The demo always works. The ground doesn't.

The demo always works

When my wife and I bought our place on Whidbey Island, the first thing we learned was that it used to be part of a golf course. The second thing we learned was that somebody had scraped the topsoil off about forty years ago. What's left is mostly sand.

You can still make ground like that look great for a weekend. Grab a flat of starts from the nursery, dig some holes, water them in, take the picture. Everything's green. Come back in August and it's a different story. Those plants were never really in the ground. They were living off the little plug of potting soil they came with, and when that ran out, so did they.

That's every AI prototype I've seen in the last two years. The demo shows up with its own potting soil: a clean export, a hand-matched customer list, one department's data from one quarter. It looks fantastic in the boardroom. Then somebody says "great, let's roll it out," and now you're asking it to grow in sand.

Ngentix is how you put the topsoil back.

Everybody's seeing the same thing

I'm not the only one who's noticed. MIT's NANDA group dug into hundreds of enterprise AI deployments last year and found only about 1 in 20 pilots producing real revenue gains. Their take wasn't that the models were bad. Companies just couldn't get them wired into how the business actually runs. The same pattern is what sits under why most enterprise AI initiatives never get past the pilot.

Gartner was blunter. They expect companies to walk away from 60% of AI projects that don't have AI-ready data behind them through 2026, and nearly two thirds of the organizations they surveyed weren't sure their data practices were up to the job.

The number that really stuck with me came from a Fivetran survey of about 400 data leaders. Even at companies that had already done the heavy lifting of centralizing their data, two thirds were spending more than 80% of their data engineering time just keeping the existing pipelines alive.

Sit with that for a second. They built the house, and now the entire crew is up on the roof patching leaks. Nobody's left to build the addition. Getting the data together once was never the hard part. Keeping it true is. That is the same trap as treating the integration platform as a pipe: the connection gets built, and then the whole job becomes keeping it from leaking.

Why the demo works

Here's the dirty little secret of a good AI demo: somebody did the hard part by hand, and nobody wrote it down.

Someone pulled a clean export out of the ERP. Someone burned two days matching customer IDs so they'd line up with the CRM. The team picked one workflow, one department and one month of data, and froze it in place. Then the model ran against that tidy little snow globe and knocked it out of the park.

There's nothing wrong with that. It's what a prototype is for. But what you actually built was a photograph of your business on a Tuesday afternoon, and the business didn't hold still for the photo.

I've been watching this same movie for thirty years, long before anybody said "LLM." We build software for the company as it looked on launch day, then act surprised when the company keeps moving and the software slowly stops telling the truth.

Then the rest of the business shows up

Production is where the demo meets the rest of the company. Now the AI has to read from the ERP, the CRM, the HR system, ticketing, billing, the warehouse system, and the three spreadsheets that actually run the place. Every one of them has its own opinion about what a "customer" is. They were never built to agree, and they don't.

And they won't sit still. A vendor ships a new API. Finance adds a field. Sales swaps CRMs. You buy a company and inherit a second ERP. Each change quietly pops another strip of the duct tape the prototype was held together with, and it isn't anybody's job to notice.

I learned how much this matters at TradeLens. We had every major ocean carrier on the platform, plus customs and trade organizations like the UN's ASYCUDA, and companies like Walmart using it. The technology worked. What sank it was trust between the carriers. You can build the smartest system in the world, but if the people at the table don't believe what's on the screen, you've built a very expensive screen.

AI makes this worse in a sneaky way. A dashboard with bad data usually looks off, so somebody squints at it and catches the problem. An AI with bad data sounds great. It takes the disagreement between your systems and irons it into a smooth, confident paragraph. Think of the most articulate intern you've ever hired, who has never once said "I'm not sure."

So the project doesn't blow up. It just parks itself in what analysts now call pilot purgatory: too promising to kill, not trustworthy enough to bet the business on.

Put the topsoil back

You can't fix sand with better seeds. On our land the fix is slow and pretty unglamorous: compost, cover crops, letting the ground rebuild what got stripped off. You do it before you plant, and you keep doing it every season, because the minute you stop, the sand starts winning again.

That's the layer missing between your systems and your AI, and it's what we built Ngentix to be. It sits in the middle of everything you run and gives your systems one shared definition of the business: what a customer is, what an order is, what an invoice is. We've defined more than a thousand of those business nouns and verbs, so your systems translate into one language instead of arguing in thirty.

Then it keeps that language true while everything around it changes. When a vendor changes an API or somebody adds a field, Ngentix catches the drift and repairs the mapping on its own, whether the system speaks REST, SOAP, EDI, MCP or agent-to-agent. The contracts, PDFs and email threads that hold half your real operating data get turned into structured records all the time, not once during some onboarding project. And every translation and every fix gets logged, so when the AI tells you something, you can follow the breadcrumbs back to the system it came from.

At Wayfair, my teams ran the systems behind CastleGate, which handled roughly a fifth of the company's global sales. At that scale, moving data around is the easy part. The real job is making sure everybody is looking at the same truth before they make a call.

None of this replaces your systems or your AI tools. Your models are going to keep getting better every quarter without any help from you. What they can't do on their own is figure out which of your systems is telling the truth today. Our shorthand for it: your systems change; we ensure the truth doesn't.

Before you fund the next pilot

If you've got an AI project parked somewhere between the demo and the rollout, stop shopping for a better model for a minute and ask your team three questions.

First, who did the data matching in the prototype, and is that work anywhere in the production plan? If the answer is "an analyst did it by hand over a long weekend," congratulations, you've found the leak.

Second, what happens to this thing the next time one of our systems changes? If the room goes quiet, the project has an expiration date.

Third, can we trace any answer it gives back to where the data came from? If not, nobody in finance or legal is ever going to sign off on acting on it. Honestly, they shouldn't.

I've been at this for about thirty years, from NikeID to ad tech to ocean shipping to e-commerce. Every wave, desktop, web, mobile, cloud, taught me the same lesson. The technology is rarely what holds you back. What holds you back is whether the information underneath it is still true next month.

AI is just the wave where you can't fake it anymore. Fix the soil first. Then plant.