Construction

The AI pilot that dies in month three of data collection

Integration costs of $500,000, contracts running five to ten years, and onboarding that takes months: why AI deals in legacy industries die before output.

AI deals in legacy industries usually fail before anyone sees output. They fail on data that does not exist in usable form, on integration money already spent, and on a sales cycle measured in years against funding measured in months. Annija Eizenarma, a founder, investor and adviser with seven years in legacy industries, has heard integration for a single system quoted "up to 500,000 or a million."

None of these are technical failures. That is what makes them expensive.

Where does the pilot actually die?

In onboarding, before any result is produced. Eizenarma describes the starting condition:

"Legacy industries are very known for having very fragmented different vendors, different solutions, softwares, Excel and everything."

The operational consequence: "They can't make data driven decisions because there's no united platform where all of the data speaks together."

And the consequence for anyone deploying into that environment:

"If you wanna even integrate an AI solution to run the business more efficiently, they will need to onboard you. And in order for them to onboard you, they will need a lot of data. And if you don't have that data, it will be really inconvenient for you and for them to work together. It will take multiple months to get all of the information about your clients, about their preferences, whatever, their projects."

This is the quietest way a pilot ends. Not on output quality — in month three of data collection, inside the exact window in which the vendor needed to prove one thing well and earn the next conversation.

Why won't they switch even when they dislike what they have?

Because they have already paid for it, and someone internally would have to defend writing that off. Eizenarma on the installed base:

"Often it's the case when there's a big fragmentation and they're almost always disappointed. They never like it. They use all the systems because they have it sold on them. And usually they have these licenses and their managers and they have contracts for five, ten years."

And the number that reframes the sales conversation:

"You have to fight against big onboarding and integration costs that legacy systems have. Like when they have all these Salesforces or whatever, they have paid a lot of money to install it in the first place. Often it's from 50 to 150, sometimes even up to 500,000 or a million, I have heard, for integrating a single system. And it's really hard for them to switch over because they have already paid, they have invested."

(Eizenarma does not state a unit for the first two figures; Artificial Reality reads them as thousands of dollars from context. These are her practitioner observations, not published benchmarks.)

Note what follows. Dissatisfaction is not the constraint — the operator is already unhappy. The constraint is that switching is a career risk for the person who signed the last contract, and no feature comparison addresses a career risk.

How long does this market really take?

Years, and Eizenarma's own company is the illustration:

"When I started the first company with my co-founder, at that time the thing we were doing, we were building smart home solutions and building automation solutions. In 2017, it was not a thing. People were talking about it, but it was not a thing. So in order for us to get to a point where we can even pay ourselves even a little salary, it took two years — because you have to educate the market first."

What those two years consisted of: "It took a lot of time on knocking on doors, meeting the people, sitting at the same table, listening to their stories, looking at their projects, to eventually say: okay, I can help you."

And a further constraint: "Often you do have to create a product to show them how it works before they can even believe it."

So the sequence is build first, educate by knocking on doors, reach a small founder salary two years in. Most AI companies are not capitalised or governed for that shape. In Artificial Reality's assessment, that mismatch is the largest single cause of failure in the category.

What does this change about how you build and raise?

Treat readiness as your product surface. If your median customer needs months to assemble data before you can start, that gap is the thing to build, not a prerequisite you wait out.
Price and scope against a five-to-ten-year contract you are not replacing. Sit beside the incumbent rather than asking anyone to write off a six-figure install.
Design the first deliverable to need almost no data. One bounded result inside weeks is what buys the second conversation.
Raise for two years of market education, or pick a market that is already educated. Those are the two honest options.
Qualify on growth intent. A pipeline of profitable, satisfied operators converts at close to zero.

Run the diagnostic on yourself before your next board meeting: how long does your median customer take from signing to running on their own data? If the honest answer is months, stop calling it onboarding friction. It is your roadmap.