EASE 9.0 · Risk & Resilience · Data Quality & ManagementWhat the Agent Can’t See, It Declines
Agents are only as good as the data they consume.
This is a business reality, not a technical footnote, and it now includes consent, tokenisation, and DPDPA by design.
Feed an agent clean, complete data and it decides well: ten thousand decisions a day, a low NPL rate, satisfied customers. Feed it dirty data at the same volume and the NPL rate climbs, collections is overwhelmed, and customers are angry. Data Quality & Management is the precondition for agents operating at scale.
Why This Matters
The Path to Progress
Data quality work fails when it is treated as a one-time cleanup project with an end date. It succeeds as a standing operational discipline, run in the same five moves on repeat.
An agent that can see a customer's current account but not their salary account elsewhere will assume no formal income and decline a loan that should have been approved.
The Business Case
Gartner estimates the average cost of poor data quality at USD 12.9 million per organization per year, a figure large enough that data quality work pays for itself even before counting the upside. This is no longer a discretionary best practice for Indian banks, either: RBI's draft Guidance on Regulatory Expectations for Data Governance (July 2026) requires every bank and NBFC to stand up a formal Data Function headed at CGM rank or above, with named Data Owners, Stewards, and Custodians accountable for data quality. It is a regulatory mandate a bank now has to satisfy regardless of whether it treats data quality as a business priority on its own.
What "fixing the gaps upstream" looks like in practice is already running in production. Under India's RBI-regulated Account Aggregator framework, live since September 2021, HDFC Bank and ICICI Bank were among the first to go live pulling consented, cryptographically-signed bank-statement and GST filing data directly rather than accepting self-reported income at face value; industry reporting on AA-adopting lenders cites loan-application turnaround falling from the prior 30–45-day norm to days, alongside a meaningfully higher share of new-to-credit approvals. It is the same underlying mechanism this page argues for, already live at scale.
Sources: RBI, Draft Guidance on Regulatory Expectations for Data Governance (July 2026); industry reporting on Account Aggregator adoption at HDFC Bank and ICICI Bank. Adoption-outcome figures as reported, not independently audited.
Where AI Agents Fit
This category is literally the first component of Appice's platform architecture: Sense is the layer that collects and unifies real-time data before any decisioning agent ever runs. No customer ever interacts with Sense directly, and this is the quietest category in the framework. Every other agent in this guide is only as good as what Sense hands it. A human reviewer's tolerance for a stale or missing field does not transfer to a system deciding in milliseconds.
Walk through the specific failure this page opened with: an applicant's core-banking record shows no salary credit, because their salary lands in an account at a different bank. A human underwriter would ask; a Credit Decisioning Agent deciding in milliseconds has no such instinct, and defaults to reading the null field as no formal income. Sense is what closes that gap before the Credit Decisioning Agent ever sees the application: it cross-references the applicant's consented Account Aggregator data, finds the salary credit at the other institution, resolves the conflict against a policy the bank defines (which source outranks which, and by what margin, before a field counts as verified), and hands the Credit Decisioning Agent one reconciled income figure instead of a blank one. Get that resolution policy wrong, and the downstream agent inherits the error at the speed of milliseconds instead of the speed of a human's second glance.
The DPDPA Compliance Layer
- Consent registry. Every data element is tagged with the purpose and duration consented to; an agent checks that tag before using the field.
- Tokenisation at capture. PII is tokenised the moment it enters the pipeline.
- Right-to-erasure workflows. A deletion request propagates automatically across every system holding a token.
The Skeptic's Answer
At most PSU banks, a data-quality programme is where good intentions go to die a slow death: a multi-year cleanup project that gets funded once, misses its own timeline, and quietly becomes an unfunded mandate the moment a more urgent priority appears. A credit head who has watched one of these initiatives stall is right to be skeptical of another one.
What is different this time is that it is no longer discretionary. RBI's draft Guidance on Regulatory Expectations for Data Governance does not ask a bank to run a cleanup project; it requires a named, CGM-rank-or-above Data Function with accountable Owners, Stewards, and Custodians, a standing organizational structure, not a task force with an end date. And the fix itself is not a one-time migration: it is Sense running continuously, flagging gaps the moment they appear rather than at the next audit cycle. A programme with no end date is the design, not a failure of one.
Governance Built Into the Decisioning Layer
Consent checks, tokenisation, and vintage-tagging are enforced inside Appice's decisioning layer itself, at the moment an agent reads a field, not as a downstream audit that catches problems after a decision has already been made. That turns data quality into a continuously improving operational metric, not a multi-year project with an indefinite deadline. A bank that cannot govern its data cannot certify its agents. RBI's draft guidance just turned that from Appice's opinion into a regulatory requirement.