Why this guide exists
India has thousands of software agencies and, since 2024, almost all of them sell AI development. Some have shipped real AI systems to real users. Many have shipped a demo. The price gap between the two is small; the outcome gap is enormous. The twelve questions below separate them in one call, and yes, we answer all twelve ourselves on ours.
The twelve questions to ask
1. What have you shipped that I can use today?
The strongest evidence is a live product or feature you can try, not a slide. Ask for links. An agency that runs its own AI products has paid its own bills for models, queues and mistakes, which changes how it estimates yours.
2. How will we know if the AI is good enough?
The right answer mentions a test set: real examples with known correct answers, measured before and after every change. If the answer is "you will see it in the demo", keep looking. A demo can be steered; a test set cannot.
3. What will it cost to run each month, not just to build?
AI features cost money per use. A good agency estimates the monthly bill from a prototype's real usage and designs to keep it down. Our article on what an AI agent costs to run shows what that estimate should look like.
4. What happens when the AI is wrong?
Every AI system is sometimes wrong. The professional answer describes fallbacks: showing sources, confidence limits, and a hand-off to a person. The amateur answer is that the model is very accurate.
5. Which model will you use, and why?
The only good answer is some version of "we will test candidates on your data and show you the comparison". A company married to one model is selling what it knows, not what you need.
6. Who owns the code, the prompts and the trained models?
You should, in writing, on final payment. Prompts and test sets are part of the deliverable; ask for them explicitly.
7. How do you protect our data?
Listen for specifics: API terms that exclude training, removing personal information before it reaches third-party models, data kept in an agreed region, NDAs as standard. Vague reassurance is a warning sign.
8. Can we see the plan for the first two weeks?
A disciplined AI project starts with discovery and a measured baseline, then a prototype on your real data. If week one is "we start building", the estimate you were given is a guess.
9. Is the price fixed, and what does it include?
Hourly billing on an uncertain AI project transfers all the risk to you. Prefer a fixed price after a paid or free discovery, with the test set and documentation included.
10. Who exactly will work on our project?
Ask to meet the engineer, not only the salesperson. In good agencies the person who scopes the project builds it.
11. What would make you tell us not to build this?
An honest company has refused projects and can tell you why: no clear right answer, too little volume, a ready-made tool that already does it. A company that has never said no will not start with you.
12. What support comes after launch?
AI systems drift as your business changes. Look for monitoring, a monthly review of failures, and a clean handover option, not a silent handoff.
Warning signs, in one list
- Accuracy promises with a number and no test set behind it
- A portfolio of demos but nothing live
- The same pitch for every problem, always ending in a chatbot
- No mention of running costs until after you sign
- Prompts and models excluded from what you own
What it should cost
Rates for experienced Indian teams generally run $15 to $40 per hour against $80 to $150 in the US, and fixed-price AI projects commonly land between the cost of a mid-range and a complex app build. The ranges in our app development cost guide apply, with the running-cost question added on top.