95% of Generative AI Pilots Deliver Zero Return. Here's the Question Businesses Should Ask Instead.

A widely discussed 2025 research report from MIT's NANDA initiative, "The GenAI Divide: State of AI in Business 2025", reported that roughly 95% of the enterprise generative AI pilots the researchers studied produced no measurable profit and loss impact. The finding circulated quickly because it matched what many executives were already experiencing: real activity, real spending, and no number they could point to at the end of the quarter.

It is worth reading the finding carefully rather than repeating the headline. The researchers described a divide, not a verdict on the technology. A small share of initiatives created measurable value. The majority did not — and the difference was rarely the model. It was how the work was chosen, scoped, and integrated.

Source: MIT NANDA, "The GenAI Divide: State of AI in Business 2025" (2025). The figure is widely repeated secondhand; read the original research before citing it.

Most Pilots Start From the Wrong End

The typical pilot begins with a capability. Someone sees a demonstration, an internal group is asked to explore it, and a use case is selected because it is convenient to build rather than because it matters financially.

That sequence produces a predictable outcome. The pilot works technically. People find it interesting. It never gets connected to a decision, a workflow, or a budget line, so no one can say what changed. When the enthusiasm fades, there is nothing to defend.

A Better Starting Question

Do not start by asking, "How can we use AI?"

Start by asking:

"What business number are we trying to move, and can AI move it?"

This single reframing filters out most of the projects that would have failed anyway. It forces three commitments up front: a specific metric, a baseline, and a mechanism by which the metric could plausibly change.

If a proposed initiative cannot name the number, it is an experiment. Experiments are legitimate — research and learning matter — but they should be funded and evaluated as experiments, not presented as business cases.

What Qualifies as a Good First Number

Strong candidates tend to share a few characteristics:

• The metric is already tracked, so a baseline exists.
• The metric has a known financial value per unit.
• The workflow is well understood and reasonably contained.
• A human remains accountable for the outcome.
• The result can be observed within weeks, not years.

Weak candidates are usually broad and abstract: "improve productivity," "become more innovative," "modernize operations." These may be real ambitions, but they cannot be measured at the level where a project is judged.

Why Customer Acquisition Is a Clear Case

Customer acquisition tends to satisfy every one of those criteria, which is why it is often a sensible place for an organization to start.

Most businesses already spend money bringing people to their websites through search, advertising, referrals, content, events, and reputation. That traffic is measured. The value of a new customer is usually known or estimable. And the failure point is easy to describe: a large share of visitors read a page and leave without calling, booking, requesting information, or identifying themselves in any way.

The funnel is explicit:

Website Visitors → AI Conversations → Qualified Leads → Appointments / Consultations → Customers → Revenue

Each arrow can be counted. That makes the value of any improvement calculable rather than rhetorical, and it makes an unsuccessful attempt visible early instead of at the end of a long program.

An AI Customer Acquisition System sits directly on that funnel. An AI Concierge engages visitors who would otherwise leave, understands what they are trying to accomplish, answers questions from approved business knowledge, identifies intent, qualifies prospects, captures leads, and moves qualified prospects toward the appropriate next action.

None of that requires guaranteeing a result. It requires instrumenting a funnel that already exists and then observing whether the numbers move.

Integration Is Where Value Is Won or Lost

The NANDA research pointed to integration and workflow fit as recurring differentiators, and that matches practical experience. A capable model that sits beside the business changes nothing. The same capability connected to scheduling, follow-up, CRM records, and the people who handle inquiries changes how work happens.

Before building, describe the handoff in one sentence: when the system produces an output, who receives it, in which tool, and what do they do next? If that sentence is hard to write, the pilot is not ready.

Practical Questions to Ask Before Funding an AI Project

1. Which number are we trying to move, and what is it today?
2. What is one unit of that number worth?
3. Where exactly does AI enter the existing workflow?
4. Who is accountable for the output, and what can they see and correct?
5. How will we know within 60 to 90 days whether it worked?

If those five questions have clear answers, the project is likely to produce evidence either way — which is the point.

The Real Lesson

The lesson of the research is not that generative AI does not work. It is that technology-led projects rarely produce financial results, while outcome-led projects sometimes do.

Choose a number that matters. Find the point in the workflow where an AI capability could plausibly change it. Instrument the path. Then judge the work by whether the number moved.