The GenAI Divide: Why 95% of Businesses See Nothing From AI While 5% Bank Real Returns

American enterprises spent an estimated US$40 billion on AI systems in 2024. According to MIT's Project NANDA, 95% of them have nothing measurable to show for it, no profit-and-loss impact at all. The researchers gave the phenomenon a name: the GenAI Divide. A thin slice of organisations extract millions in value; the vast majority run pilots that produce demos, headlines and precisely zero dollars.

If you've read our earlier piece on pilot purgatory, this is the companion question: not why do pilots stall before production, but why do even the launched ones so often earn nothing. The research is blunt about the answer, and it isn't the models.

The two failure causes hiding under "AI didn't work"

Strip the post-mortems down and the same two causes keep appearing.

Cause one: the data was never fit to build on

Eighty-five per cent of failed projects trace back to data quality, and only about one organisation in eight has data that's actually AI-ready. For an SMB, this isn't abstract. Your "data infrastructure" is your CRM, your accounting file and the spreadsheets in between. If the CRM has duplicate contacts, dead pipelines and lifecycle stages nobody trusts, then every AI initiative you fund is a renovation built on wet sand. The model summarises garbage fluently. Fluent garbage is worse than obvious garbage, because people act on it.

Cause two: nobody defined what success was worth

The MIT Sloan finding is the one I'd frame and hang on the wall: 61% of projects were approved on projected returns that no one ever went back and measured. And the follow-on stat explains why it matters. Projects with quantified success metrics defined before launch succeed at 54%; projects without them succeed at 12%.

Writing down the number you expect, before you spend, more than quadruples your odds. It's the cheapest intervention in all of AI.

What the 5% do differently

The organisations on the right side of the divide aren't smarter about AI. The research says they're more disciplined about everything around it. Three habits recur.

They fix the substrate first. The winners built on data infrastructure designed to support production use, for an SMB, that means the unglamorous portal work: dedupe, lifecycle definitions, dead-workflow cleanup, reporting that reconciles with the bank account. Every dollar of that work also pays off with zero AI attached, which is why it's the correct first dollar.

They define the metric before the pilot. Not "improve productivity". A number. Minutes per quote. Response time to a new enquiry. Percentage of settlements with complete documentation. One workflow, one number, measured before and after.

They treat adoption as work, not gravity. The failed projects assumed end users would simply start using the thing. The successful ones assigned an owner, changed the process the tool lives inside, and checked usage weekly. Tools don't create adoption. Owners do.

A 30-day version for an SMB

1. Week 1: pick one workflow and one number. Choose a single revenue-adjacent process that annoys everyone, and write down its current cost: hours, response time, error rate. That baseline is your entire business case.
2. Week 2: audit the data that workflow runs on. Usually a handful of CRM objects and fields. Fix those, not the whole portal, just the fields this workflow reads and writes.
3. Weeks 3–4: run the smallest AI (or automation) that moves the number. Often the honest answer is deterministic automation with a small AI step inside it, not an AI platform. Ship it to the people who do the work, with an owner.
4. Day 30: measure against the baseline, then decide. Better? Expand. Flat? Kill it, keep the cleaned data, pick the next workflow. Either way you've spent weeks and hundreds of dollars learning what enterprises spend quarters and millions not learning.

The takeaway

The GenAI Divide isn't a technology gap. The 95% and the 5% are largely buying the same models. It's a foundations-and-discipline gap: clean data underneath, a number defined up front, an owner driving adoption, and the honesty to measure afterwards.

All four of those are available to a twelve-person business in Parramatta just as much as to a bank. Arguably more so. You can fix your data layer in weeks, not years. The divide is real, but which side you're on is a choice.