TL;DR

  • A traditional data governance audit checks the data a model was built on. It rarely checks the data an AI model keeps consuming after launch, and that gap is where the real exposure sits.
  • Four input types slip past most audits by default. Unstructured text, real-time streams, synthetic data, and third-party or alternative data each enter a model’s pipeline without passing through the same onboarding checkpoint structured data goes through.
  • The most dangerous gap is the one nobody approves. A model’s own outputs feeding back into its training data are not a new data source anyone signs off on, so they sit outside governance entirely.
  • Regulators are already signaling where this is heading. The Basel Committee has told bank boards to intensify oversight of data governance and risk data aggregation, a signal ahead of any binding rule for a mid-tier bank.
  • A short checklist separates a bank that can name every input feeding its AI models from one that can’t. Five items, covered below.
  • This piece covers the inputs themselves. Read Data Governance Models for AI-Driven Institutions for the governance structure, centralized, decentralized, or federated, that decides who is accountable for auditing them.

A bank that passed its last data governance audit with no findings has good reason to trust its data. That confidence covers less ground than it used to. Most data governance programs were built to audit the data a model was trained on: sourced once, documented, approved. An AI model keeps consuming data long after that approval, often from sources that never went through the same checkpoint. This piece names the input types that slip through, the one gap that grows without anyone approving it, and what a governance program has to add to cover what an AI model runs on today.

Why traditional data governance misses the inputs an AI model uses

A traditional data governance program is built around a discrete event: a data source gets identified, documented, and approved before a model is trained on it. That model’s inputs are then fixed until someone deliberately retrains it, and the audit trail follows that same fixed shape, a list of approved sources, checked periodically.

An AI model doesn’t consume data on that schedule. A generative model can pull from a document repository that grows every day. An agentic system can pull a live data feed mid-decision. Neither event looks like the deliberate, documented retrain a traditional audit is built to catch, so neither one triggers the checkpoint that would normally bring a new input under governance.

Four input types that slip past a traditional data governance audit

Unstructured text. Collection notes, customer emails, call transcripts, and scanned documents increasingly feed generative models directly, without the schema, field definitions, or source documentation a structured data source requires before it clears governance.

Real-time streams. A transaction feed or a behavioral data stream can inform a decision the moment it arrives, well before a periodic data-quality review would ever touch it. By the time the next audit cycle runs, months of that input have already shaped decisions with no record of what came through.

Synthetic data. Data generated to fill a gap in training data, common where labeled examples of a rare event like fraud are scarce, doesn’t exist in the source systems a traditional audit checks. Its accuracy and representativeness depend entirely on the process that generated it, a process most governance frameworks were never built to review.

Third-party and alternative data. A vendor feed gets vetted once at onboarding. The vendor’s own data quality, coverage, and methodology can drift afterward, and a governance program built around a one-time vendor approval has no mechanism to catch that drift unless someone builds one in.

The feedback loop: the input gap that grows on its own

The gap that matters most doesn’t come from outside the institution at all. A model’s own output, a collections outcome, a fraud flag, a churn prediction, often becomes part of the data that trains the next version of that same model or a related one. Nobody signs off on this the way they’d sign off on a new vendor feed, because it doesn’t look like a new data source. It’s the model’s own decisions becoming tomorrow’s training data.

That matters because a model trained partly on its own prior decisions can reinforce whatever bias or error existed in those decisions, and it does so without ever crossing the approval checkpoint a genuinely external input would have to clear. An institution that governs its third-party feeds carefully while treating this loop as internal plumbing has closed the wrong gap.

What BCBS 239 signals about board-level expectations, and why it matters before it is binding

The Basel Committee on Banking Supervision’s November 2023 progress report on its Principles for effective risk data aggregation and risk reporting states plainly that “bank boards should prioritise and intensify their oversight of data governance, including the development, implementation, and maintenance of robust data governance frameworks, risk data aggregation and reporting.”

BCBS 239 itself formally applies to the 31 global systemically important banks the Committee has designated. It doesn’t bind a community bank or credit union directly. That doesn’t make the signal irrelevant to a smaller institution. It shows where board-level expectations for data governance are heading at the top of the industry, and an examiner reviewing a mid-tier bank’s data practices is reading the same industry direction the board members at the largest banks are already being told to intensify oversight of.

A short checklist: which inputs feeding your AI models have never been audited

  • Every unstructured data source (documents, transcripts, emails) a generative model draws on, documented the same way a structured source would be.
  • Every real-time or streaming input, with a monitoring cadence that matches how often it changes, separate from the periodic schedule built for static sources.
  • Any synthetic data in the training set, with the generation process itself documented and reviewed.
  • A refresh check on every third-party or alternative data feed, beyond the one-time vendor approval at onboarding.
  • A named checkpoint for a model’s own outputs re-entering training data, so the feedback loop gets the same scrutiny a new external source would get.

A bank missing two or more of these can pass a traditional data governance audit while still running an AI model on inputs nobody has reviewed.

How iTuring extends lineage beyond the inputs a traditional audit checks

iTuring’s Data Accelerator tracks lineage for every input a model consumes, structured or unstructured, one-time or streaming, external or generated by the model’s own prior outputs, using the same audit trail regardless of which of these categories the input falls into. A feedback loop that feeds a model’s own outputs back into training data is tracked with the same lineage as a newly onboarded vendor feed, so it never becomes the kind of gap that only surfaces when an examiner or a validator asks where a specific input came from.

The same lineage discipline supported the 116% improvement in collections recovery iTuring delivered for a leading NBFC in India, where every input feeding the model, including the outcomes of its own prior recommendations, carried a documented source from the start.

An audit that only checks the data a model was trained on misses most of what an AI model runs on today. Book a 15-minute discovery call to see how iTuring tracks lineage for every input your models consume, including the ones a traditional audit never reaches.

Sources

  1. Basel Committee on Banking Supervision, “Progress in adopting the Principles for effective risk data aggregation and risk reporting,” Bank for International Settlements, November 2023. https://www.bis.org/bcbs/publ/d559.pdf
  2. iTuring.ai case study, “Improve Collections and Optimize Efforts.” https://ituring.ai/case-study/improve-collections-and-optimize-efforts/
  3. iTuring.ai, “Data Governance Models for AI-Driven Institutions.” https://ituring.ai/data-governance-models-compared/