Two of the most credible research shops in the world published the same finding this year, from two completely different angles: almost every company running AI pilots is getting nothing back for it. We read both reports closely because it changes how we scope every engagement we run for operators.
MIT's Project NANDA reviewed more than 300 publicly disclosed AI initiatives, ran 52 structured interviews, and surveyed senior leaders at four conferences. Their July 2025 report, The GenAI Divide: State of AI in Business 2025, found that despite $30 to $40 billion in enterprise AI spend, 95% of generative AI pilots showed no measurable effect on profit and loss. Not underwhelming. No measurable effect at all.
Deloitte's own 2025 State of AI in the Enterprise survey, polling more than 3,200 business and IT leaders across 24 countries, landed close to the same place from a different angle. Workforce access to sanctioned AI tools jumped from under 40% to roughly 60% of employees in a single year. That number looks great in a board deck. But only 25% of companies report AI having a transformative effect on the business, and just 25% have moved 40% or more of their AI pilots into production. Adoption went up. Impact barely moved.
We run luup on Claude, we build these systems for clients, and this is the gap we see on almost every discovery call. A company has ChatGPT seats for the whole team, maybe a chatbot on the website, maybe someone in ops experimenting with agents on the side. None of it is touching a number anyone tracks. That is not a technology problem. It is a scoping problem.
Why the 95% fail: no number, no fix
The pattern in MIT's interviews was consistent: teams built something that looked impressive in a demo, then had no way to prove it saved anyone time or money because nobody named a metric before they started building. A polished internal tool with no baseline to compare against is a science project, not a system.
Contrast that with the 25% Deloitte found reporting a transformative effect from AI. They are not running more pilots than everyone else. They are running fewer, better-scoped ones, each tied to a specific line item. That is the entire difference between a pilot and a system: a pilot proves the tech works, a system proves the number moved.
The three numbers worth chasing
Every AI project we scope for a client comes down to one of three things: time saved, mistakes cut, or money made. If a proposed project cannot be tied to one of those three before a single line of automation gets built, we do not build it. That discipline is what separates the minority seeing transformative results from the majority still running shallow pilots.
- Time saved: the report someone rebuilds from scratch every week, the inbox someone triages by hand every morning, the intake form someone re-types into three systems.
- Mistakes cut: the manual data entry that occasionally sends the wrong invoice, the missed follow-up that costs a deal, the compliance step someone forgets under deadline pressure.
- Money made: the lead that goes cold because nobody replied fast enough, the upsell nobody remembered to offer, the quote that took three days instead of three minutes.
Speed matters more than most operators assume here. Our own build data on lead response systems keeps landing on the same threshold: leads contacted within 5 minutes convert dramatically better than leads contacted an hour later. That is a number. A chatbot with no response-time target attached to it is not.
Why in-house skills are getting harder to hire, and more expensive
The labor market data backs up why this gap exists. Accenture's own research shows most companies' AI investment plans are running well ahead of their organizational readiness to actually use the tools they bought. And PwC's 2025 Global AI Jobs Barometer, built from close to a billion job postings across six continents, found the wage premium for AI skills hit 56%, more than double the prior year's premium. That premium is not paid for people who can open ChatGPT. It is paid for people who can walk into a business, find the one process actually bleeding time or money, and build something reliable against it.
For a 10 to 50 person company, that skill almost never exists in-house yet, and hiring for it against a 56% wage premium in a market this tight is a slow, expensive bet. That is the exact gap our AI Concierge assessment exists to close: a fixed-scope engagement that finds the biggest cost drain in your business, prices out the fix, and gets you a working system without a six-month hiring search. It is priced at 999 euros and the fee is credited to the build if you move forward, so the diagnostic itself costs you nothing extra once you commit.
What we build against instead of pilots
Every engagement follows the same four steps, whether the client is us running it or an operator running it internally.
1. Name the constraint before touching a tool
List every manual, repeatable task your team did last week. Circle the one eating the most hours or causing the most cost, then write down the actual number: 10 hours a week, 3 missed follow-ups a month, whatever it is. Our own client conversations consistently surface at least 10 hours a week of admin leak in businesses this size once someone actually maps it out.
2. Build the fix against real documents, not a demo
The fix is not a generic chatbot. It is a system fed your actual templates, your actual checklists, your actual historical examples, tuned until it gets the specific task right every time, not most of the time. This is the difference between a Second Brain that actually knows your business and a generic assistant that has to be re-explained every session.
3. Prove the number moved
Before and after, side by side. If the Friday report used to take 4 hours and now takes 20 minutes, that is 3.5 hours back every single week, and you can say that in one sentence to whoever owns the budget. This is also exactly what we surface for clients with our revenue leak heatmap: a concrete picture of where the hours and euros are actually going before any build starts.
4. Only then, expand
Once one system is live and proven, the next biggest-cost task gets the same treatment. This is how a business builds toward the small group Deloitte found actually seeing transformative results, without pretending it is already there. Our case studies follow this exact sequence: one proven system first, expansion second.
| Approach | What gets built | Outcome |
|---|---|---|
| Typical pilot | General-purpose chatbot or automation demo | No baseline, no proof, usually abandoned within months (MIT: 95% no measurable P&L impact) |
| luup system build | One named process, one named number, fed real documents and checklists | Before/after comparison, provable ROI, becomes the template for the next system |
What to do this quarter
Do not start with a tool. Start with a list. Get your team to write down every manual task that ate their time last week, circle the worst one, and put a number on it before anyone opens an AI product. If you already have three AI subscriptions running with nobody able to tell you what they saved, that is the tell you are inside the 95%, not the 25% seeing real impact.
If you want a second set of eyes on which process is actually worth fixing first, that is what the assessment is for. If you would rather see how this plays out end to end, our automation and voice agent builds both follow this exact named-number discipline, and the code and files are yours, 100%, from day one. Nobody at luup is holding your system hostage behind a subscription.
The market data is not subtle here. Companies that treat AI as a demo stay stuck in the 95%. Companies that treat it as a system, built against one number at a time, are the ones showing up in Deloitte's 25% actually seeing transformative results. That gap is not going to close on its own, and it is widening every quarter someone keeps running pilots instead of naming the constraint. More of our thinking on this is on the blog.