Construction

How to judge an AI vendor when you cannot judge the software

A $20 million business run on sticky notes cannot tell a real AI product from a wrapper. Annija Eizenarma on what to evaluate instead of the technology.

You probably cannot tell a substantial AI product from a thin layer over a language model, and you are not expected to. Annija Eizenarma, a founder, investor and adviser with seven years in legacy industries, puts the gap plainly: "legacy industry operators work with sticky notes and they're doing like twenty million a year." So evaluate what you can actually judge — one delivered result, and the people selling it.

Why can't you tell a real product from a wrapper?

Because the distinction lives in a technical vocabulary that has no bearing on how you run your business. Asked how a legacy operator should tell a substantive product from a thin layer over a model, Eizenarma did not offer a test. She said there is not one available to this buyer:

"At the first glance the legacy industry buyer really can't tell if it is a [wrapper] or not, right? Because it's still very far from their understanding of how the things work."

The twenty-million-on-sticky-notes contrast is the whole problem in one sentence. A sophisticated business operator is not a software evaluator, and no amount of technical documentation closes that distance.

The consequence is that evaluation gets replaced by trust, which means the market rewards whoever is best at earning trust rather than whoever built the better product. Real products and thin wrappers compete on equal footing at the first meeting.

What can you evaluate instead?

One bounded result, delivered with minimal onboarding, and what actually happened rather than what was claimed. Eizenarma's advice on this point is aimed at builders — prove one thing well before proposing more. Artificial Reality's inference is that it inverts cleanly into a buyer-side test.

The mechanics matter. Ask for one specific job, name the output you expect, cap the data you hand over, and set a date. A vendor whose product is real will take a narrow scope, because a narrow scope is how they intended to sell to you anyway. A vendor who needs six months of your records before anything appears has told you where their pilot will end.

Why is an AI-forward pitch a warning sign?

Because the vendors who understand this market lead with the outcome, not the technology. Eizenarma is direct about how the framing lands:

"Legacy industry buyers, they really don't want AI for AI. They're truly, in a way, skeptical for what AI stands for, and in many ways they really don't want anything to do with it."

Her prescription to builders is to invert the vocabulary entirely: "as little tech words as possible, more focusing on the solution itself and the end outcome." A vendor still selling you the model rather than the result has not spent time with businesses like yours — which is a fact about the vendor you can assess without understanding a line of their code.

What should you ask about why they chose your industry?

Ask where the founder's knowledge of your work comes from, because the answer predicts whether they will still exist in two years. Eizenarma describes the common origin story she considers a warning:

"Often I see startups coming up with business idea based on the market size, where they're sitting with AI and asking ChatGPT: 'Okay, ChatGPT, tell me how I can become a millionaire, which is the biggest market, where is biggest market share where it can build, where are many problems in which industry?' And I think it's a one-way, but not necessarily the smartest way."

Her objection is about endurance rather than taste: "It is not an easy journey to sell into these industries. And I think that's where often companies die off." The qualifying conditions she names are "really good insights from your own work experience, from your family's work experience, or somebody really close that you can relate with the problem" — and being passionate about it "beyond the market size, beyond the money that you see potentially there for the market."

That is stated as guidance to founders. Artificial Reality's inference is that it is equally a buyer's question, because a vendor who selected your industry from a list is a vendor who may abandon it, taking your implementation with them.

Five questions to ask before you sign

What is the one job this does? If the answer takes ten features, you are being asked to evaluate ten claims at once.
Can you describe my week before I explain it? The vendors who succeed here learned the physical workflow first.
What can you deliver in three weeks with the data I already have? Then judge the result, not the demo.
Does this plug into what I already run? Building on top of existing systems is the pattern Eizenarma describes as working.
Why this industry, and how do you know it? Market size is an answer. It is not a good one.

And if you do not want to change anything?

That is a legitimate position, and Eizenarma treats it as one. She describes calls where an operator tells her plainly: "I'm exactly where I want to be, never call me again, I don't want to grow. And that's fine. That's their decision." Her conclusion, aimed at vendors, is that "doesn't matter how good product is, you will not change somebody's perspective on their business."

So decide first whether you are trying to grow, then buy accordingly. If you are, pick the single task costing you the most hours this month and run one three-week test against it. If you are not, you are allowed to say so, and the pressure in your inbox is a sales problem, not yours.