The 90-minute test that separates real AI products from expensive chatbots

I evaluate technology claims for a living: first as a physicist, where the instrument either measures the thing or it doesn't, and now on the venture capital team at SCB 10X, the venture arm of one of Thailand's largest banking groups, where I sit through AI product pitches every month. Most of them are the same product wearing different logos.

That experience is about to become relevant to a lot of Thai businesses at once. This month, AIS and Microsoft launched "AI Ready for SMEs": packages, training roadshows across seven regions, and free workshops for the first thousand subscribers. The Thai SME Council, present at the launch, counts 3.13 million small and medium-sized businesses in this country: 99.6 percent of all Thai firms. A wave of AI buying is starting, and the sellers are better prepared than the buyers.

The short version: AI tools that look identical in a sales demo can have opposite fates. Some compound in value for years; others are interchangeable and quietly die. There is a test that tells them apart, it takes about ninety minutes, and it requires no technical background at all.

A race has the wrong finish line

You have probably read that Thailand is behind. A global adoption survey published in late May ranked us 89th, with 10.7 percent of the population using AI against a world average of 16.3 percent. The numbers are real. The usual conclusion, that the answer is to adopt faster, is the part I'd push back on.

For a business, the risk isn't adopting AI too late. It's buying the wrong thing, watching it produce nothing, and concluding that AI was hype all along. A wasted pilot doesn't just burn budget; it burns the organization's willingness to try again. Selection, not speed, decides whether your AI spending compounds or evaporates.

The shop and the building

Here is the one piece of theory you need, and it fits in a paragraph.

Think of any AI product as a shop in a rented building. The building is the underlying AI engine, built by a handful of global companies and rented out to everyone. The rent is collapsing: the computing power behind AI answers costs a small fraction of what it did two years ago, and the price keeps falling. That sounds like good news for the shop, and it is. But it's equally good news for every competing shop, because they all rent the same building.

So the only question that matters when you evaluate an AI product is: what does this shop own that isn't the building? Your files and history inside the product. Connections to the systems where your work lives: your accounts, your inventory, your customer records. The daily routines that would hurt to rebuild somewhere else. Call it the workflow layer. It is the only layer worth paying for.

The companies that own workflow are doing extraordinarily well. Cursor, a programming tool that became the place its users do their daily work, went from its first revenue to two billion US dollars a year in roughly three years. That is the fastest climb in business-software history, according to TechCrunch and Bloomberg reporting in March 2026. The best legal AI keeps 98 percent of its customers year over year, by the company's own published figures.

The companies that own nothing but a logo do not. Last year an American sales-automation startup that had raised 74 million dollars from top-tier investors was found, in a TechCrunch investigation published in March 2025, to have displayed customer logos without permission (two well-known software companies stated flatly they were never customers) and, according to former employees in the same investigation, to have claimed 14 million dollars in annual revenue of which roughly 3 million came from contracts that survived the trial period. In its product category, vendor surveys estimate that half to two-thirds of customers walk away within a year. Nothing held them, because the product owned nothing.

The ninety-minute test

Before you sign anything, book one demo session and run five questions. Stop at the first clear failure.

1. Go off script. Let the vendor show their best demonstration. Then ask the next question a real employee would ask: a follow-up, a comparison, a "what about last month's numbers?" Polished demos are rehearsed paths; real work leaves the path in the first five minutes. A good product survives the detour.

2. Ask: "Where did that answer come from?" Any serious AI product can show its sources: the document, the record, the database it drew from. If the answer to this question is a shrug, or a vague gesture at "our AI," walk. You cannot run a business on answers nobody can trace.

3. Open ChatGPT next to it. ChatGPT costs 20 dollars a month. Give it the same task you just watched the vendor demonstrate. If you can't tell the outputs apart, you're being asked to pay a large markup for someone else's AI with a logo on top. The vendor must beat the building, not just decorate it.

4. Test your five can't-live-without features. Before the meeting, ask the staff who'd use the tool daily: what five things can we not work without? Invoicing in Thai. Stock counts that match reality. A report your accountant accepts. Then test each one, live. A tool that's brilliant at nine things you don't need and missing one you do is not 90 percent of a solution. It's zero percent.

5. Ask what it remembers after a month. Where do your files, your history, your customizations live? Do they connect to the systems you already run? This is the leaving-cost question, and here, unusually, you want the cost to be high. A tool that accumulates your working state is a tool with a reason to exist next year.

Ninety minutes. No engineering degree required, only the willingness to be unimpressed.

The good news in the collapsing price

Notice what this test measures. It never asks how clever the AI is: the intelligence is rented, identical, and getting cheaper for everyone. It asks what the product builds around the intelligence, and that is exactly the part a businessperson can judge better than an engineer. It is also, for what it's worth, the part we weigh most heavily when we decide where to invest.

Here is a prediction you can hold me to. Of the AI tools being sold to Thai businesses this month that would fail this test, I expect more than half to be discontinued, acquired, or repriced out of the market within eighteen months, by December 2027. If that turns out wrong, I'll write the follow-up explaining why.

The collapsing price of AI was never the threat. It's the opportunity: it means the expensive part is now free, for you and for every tool you might buy. What was never free is the part that holds your business's daily work. Pay for that, and only that.

Asst. Dr. Tanwa Arpornthip is Senior Advisor to the Venture Capital Team at SCB 10X, where he provides insights on emerging technologies, innovation ecosystems, and startup development. He is also a lecturer at the Faculty of Technology and Environment, Prince of Songkla University, Phuket Campus.