A fixed quote before any discovery
Someone quoting ₹2 lakh for "an AI chatbot" without asking what it connects to is guessing. You'll pay the difference later in change requests.
Buyer's guide · Updated August 2026
How to choose an AI development company in India
Right now almost every agency in India has an AI page. Most of them are calling an API and charging for engineering. These ten questions tell you which is which in about forty minutes — and we're handing them to you knowing you might use them on us.
Short answer
Ask to see one AI system running in production, ask how they measure whether it works, and ask what breaks first as it scales. Teams that have actually shipped answer all three in specifics within a minute. Then run a paid two-to-four week proof of concept on your own messy data with the pass mark agreed in writing beforehand.
The shortlist call
Ten questions, and what the answers mean
You don't need to be technical to use these. You're not judging whether the answer is correct — you're judging whether it's specific. Vagueness is the signal.
- 01
“What happens when the model gets it wrong?”
Every AI system fails sometimes. This question separates people who've shipped one from people who've demoed one.
Good answer
They describe confidence thresholds, fallback paths, and where a human reviews output before it reaches a customer.
Walk away
"Our accuracy is 99%." That's a benchmark number, not a failure plan.
- 02
“How will you measure whether this works?”
Without evals you have no way to know if a change improved the system or quietly broke it.
Good answer
They talk about a test set built from your real data, and a score they can re-run after every change.
Walk away
They talk about how impressive the demo feels, or point at the model provider's benchmarks.
- 03
“What does this cost to run per month at our volume?”
Token costs scale with usage. Plenty of AI projects launch fine and become unaffordable at real traffic.
Good answer
They ask your expected volume first, then give a range, and explain what drives it up.
Walk away
They can't answer, or wave it off as negligible without asking a single question about volume.
- 04
“Who owns the code, the prompts and the data?”
Some agencies keep the system on their infrastructure so you can't leave. Find out before you sign, not after.
Good answer
You own everything. It's in the contract. They'll hand over the repository.
Walk away
Hosted only on their platform, or the prompts are described as their proprietary IP.
- 05
“Show me something you built that's live right now.”
The single most useful question on this list. Anyone can build a demo in a weekend.
Good answer
A working URL or a screen-share of a production system, with honest notes on what's still rough.
Walk away
Only slide decks, mock-ups, or a video that never shows the system handling an unexpected input.
- 06
“What data do you need from us, and where does it go?”
If your customer data is going into a third-party API, you need to know which one and under what terms.
Good answer
A clear list, named providers, retention terms, and whether anything is used for training.
Walk away
Vague reassurance that it's "completely secure" without naming a single provider.
- 07
“What would make you tell us not to build this?”
A vendor who has never turned down work will sell you AI whether or not you need it.
Good answer
A real example of a project they declined, and the reason. Often it's "a rules engine would've done this".
Walk away
"AI can solve any problem." You're talking to sales, not engineering.
- 08
“Who is actually writing the code?”
In India specifically, the person in the pitch is frequently not the person on the project.
Good answer
Named engineers, their experience, and you meet them before signing.
Walk away
"Our team of 50+ experts." Nobody named, no CVs, no introductions.
- 09
“How do we hand this over if we part ways?”
Ask on day one. Asking during a dispute is too late to negotiate anything.
Good answer
Documented handover, repository access, a runbook, and a transition period in the contract.
Walk away
They've never thought about it, or it becomes an awkward conversation.
- 10
“What breaks first as we scale?”
Good engineers know the weak point in their own design and will tell you. It's the strongest signal on this list.
Good answer
A specific answer — retrieval quality at document volume, rate limits, latency under concurrency.
Walk away
"Nothing, it scales infinitely." Nothing scales infinitely.
Red flags
Six things worth ending the call over
Any one of these on its own is a conversation. Two together, and you should be looking at someone else.
Buzzwords instead of architecture
If you ask how it works and get "advanced machine learning algorithms", ask again. Real engineers will happily draw you the data flow.
