AIBuying guideVendors
How to choose an AI development company in India
Every software firm added AI to its homepage in the last two years. Eleven questions that separate the ones who have shipped it from the ones who have read about it.
- Author
- Kentron Technologies
- Published
- Reading time
- 6 min read
Between 2024 and 2026 practically every software company in India added AI to its capabilities page. Some of them had shipped a language model into production and supported it afterwards. Others had built a demo. The gap between those two groups is enormous and almost invisible from a proposal, because both can produce an impressive fifteen-minute presentation. These are the questions that separate them, and we have written them so they work regardless of whether you end up hiring us.
Ask to see something running in production
Not a demo, not a video, not a prototype on a laptop. Software with real users, which has been running long enough to have had problems. The follow-up question is the one that matters: what broke, and what did you change?
A team that has run an AI feature in production will answer immediately and specifically, because the experience is memorable. They will tell you about the provider outage, the prompt that worked until the model was updated, the cost that was triple the estimate in the first month. A team that has only built demos will answer in generalities.
The eleven questions
- Show me an AI feature you have in production, with real users. How long has it been live?
- What happens when the model provider has an outage? What does the user see?
- What is my running cost per conversation, per document or per month, separately from the build cost?
- Whose API keys and whose accounts are these? Who receives the provider invoice?
- Where is my data processed, and is any of it used to train a model?
- How do you know the feature is working? Show me the evaluation, or tell me honestly that it is manual.
- Show me a transcript or log of a case where the AI got it wrong, and tell me how you found out.
- What is the plain fallback if the AI path is unavailable?
- Who owns the code and the prompts at the end, and will you hand over a deployable repository?
- What does support look like in month thirteen, and what does it cost?
- What would you refuse to build with AI, and why?
The last question is the most informative one on the list. Anyone who cannot name something they would refuse is selling rather than engineering. Our own answer is that we will not put a model in a position to make a clinical or financial decision unreviewed, and we will not build handwritten prescription extraction that posts without a clinician's approval.
Signals that someone has actually shipped
| They say | Which suggests |
|---|---|
| 'Here is the cost per conversation at your volume' | They have paid a provider bill |
| 'This is the fallback when the API is down' | They have been through an outage |
| 'We log every call and here is a wrong one' | They have debugged a real failure |
| 'That should be automation, not AI' | They are solving your problem, not selling theirs |
| 'We would not automate that part' | They have judgement about risk |
Warning signs
- An accuracy percentage with no mention of which documents, which fields or whose data it was measured on.
- A flat monthly fee covering unlimited model usage, with no cap disclosed. Someone is carrying a risk, and it will be priced in eventually.
- No answer on data processing location or training usage. For patient or financial data in India this should be immediate and specific.
- A proposal where every problem is solved by a chat interface.
- Reluctance to show a failure. Every production AI feature has failed; a vendor who implies otherwise is managing you.
On data, specifically
Ask where data is processed and whether it is used to train a model. Major providers offer terms where business API traffic is not used for training, but this varies by provider, by plan and over time, so the answer should be a specific statement about your contract rather than a reassurance. If your data is patient records or financial records, get it in writing.
The DPDP Act 2023 makes you the data fiduciary. A vendor's casual approach to where processing happens becomes your obligation, not theirs, which is a good reason to be the difficult customer on this point.
Judge the diagnosis, not the proposal
The strongest signal in a first conversation is whether the vendor tries to make your problem smaller. A team that tells you a chunk of your requirement should be plain automation rather than AI, or that one part is not worth automating at all, is a team optimising for the project working rather than for its size.
A vendor who shrinks your scope in the first meeting is worth more than one who expands it.
We publish our prices, our process and our case studies with no invented percentages for the same reason: it lets a buyer decide in minutes rather than afternoons. If the eleven questions above are useful to you and you ask them of someone else, the article has done its job.
Frequently asked questions
What should I ask an AI development company before hiring them?
Ask to see an AI feature running in production with real users, then ask what broke and what they changed. Ask for the running cost per conversation or document separately from the build cost, what the fallback is during a provider outage, where your data is processed and whether it trains a model, and what they would refuse to build with AI. The last question reveals judgement faster than any portfolio.
How do I know if an AI vendor has real experience?
They answer failure questions specifically and without hesitation, because production failures are memorable. They quote running costs at your volume, describe a concrete fallback for provider outages, can show a logged case where the model was wrong, and will sometimes tell you that part of your requirement should be plain automation instead. Vendors with only demo experience answer these in generalities.
Will my data be used to train the AI model?
It depends on the provider, the plan and the contract, which is why the answer must be a specific written statement rather than a reassurance. Major providers offer business terms where API traffic is not used for training, but this varies and changes. Under the DPDP Act 2023 you are the data fiduciary, so a vendor's vagueness here becomes your legal exposure.
Should I hire an AI specialist or a general software company?
Prefer whoever has shipped and then supported the specific thing you need. Most AI projects are mostly ordinary software engineering, with integrations, permissions, logging and a user interface surrounding a comparatively small model call, so a team that cannot build and run reliable software will not be rescued by model expertise. Ask who will answer the phone in month thirteen.