Text to SQL Solution in India: costs, delivery models and data rules

A text to SQL solution built in India costs ₹8 lakh to ₹25.6 lakh, or $12,500 to $38,500, for a production deployment covering two or three data domains. The spread is driven by warehouse readiness, the number of metric definitions that must be written, and whether data residency rules force a self-hosted model rather than a hosted API.

This article sets out what the money buys, how the four common delivery models compare, which Indian data rules actually bite on a querying layer, and how to judge a Bengaluru partner against a global consultancy without relying on rate cards alone.

What you are buying, and why the range is wide

A text to SQL solution is a system that turns a business question typed in English or an Indian language into validated SQL, executes it against your warehouse under the asker's own permissions, and returns the answer with the query shown. The AI part is the smallest part. The expensive part is agreeing what your business means by its own words.

That is why two companies with identical headcount get quotes that differ by three times. If your finance team already has a dbt project with tested metric definitions, a large slice of the work exists. If "revenue" means one thing in the billing system, another in the CRM and a third in the board pack, the project starts with reconciliation, and that is data work priced as data work.

Indian deployments add two specific costs more often than Western ones: multilingual questions, where a query arrives in Hindi or Kannada and must be resolved against English column names, and reporting entities that split by GST registration, which quietly doubles the number of dimensions the semantic layer has to carry.

The third variable is language coverage in the interface rather than the query. A sales manager in Coimbatore who types a question in Tamil expects the answer, the axis labels and the caveat text to come back in Tamil, and that is presentation work layered on top of the SQL generation. It is not difficult, but it is rarely in the first estimate.

Which delivery model should you choose?

The short answer is that a domestic product engineering partner is the best value for a first production deployment, an in-house build makes sense only if you already employ data engineers with spare capacity, and a global consultancy is worth its premium only when a group-level contract requires it.

The column that decides most arguments is the fourth one. Whoever owns the semantic layer owns the ability to change what your numbers mean, and if that is a vendor, every metric change becomes a change request. The general pricing picture is covered in software development pricing in India versus the US.

The data rules that actually apply

DPDP Act 2023

The Digital Personal Data Protection Act 2023, published by the Ministry of Electronics and Information Technology, governs the processing of digital personal data in India and requires a lawful purpose, notice, security safeguards and limits on retention. The text and associated material are on the MeitY data protection framework page. For a querying layer the practical consequences are narrow but firm: personal data must not leave the access boundary of the person asking, and you must be able to show who asked what.

That translates into two engineering requirements. Queries execute under the identity of the user, not a shared service account, which in Postgres means row-level security rather than application-side filtering. And every question, generated query and result count is logged. We cover the pattern in row-level security for AI analytics and the compliance detail in text to SQL solution, security and the DPDP Act.

Sectoral rules: RBI, IRDAI and health data

Banks, NBFCs and insurers carry obligations beyond DPDP, including outsourcing controls, audit access and in some cases storage of payment data within India. If you are regulated, the question is not whether the model is clever but whether your auditor can trace a decision. Practically, regulated clients either self-host an open-weight model or restrict the querying layer to aggregated, de-identified marts. Both are workable. What does not work is discovering the constraint after the system is live and being asked to retrofit an audit trail onto a design that never had one.

Residency and where the model runs

Schema names, column names and sample rows are what a text to SQL system sends to a model, and schema is often more sensitive than people assume: a column called restructured_loan_flag tells a story on its own. Where residency is required, the options are a model hosted in an Indian region, or a self-hosted open-weight model on your own infrastructure through private agentic AI from $31,500 or ₹20.8 lakh plus infrastructure. Neither is exotic now, but both change the budget, so decide early. The data residency glossary entry is a short briefing for non-technical stakeholders.

GST, contracting and currency

We invoice Indian clients in INR with GST and international clients in USD, and the same scope carries the same fixed price in both. For Indian buyers the practical points are that the fixed-price structure suits a capex approval cycle, that model usage runs on your own accounts so it lands as opex you can see, and that a multi-entity group usually needs the reporting entity dimension agreed before the build rather than after. Groups with several GST registrations and a shared chart of accounts should expect the entity dimension to appear in almost every generated query, which makes it a semantic layer decision rather than a reporting preference.

How to judge a domestic partner

• Ask for a failed question list from a previous deployment, not a demo on a sample database.
• Check who writes the golden question set: it should be your analysts, with the vendor measured against their answers.
• Confirm the semantic layer is yours in the contract, in a portable format, with documentation.
• Test the permission story: ask what happens when a regional manager asks a national question.
• Require your own model and warehouse accounts, with budgets and dashboards configured for you.
• Ask about IST overlap if your data team sits in Europe or the US, and get the working hours in writing.
• Insist on a query cost ceiling: row limits, statement timeouts and an estimate before execution.
• Agree the acceptance gate as a percentage on a named question set, before the build starts.

Cost, timeline and what we charge

Our natural language data querying programme starts at $12,500 or ₹8 lakh and runs to $38,500 or ₹25.6 lakh. A ten-day Sprint Zero at $3,250 or ₹2 lakh, credited against the build, produces the metric inventory, the permission model and the first golden question set, which is usually enough to make the full quote precise. Typical builds take eight to twelve weeks, including a period where the system answers alongside existing dashboards and is measured against them.

After launch, Care Plans run from ₹68,000 a month for business-hours cover in IST to ₹3.4 lakh a month for round-the-clock cover with a named engineer, plus ₹40,000 for the AI add-on that covers evals, prompt regression and cost monitoring. Figures are on the pricing page, and our delivery base is described on the Bengaluru page.

When building in India is the wrong choice

If your contract with a European customer forbids any processing outside the EU, an Indian build team can still design and deliver, but the running system has to sit elsewhere, and that is a conversation to have before the statement of work, not during user acceptance testing. If your organisation has no data owner who can settle metric disputes, no delivery model in any country will save the project.

And if you have fewer than twenty users asking a stable set of questions, the honest recommendation is a good set of dashboards. We say no to this build more often than people expect, because a querying layer earns its cost through the questions nobody thought to build a dashboard for.

A comparable engagement

Regulated Indian deployments usually start with the constraint rather than the feature. Our KYC document intelligence work with an NBFC describes private document intelligence built for loan onboarding, and the architectural instinct carries over: keep sensitive material inside the customer's boundary, prove access controls, and make every step auditable before anyone discusses model choice.

Related reading

How much does it cost to build an AI product in India gives the wider budget picture, DPDP Act 2023 and AI explains the obligations in plain terms, and the semantic layer post covers the artefact that decides whether your build is cheap or expensive.

Buy the semantic layer and the permission model; the natural-language part is the easiest thing on the invoice to replace later.