An AI demo is designed to impress. It uses clean examples, it runs on the supplier’s best day, and nobody asks what happens in month seven when the bill doubles or the supplier is bought by someone else. The time to ask those questions is before you sign, when you still have bargaining power and the supplier wants your business.
Below are the questions worth asking, grouped by topic, with notes on what a good answer sounds like.
Questions about your data
- Where will our data be stored, and in which country? A good supplier answers plainly. “In the cloud” is not an answer.
- Will our data be used to train your models or anyone else’s? You want a clear no, written into the contract, not just said in a meeting.
- Which outside AI providers will see our data? Many AI products send your text to a large model provider behind the scenes. You should know which ones.
- Who on your side can see our data, and is that access logged?
- If we leave, how do we get all our data back, in what format, and how quickly is it deleted from your systems?
If you hold customer details such as national identity numbers, phone numbers or loan histories, these answers matter a great deal. Sri Lanka has a Personal Data Protection Act, and your lawyer can tell you what your obligations are when a supplier handles that information for you.
Questions about cost
AI pricing is often harder to predict than ordinary software. Many products charge partly by use, so a busy month costs more.
- What is the full monthly cost at our expected volume, and at double that volume? Ask for both numbers in writing.
- Is there a spending cap, and what happens when we reach it? A good system stops or warns. A poor one keeps running and bills you.
- What costs are not in this quote? Setup, training, extra users, integrations, support hours, model usage.
- How much notice do you give before raising prices?
- Is the price in rupees or dollars? A dollar price means your cost moves with the exchange rate.
A worked example: a quote of USD 200 a month sounds fixed. If it includes 10,000 conversations and your WhatsApp volume doubles in the festive season, you may pay overage charges for those weeks. Ask how overage is calculated before it happens.
Questions about control and safety
- What can the system do without a person approving it? Anything that sends money, messages customers or changes records should be approvable by a person, at least at first.
- Can we see a log of every action the system took, and why? If the answer is vague, walk away. Without logs, you cannot investigate mistakes or answer customer complaints.
- How does it behave when it does not know the answer? Ask for a demonstration of a question it cannot answer. A good system says so and hands over to a person. A poor one makes something up.
- Can we switch it off instantly, and who can do that?
- How do you test changes before they reach our system?
Test it on your own material
Ask to run the tool on twenty or thirty real examples from your business: actual customer messages, actual invoices, actual forms. Include the awkward ones, such as messages mixing Sinhala and English in English letters, or blurry photographs of documents. Count how many it handles correctly. This tells you far more than any demo.
Questions about ownership and exit
- Who owns anything built for us: the configuration, the instructions, any code?
- Which accounts are in our name? Ideally, cloud and AI provider accounts belong to you, with the supplier given access. If they belong to the supplier, you are locked in.
- Can we change the underlying AI model later? Models improve and prices change. A system tied to one model provider has less room to adapt.
- If your company closes or is sold, what happens to our system?
- What is the minimum contract length and the notice period?
Answers that should worry you
Some responses are warning signs, however good the demo was:
- “It is nearly always accurate.” Ask: on what data, measured how? Accuracy on someone else’s examples tells you little about yours.
- “You will not need anyone to check it.” Every AI system makes mistakes. The question is whether they get caught.
- “We cannot share that, it is proprietary.” Reasonable for their internal code. Unreasonable for where your data goes or what the system did.
- “We can start straight away, no need to look at your process.” Good suppliers want to understand the work before promising anything.
Be fair in return. A small supplier may not have polished answers to every question, and that is not always a bad sign. What matters is whether they answer honestly and put the important points in writing.
This list will not protect you from every problem, and it is no substitute for legal advice on the contract itself. But a supplier who is comfortable with all of these questions is usually one you can work with. Print them, take them to the next meeting, and note which ones get a clear answer.