Hiring an AI development company comes down to verifying three things: they have shipped AI to production and kept it there, they can say no to you, and you can check their claims yourself. Everything in this guide, the ten questions, the red flags, the scoring rubric, is machinery for testing those three, and every test in it works on us too. We would not publish a checklist we fail.
In 2026 every agency in India is an "AI development company", which is precisely why the filter matters: the label now carries no information, and the verification has to come from you.
Ten questions, and the answers you want to hear
Ask these in the first call, in roughly this order. The pattern in the answers matters more than any single one.
- "What have you shipped to production, and is it still running?" You want named products or clients, and the phrase "still running". Demos and hackathon wins are a different trade.
- "What does it cost to run per month?" Teams that have operated AI answer in numbers and caveats; teams that have only built it answer in shrugs. The whole cost structure is public knowledge now, we published ours, so silence here is a tell.
- "How do you measure whether the AI is right?" The words you want: evaluation set, test cases, baseline, threshold. The words you do not: "we test it thoroughly".
- "What happens when the model is wrong?" Every honest answer includes hand-off to a person, action limits and logging. A vendor who says their system does not make mistakes has never run one.
- "When would you tell us not to use AI?" The single most predictive question on this list. A firm that cannot name a case where they said no has never said it.
- "Who owns the code, prompts, data and evaluation sets?" The only acceptable answer is you, on final payment, including the third-party accounts. Anything else is a switching cost being built in front of you.
- "Which models do you use, and how do you decide?" Good answers name trade-offs and testing on your data; bad answers name one vendor for everything, which usually means a reseller margin rather than an engineering decision.
- "Where does our data go?" You want a clear map: what leaves your systems, what is stored where, what the model provider sees, and an NDA offered without prodding.
- "Who exactly works on our project?" Senior people sell and juniors build at plenty of firms. Ask to meet the engineer, not the account manager.
- "What happens after launch?" AI features degrade without maintenance, models deprecate, prompts drift. A firm with no post-launch answer is planning to hand you a slowly-expiring product.
Red flags, and what each one actually means
| The flag | What it usually means |
|---|---|
| Guaranteed accuracy or "no hallucinations" | They have not run AI in production, or hope you have not |
| A quote before any hard questions | The real price arrives later, as change requests |
| Everything is "agentic" | Marketing has outrun the engineering |
| No questions about your data | The most failure-prone part of your project is being skipped |
| Case studies with no numbers and no names | Possibly real, unverifiable by design |
| They keep the repository until full payment of everything ever | Leverage, not partnership |
One softer flag worth naming: a portfolio that is all chatbots. It is the easiest AI project to demo and the least representative of the harder work, integrations, evaluation, guardrails, that decides whether production survives contact with reality.
How to verify claims without being an engineer
This is the section most hiring guides omit, because it invites scrutiny. Ten minutes per vendor:
Check their products are real. If they claim their own AI products, open them, sign up, use them. A product page that 404s or a "coming soon" tells you what their production experience is. Ours are listed with links; the same test applies.
Run a Lighthouse check on their own website. It is a free test in any Chrome browser, and it answers a simple question: does this firm apply its claimed engineering standards to the one product it fully controls? We rebuilt ours when it failed that test and published the before and after, scores included, precisely so this check works on us.
Call one client. Not the written testimonial, a phone call. Ask what went wrong during the project, because something always does, and how the firm handled it. That answer is the entire relationship in miniature.
Ask to see an evaluation report. Any firm that measures AI quality has produced one and can show a redacted example. Blank looks here mean question 3 was answered with vocabulary rather than practice.
Pricing models, and when each one serves you
Fixed price after paid discovery fits defined projects: you pay a small amount for a week of hard questions, then a number that does not move. The discovery fee filters both sides, which is the point. This is our model, for reasons this guide has probably made obvious.
Time and materials fits genuinely exploratory work where scope cannot be honestly fixed, and it demands more supervision from you; insist on weekly demos and the right to stop.
Dedicated team fits ongoing product work past the first project, priced monthly per person; the trade-offs are on our dedicated teams page.
The model to refuse is the fixed price quoted in the first call, before anyone has asked about your data or your edge cases. It is not a price, it is an anchor, and the difference will be recovered from you later.
What a good discovery produces, so you can judge one
Since discovery is the gate between you and a fixed price, know what you are owed at the end of it. A real discovery week produces, in writing: the scope with its exclusions named, the data map for anything AI touches, the evaluation plan with a baseline number, the risk list with the expensive edge cases priced in, and a fixed quote with a timeline whose assumptions are visible. If any of those is missing, the price you were quoted still contains guesses, and guesses are always priced in your disfavour.
A useful smell test: discovery should generate questions that are annoying to answer, about your refund edge cases, your data retention, who owns which decision. A pleasant, frictionless discovery is a vendor gathering requirements for a demo, not a build.
The contract terms that matter more than the price
Four clauses decide how bad the bad scenario gets, and they cost nothing to ask for up front.
- IP assignment on payment, covering code, prompts, evaluation sets and designs, with third-party accounts, stores, cloud, model providers, registered to you from day one.
- A data schedule naming what the vendor may access, where it may be processed and what is deleted at exit. In AI work this is not boilerplate; it is the substance.
- An exit clause with handover, a fixed number of hours of documented transition help, priced now, not negotiated during a dispute.
- Warranty and defect terms, a defined period where broken means fixed without a change request, so the launch-week bug is not a new invoice.
None of these is exotic; every serious firm in India signs them routinely. The ones who resist are answering your question 6 honestly, just not in words.
A scoring rubric you can copy
Score each vendor 1 to 5 per row, multiply by the weight, and total. The weights encode a decade of watching where these projects actually fail.
| Criterion | Weight | What a 5 looks like |
|---|---|---|
| Production evidence | x3 | Named, running systems you verified yourself |
| Measurement practice | x3 | Evaluation sets shown, baselines discussed unprompted |
| Honesty under pressure | x2 | Said "that will not work" at least once in the sales call |
| Ownership and data terms | x2 | Everything yours on final payment, in writing |
| Communication | x1 | The engineer answered plainly, the same day |
Anything under 35 of 55, keep looking. The weighting is the message: production evidence and measurement outweigh the demo, the deck and the discount combined.
Questions people ask
How much does an AI development company cost in India?
Hourly rates for experienced Indian teams run $15 to 40 against $80 to 150 in the US, and serious AI projects typically start in the single-digit lakhs; the running costs are separate and worth understanding before you commit. Cheaper exists, and question 1 above is how you find out what it buys. If you are staffing a team rather than buying a project, the monthly rate by role is in our guide to the cost of hiring developers in India.
Should we hire an AI company or build an in-house team?
For a first project, a firm is usually faster and cheaper than assembling a team before you know what you are building; the sensible sequence is an external build with clean handover terms, then in-house growth on top of it, which is exactly what ownership question 6 protects.
How do we know if we are even ready for an AI project?
Run the 15-minute AI readiness self-test before talking to anyone, including us. A vendor who starts building below a readiness score that justifies it is answering question 5 for you, badly.
What should the first engagement look like?
Small, measured and reversible: a two-week readiness audit or a single scoped feature with an evaluation set, not a six-month platform. Any firm that resists starting small is telling you how the big project would go.
The test that summarises the guide
Every claim in a sales deck is a claim someone chose not to make checkable. The firms worth hiring make the opposite move: numbers you can verify, products you can open, references you can call, and a visible willingness to say no. Apply that test uniformly, us included, and the shortlist mostly builds itself.