No production references
An AI practice with nothing live is learning on your budget. That can be fine — at a discount, with your eyes open, never at full price.
Accuracy claims without a test set
"95% accurate" means nothing unless they can tell you accurate against what, measured how, on which data.
Refusing a paid pilot
Any competent team will take a scoped, paid proof of concept. Insisting on a full engagement upfront is a confidence problem — theirs.
The pitch team vanishes after signing
Ask who runs the project on day one. Get the names in writing. This is the most common complaint we hear about the previous vendor.
Before the big contract
Run a paid pilot. It's the cheapest information you'll buy.
Two to four weeks and a couple of lakh tells you whether a team can actually build. Finding out six months into a ₹20 lakh engagement costs a great deal more than that.
Run it with two vendors at once. Same task, same data, same pass mark. The difference in what comes back will be obvious, and it costs less than one bad hire.
- 1
Pick the narrowest useful problem
Not "an AI assistant for our business". Something like: read incoming supplier invoices and pull out five fields correctly. One task, one measurable output.
- 2
Give them your messy real data
Not the clean sample. The scanned PDFs, the mixed-language entries, the ones your own team argues about. Clean data proves nothing.
- 3
Agree the pass mark before they start
Write it down: "correct on 90 of 100 real invoices". Deciding what counts as success afterwards always ends in an argument.
- 4
Pay for it, and cap it at 2–4 weeks
Free pilots get junior attention and no deadline. Paid ones get engineers. ₹75,000 to ₹2,50,000 is normal, and it buys you the truth cheaply.
- 5
Ask for the evaluation, not just the demo
The demo shows the wins. The evaluation shows the failures. A team that hands you both without being pushed is the team to hire.
Questions
What buyers ask before shortlisting
Working through a vendor decision? Send us the shortlist — we'll tell you what we'd ask, even if we're not on it.
How do I know if an AI development company is actually capable?
Ask to see a system running in production right now, and ask what breaks first as it scales. Teams that have shipped AI will answer both immediately and specifically — they'll name the weak point in their own architecture. Teams that have only built demos deflect to accuracy percentages, client logos, or team size. One production URL tells you more than an hour of pitching.
What does AI development cost in India?
A scoped proof of concept costs ₹75,000 to ₹2,50,000 over two to four weeks. A production AI agent or RAG system typically costs ₹5,00,000 to ₹25,00,000 depending on integrations and data complexity. Budget separately for monthly running costs — model API usage, hosting and monitoring commonly run ₹10,000 to ₹1,00,000 a month depending on volume. Any quote that ignores running cost is incomplete.
Should I hire a specialist AI company or a general development agency?
It depends on where the difficulty sits. If the hard part is the AI — retrieval quality, evaluation, agent reliability — hire specialists. If the hard part is the surrounding product, and AI is one feature inside it, a strong product team that understands AI serves you better. The worst outcome is a general agency treating AI as an API call, because that's exactly what breaks at real volume.
What should be in an AI development contract?
Five things: clear ownership of code, prompts and data; the model providers named explicitly along with data retention terms; acceptance criteria stated as measurable evaluation scores rather than opinions; a documented handover process with repository access; and who covers model API costs during development. Missing ownership terms and undefined acceptance criteria cause most disputes we hear about.
How long does an AI project take?
A proof of concept takes two to four weeks. A production system takes three to six months, and that time goes mostly into evaluation, edge cases and integration rather than the model itself. Any vendor promising a production-grade AI agent in two weeks is either shipping a demo or planning to skip the evaluation work that keeps it reliable.
Is it worth building AI if we're a small business?
Only when you have a repetitive task with enough volume to matter and enough consistency for a system to learn the pattern. Answering the same forty customer questions daily qualifies. A task you do twice a month does not — a checklist is cheaper and more reliable. The right question isn't "can AI do this" but "does this happen often enough to be worth automating".
Use this list on us
Bring all ten questions to the first call. We'll show you something running in production, tell you what breaks first, and say so plainly if a rules engine would solve your problem for a tenth of the price.
Book a technical call
