, even with zero technical skill, can look at a proposed AI solution and tell you in minutes — not months — whether it fits the real service or not. The feedback loop collapses from a quarter to a conversation. The single most expensive thing about function projects, the wait, simply goes away.
3. There's no black box. When the task is a service you understand, the work is honest and visible: describe the delivery process accurately and encode it in the AI agent's instructions. You're not approving a mystery; you're writing down what you already know and checking the machine follows it.
4. The customer is your quality control — for free. This is the most important one. If the AI gets a service wrong, you don't wait for a dashboard. The customer tells you, immediately and loudly, because their money and outcome are on the line. They are ruthlessly invested in quality and will not let a mediocre AI through. You've recruited the most demanding QA team imaginable, and you're not paying them.
And the part that ties it back to money: services are where the 80% of value and revenue lives. A win there shows up on the P&L — the exact thing 95% of companies are failing to produce. You also get, almost for free, the one success factor the MIT data kept pointing to: a defined outcome stated before you build. The outcome is obvious — a satisfied customer who paid.
The deeper logic
Step back and the whole thing clicks. The research is unanimous that AI rarely fails on capability. It fails on three things: no defined outcome, weak feedback, and no buy-in. Starting with a service fixes all three at once. The outcome is the customer's satisfaction and the revenue attached to it. The feedback is daily, from both the expert who runs the service and the customer who receives it. The buy-in is structural, because the person who owns the service genuinely wants a tool that makes their work better. Services-first isn't a preference. It's the structural antidote to the 95%.
This is also why I keep returning, on this site, to concrete cases over abstract promise — like the hour in which one person and an AI closed a live security breach that would have cost a firm tens of thousands. The leverage is real when the work is well-defined and the result is visible. It evaporates when the target is a fuzzy internal process no one can score.
The honest caveats
I'm not telling you to point AI at your highest-stakes service tomorrow morning. Some services are regulated or carry real risk to customers — there the bar for human oversight and governance is higher, and that's a separate discipline I've written about in what AI compliance actually costs. Start with a service that's well-understood and forgiving, prove the loop, then climb.
And there's a craft to the first step that I've deliberately skipped here: how to describe a service precisely enough to hand it to an AI without losing the judgment that makes it good. That's the make-or-break, and it deserves its own piece. It's also close to the work of building an AI-enabled service business from the operator's side rather than the vendor's.
So if you take one decision from this: before you buy another back-office tool that won't show a number for three months, look at the services your customers already pay you for, and ask which one you understand well enough to describe out loud. That's where your first real AI return is hiding — not in the half of the company the customer never sees, but in the half they pay for.
I'm building a standard for exactly this — how to describe and govern an AI-run service so it stays accountable to the customer. If that's a problem you're wrestling with, I'd genuinely like to hear how you're approaching it; reach out through the contact on this site. The companies that get this order right won't just avoid the 95%. They'll quietly pull away from the competitors still piloting chatbots in the back office.
