2024 is a gold-rush year for AI, and gold rushes are wonderful for the people selling shovels. Right now a large share of the "AI companies" pitching Indonesian businesses are thin wrappers around someone else's model, dressed up with a logo and a slick demo. Some are excellent. Many are one clever prompt away from having nothing to sell.
Choosing an ai vendor in this environment is a real risk, not because AI does not work, but because it is unusually hard to tell what you are actually buying. The demo always looks magical. Your job as a buyer is to look past the demo and ask what the vendor truly owns, what happens when the ground shifts under them, and how you get out if it goes wrong.
This is a checklist I use when I help clients evaluate these pitches. None of it requires you to understand machine learning. It requires you to think like a buyer of any critical supplier, and to be a little skeptical when something feels too easy.
Ask what they own versus what they resell
The first and most clarifying question when choosing an ai vendor: which part of this is yours, and which part are you renting from a company I have never heard of?
Most AI products today sit on top of a foundation model from a big provider. That is fine. What matters is how much value the vendor adds above that, and whether they are honest about the arrangement.
- A thin wrapper takes your question, forwards it to someone else's model, and returns the answer. If that is all it does, you are paying a markup for something you could nearly buy direct.
- A real product adds meaningful work on top: your data integrated properly, workflows built around your process, guardrails, evaluation, ongoing tuning. That is worth paying for.
Ask directly: "If the underlying model you use disappeared tomorrow, what would you still have?" A good vendor has a real answer about their own data, integrations, and know-how. A weak one gets uncomfortable, and that discomfort is your information.
Pin down data handling before anything else
This is the one I refuse to skip. You are about to feed a supplier your customer records, your contracts, maybe your financials. Where does that go?
Get clear, written answers to:
- Where is our data stored, and in which country?
- Is our data used to train their or anyone else's models? If yes, walk away or get it excluded in writing.
- Who on their side can see it, and how is access controlled?
- When we leave, is our data deleted, and can we get proof?
A serious vendor answers these crisply because they have thought about them. Vague reassurance like "don't worry, it's secure" is not an answer, it is a red flag. This is the same discipline behind ordinary security basics small businesses keep ignoring: the boring questions are the ones that save you.
Understand your model dependency and its pricing
Here is a risk unique to this moment. Many AI vendors have costs tied to an underlying model whose price they do not control. In 2024, model pricing has been volatile, sometimes dropping, sometimes changing terms overnight.
So ask what happens to your bill if the model they depend on doubles in price, or changes its terms. Do you absorb it? Are you protected? Does their whole product break if the provider changes a policy?
You are not trying to trap them. You are checking whether they have thought past the demo to the boring reality of running a service on rented foundations. A vendor who has a clear answer here is one who plans to be around next year.
Demand proof beyond the polished demo
Every AI demo works. It was built to. The demo is a sales asset, not evidence that the thing will work on your messy real data.
So do not buy on a demo. Insist on one of these instead:
- A paid pilot on your own data. Small, time-boxed, with a clear success measure agreed up front. Pay for it. A paid pilot filters out vendors who cannot actually deliver on real inputs.
- A reference you can call. An existing client, in a similar situation, who will talk candidly. If they have no reference willing to speak, ask why.
- A test on data they have not seen. Hand them a realistic sample and watch how it performs, and how they respond when it stumbles.
The paid pilot is the strongest filter I know. It costs a little, it saves a lot, and the vendors who resist it are usually telling you something important about themselves.
Write the exit into the contract
Optimists sign contracts imagining success. Professionals sign them imagining the divorce. Before you commit, know exactly how you leave.
| Exit question | What to require |
|---|---|
| Our data | Full export in a usable format, on demand |
| Notice period | Reasonable, not a trap that locks you for a year |
| Handover | Documentation and cooperation if we move providers |
| Lock-in | Understand how hard it is to leave before you enter |
If leaving is designed to be painful, that pain was designed on purpose. Know it going in.
Practical takeaway
Choosing an ai vendor in a hype year comes down to seeing past the demo to the substance underneath.
- Ownership: what do they have if the model they rent vanishes?
- Data handling: written, specific answers, never vague reassurance.
- Model dependency: who eats the cost if the foundation shifts?
- Proof: a paid pilot on your data beats any demo.
- Exit: know how you leave before you sign.
Run this list and most thin wrappers eliminate themselves, because they cannot answer it. The strong vendors welcome the questions, which is exactly why they are strong. If you want someone in your corner who reads these pitches for a living, that is the kind of engagement I take on as a technology partner.