What AI costs in banking and insurance: budgets that hold up

A first production AI system in banking or insurance usually costs between $12,500 and $45,500 to build, or ₹8 lakh to ₹30.4 lakh, depending on how many core systems it touches. The line that breaks budgets is rarely the model. It is integration with the core platform, the human review desk that regulation implies, and the audit evidence you need from day one.

Below is how we price this work, which use cases sit at which end of the range, the five cost lines that naive budgets omit entirely, what running one costs each month, and the sequence that stops a board approving a number that turns out to be a third of the real one.

Why BFSI budgets behave differently

In most industries the AI budget is a build cost plus an inference bill. In banking and insurance three additional forces push the number up, and pretending otherwise is how projects get cancelled in month five.

The first is integration surface. A claims triage agent that sounds like one system is a conversation with a policy administration platform, a document store, a payments rail and a fraud engine. Core banking and policy systems rarely offer clean APIs, so the integration is often the largest single engineering line in the estimate.

The second is the review layer. A model that classifies a document, scores a risk or drafts a customer communication cannot act unsupervised on day one. You are budgeting for an exception queue, the people who work it, and the tooling that lets them work it fast. Teams that skip this line discover it in user acceptance testing, when the number is hardest to add.

The third is evidence. Regulated deployments need to show which model version produced which output from which inputs, with retention. That is an engineering requirement, not a policy document, and it belongs in the build estimate. Our overview of AI in fintech and BFSI sets out how these constraints shape architecture rather than merely adding paperwork.

What does an AI build cost in banking and insurance?

These are Eazyware's published starting prices and the use cases they typically carry in this sector. Every figure appears on the pricing page; ranges widen with the number of systems touched, not with the sophistication of the model.

The five lines a naive budget misses

When a quote comes in dramatically below these numbers, one or more of the following has been left out. Each is unavoidable in a regulated deployment.

• Core system integration. Reading from and writing to a core banking or policy administration platform, often through a middleware layer or a batch file, is frequently thirty to forty per cent of the build.
• The exception desk. Somebody works the queue of low-confidence cases. Budget the tooling and the headcount, and expect the queue to be largest in the first quarter.
• Audit and lineage. Immutable records of input, model version, prompt version, output and human decision, retained for the period your regulator expects.
• Deployment mode. If data cannot leave your perimeter, you are costing GPUs and the people who run them, not an API bill. That decision is unpacked in private AI for banks.
• Model change management. Providers deprecate models. Regression suites, re-evaluation and a re-approval path are recurring work, not a one-off.

What does it cost to run each month?

Three costs run in parallel after launch, and in BFSI the token bill is usually the smallest of them.

Inference

For a document or classification workload at realistic volumes, hosted model usage typically lands in the hundreds of dollars a month rather than the thousands, and routing cheaper models at the easy cases cuts it further. You pay this through your own provider accounts, which keeps the spend visible. Model it against your actual volumes with the LLM inference cost calculator before assuming either extreme.

Self-hosted infrastructure

If residency or policy pushes you to open-weight models inside your own account, GPU instances are a fixed monthly commitment whether or not volume arrives. This flips the economics: self-hosting rewards high, steady throughput and punishes pilots. A self-hosted agentic deployment starts at $31,500 or ₹20.8 lakh plus infrastructure for precisely that reason.

Maintenance and evaluation

An Eazyware Care Plan starts at $1,000 or ₹68,000 per month for business-hours cover, $2,500 or ₹1,60,000 for 24x5 and $5,250 or ₹3,40,000 for 24x7 with a named engineer. AI systems add $750 or ₹40,000 per month, which buys evaluation runs, prompt regression, cost monitoring and re-indexing. In a regulated environment this is not optional: a retrieval system quietly degrades as documents change, and nothing in your uptime dashboard will tell you.

Sequencing spend so the budget holds

The budgets that survive review are staged, and each stage buys information that reprices the next one.

Start with a three-week AI POC Sprint at $6,250 or ₹4 lakh on the single hardest case: the document type your operations team argues about, the claim category with the most rework. A proof run that fails is cheap information. A proof run that passes gives you an accuracy figure to put in the business case instead of a vendor's benchmark.

Only then commit the build, and commit it one workflow at a time. Launch in shadow mode, where the system proposes and a person disposes, until the acceptance rate justifies moving the approval threshold. The gap between a demo and something that survives an audit is most of the project, and our note on total cost of ownership for AI systems puts numbers on the years after launch.

When the honest answer is not to spend

Three situations where we advise against a build, and say so before quoting.

When the process is genuinely deterministic. Rule-based eligibility checks, tariff calculations and mandate validations do not need a language model. Automation, yes. AI, no. The rules engine is cheaper, faster and explainable without effort.

When the data is not there. If your claims history lives in scanned PDFs with no consistent structure and no labelled outcomes, the first spend is data engineering. Buying a model first produces a demo and a disappointment.

When nobody will own the exception queue. An AI deployment in a regulated process creates work before it removes work. If operations has not agreed to staff the review desk for the first two quarters, the project has a budget but not a plan.

What this looked like for one lender

An NBFC came to us with onboarding that took days because KYC and income documents were checked by hand. We built document intelligence inside their own cloud account, with confidence thresholds routing uncertain cases to a human queue and every decision recorded against a case ID. The engagement is documented in the KYC document intelligence case study, and the mechanics of accuracy, exceptions and audit are covered in our guide to KYC document processing with AI. The budget held because the exception desk and the audit trail were in the original estimate rather than discovered later.

Before you take it to the board

• Name the single workflow, with its current cost per case measured in rupees and minutes
• List every core system the workflow touches and confirm whether each has an API
• Decide the deployment mode, hosted or inside your perimeter, before you price anything
• Budget the exception desk explicitly, with headcount for two quarters
• Agree who signs off model changes and how often re-approval happens
• Include a Care Plan tier and the AI add-on in the first-year total
• Set the accuracy threshold at which the system may act without approval
• Fund a proof sprint before the build, and be willing to stop after it

Related reading

Our piece on insurance claims triage with AI agents walks the workflow that carries the largest savings in general insurance, and RBI guidelines and AI covers the outsourcing, localisation and audit expectations that shape the architecture. The Reserve Bank of India publishes its master directions on outsourcing of IT services at rbi.org.in, and the audit-rights and exit clauses there are worth reading before you sign any vendor arrangement. If you want a costed sequence for a specific workflow, send us the process.

Budget the integration, the review desk and the evidence, and the model itself turns out to be the cheap part.