Hiring a Head of Data Science and AI in fintech

What the role owns

Credit and risk models. Scorecards, underwriting models, limit and pricing models, early warning. These sit under the Chief Risk Officer's accountability for outcomes, which makes the working relationship between the two the most important one in the job.

Fraud and transaction risk models. Inline, latency-bounded, adversarial. A different discipline from credit modelling and frequently a different team, described in real-time fraud and risk decisioning.

Collections and recovery models. Propensity, channel allocation, settlement optimisation. Commercially valuable and conduct-sensitive at the same time.

Growth and personalisation models. Targeting, cross-sell, pricing experimentation, churn. The least regulated part of the remit and often the largest in headcount.

The platform. Feature store, training and serving infrastructure, experiment tracking, the model registry. Sometimes owned here, sometimes with engineering, and the boundary needs to be explicit before the hire.

Model governance. Validation, documentation, approval, monitoring, drift detection and the criteria for withdrawing a model from production. This is the part that separates a regulated-sector leader from a general one, and the part most often missing.

AI adoption more broadly. Increasingly the role is also asked to own large language model applications across service, operations and engineering productivity, which is a genuinely different problem from predictive modelling and should be scoped deliberately rather than assumed.

Why governance is the job

In an unregulated business, a model that underperforms costs money and gets retrained. In a lender, a model that underperforms has declined people who should have been approved, approved people who should not have been, and produced a book whose losses arrive over the following four quarters. The regulator holds the institution accountable for that, and the institution's defence is its governance.

What that requires in practice:

Validation independent of the builder. Someone other than the modelling team has to be able to say the model is fit for use, and the organisation has to act on it when they say otherwise.

Lineage and reproducibility. Which model version, trained on which data, with which features, approved by whom, deployed when. Without this, a decision cannot be explained and the lending stack's replay requirement cannot be met.

Monitoring with thresholds and consequences. Drift in score distribution, feature distribution and outcome performance, with defined levels at which somebody is obliged to act.

Withdrawal criteria agreed in advance. The single most revealing thing to ask a candidate. What would make you pull a model out of production, and who decides? Teams that have not agreed this in calm conditions will not agree it in a crisis.

Explainability proportionate to the decision. A declined loan application needs a reason code the customer and the regulator can both follow. A marketing propensity score does not. Treating these identically either cripples the growth work or exposes the credit work.

Fairness and alternative data scrutiny. Indian credit models draw on bank statements, device signals, employment and account aggregator feeds because bureau data is thin for many borrowers. Each source carries reliability and bias questions, and the leader has to have examined them rather than assumed them away.

Which backgrounds transfer

The fourth row is where most of the hiring decisions get made and where most of the failures come from. Candidates from large consumer internet companies are frequently the strongest modellers available and arrive with an instinct that a better model is always the answer. In a regulated lender, a better model that cannot be explained, validated or withdrawn on agreed criteria is worse than a simpler one that can. Candidates who absorb that quickly do very well; candidates who treat it as bureaucracy create exposure.

How to assess

Ask about a model they withdrew. The best question for this seat. Which model, what triggered it, who decided, what replaced it in the meantime. Leaders who have governed models have done this. Those who have only built them have not.

Ask how they would validate a model they did not build. This is the day job in a regulated entity. Strong answers cover validation data, challenger comparison, stability testing, monitoring design and the conditions for rejection.

Probe training and serving skew. "Has a model ever performed worse in production than in backtest, and why?" Everyone who has deployed at scale has this story, and the answer reveals how deep their production experience actually goes.

Test the alternative data reasoning. "Which of your data sources worried you most, and what did you do?" Candidates who have thought about bias and reliability have a specific answer. Those who have not will talk about accuracy.

Check the CRO relationship. "The Chief Risk Officer does not accept your model. What happens?" The right answer treats that as legitimate and describes a process. The wrong answer treats it as an obstacle.

Scope the generative AI question honestly. If the role includes large language model applications, ask what they have actually deployed and what it cost. The gap between pilot and production is wide and most organisations are still on the wrong side of it.

The market right now

Adoption is broad and shallow. A Zeta survey of 40 CXOs across 18 Indian banks and NBFCs, published September 2026, found 70% of chief data officers reporting AI at selective or scaled deployment and only 30% at scale, 79% with real-time data platform or API-first adoption, and 80% of technology leaders using AI in testing and quality assurance. It also found security and data privacy as the leading barrier and roughly 40% of chief data officers unable to identify high-return use cases. The capability is being bought faster than the judgement about where to apply it, which is precisely the gap this role exists to close.

The governance framework is still forming. The RBI's committee on a Framework for Responsible and Ethical Enablement of AI reported in August 2025, and the recommendations remain under consideration rather than operative. Hiring for AI leadership in Indian financial services is therefore running ahead of the rules it will eventually have to satisfy. The sensible posture is to hire someone who builds governance that will survive the framework rather than one who waits for it.

Compensation competes with technology, not with banking. Published benchmarks put a Head of Data Science at roughly ₹1.2 crore to ₹2.5 crore at senior fintech scale, with machine learning engineers at five to eight years in the ₹38 lakh to ₹65 lakh range. These candidates hold alternatives at product companies and global capability centres, and offers calibrated to a financial services banding exercise lose.

The pool is deep in modelling and thin in accountability. India has an abundance of capable data scientists. The subset who have been personally accountable for automated decisions in a supervised entity is small, and that subset is the actual market for this role.