Leadership1 publisher2 min readPublished
Trintech's CTO raises the data bar for AI agents that execute finance decisions
Sunil Padiyar, Trintech's CTO, argues that data fit for a human reviewer falls short once AI agents execute finance decisions themselves. In his account, context and decision records become part of data quality, though insisting on perfect data would stall deployment.
The Board Room · Leadership desk
What happened
- Trintech CTO Sunil Padiyar separates reporting-grade data, used by a person to decide, from action-grade data carrying enough accuracy, context and control for a system to act safely.
- In his finance example, a human reviewer adds what the ledger omits: that a variance is normal at quarter-end, that an account is material, or that an unwritten policy applies.
- An agent handling an exception, he writes, must know whether it is reconciled, which policy and materiality threshold apply, where the data came from and whether a human must approve.
- He adds decision lineage to data lineage: a record of what an AI system considered, which rules applied, why it chose an action and what happened afterward.
Compiled by The Board RoomSomething wrong?How this is made
Why it matters
- cost Judgement a controller built over years has to be turned into thresholds, policy flags and approval rules an agent can read, and finance teams do that work before any agent saves them time.
- exposure An entry with correct numbers can still violate policy or bypass a control, and the controller who let an agent post it answers for the error without a reviewer positioned to catch it.
- precedent Once auditors ask why an agent created an entry, evidence and approval records become part of the entry, and lineage built only to trace data sources will fall short.
Padiyar starts from the investment enterprises have already made. For years they have put heavy money into data quality, with the goal of giving people accurate information for better decisions, he wrote [15]. The conversation about preparing data for AI "often centers on cleanliness and accuracy," he wrote, and then he drew the line: "Those things remain fundamental, but action-grade data requires more" [14]. His shortest version of the problem: "Data that is good enough to inform a human decision may not be good enough for a machine to execute one" [3].
The trade-off concerns the person who reviews the work. A company can pay to write that person's judgement into the data, or it can keep the reviewer and let the agent recommend. Padiyar describes the second arrangement as the one enterprises have always run. "People have historically compensated for what the data does not say," he wrote [5]. That holds while a person makes the call. It "becomes much more problematic when AI is making or executing the decision itself" [5].
Trintech sells AI financial close software, according to the author line on Padiyar's column [1]. An argument that customer data must be upgraded before agents can act also describes work that comes before a sale, and buyers can weigh it with that in mind. The audit trail is the part of his case that holds whoever is selling. Padiyar frames it as a design question: "This is where governance stops being something applied around AI and becomes part of the architecture that enables AI" [13].
We do not know yet what action-grade data costs. Padiyar makes a qualitative argument, and he sets his own limit on it: "If organizations conclude that every piece of enterprise data must be perfect before they can deploy agentic AI, very little will ever get deployed," he wrote [10]. In my view, that points the spending at the specific processes where an agent will execute, and leaves reporting-grade data to serve the rest. It also sets the sequence. Whichever processes a finance team lets an agent execute this quarter will need their context and decision records in place by the next close.
What to watch
- Whether external auditors or audit standard-setters publish expectations for documenting AI-created journal entries, including approvals and supporting evidence.
- Whether Trintech or other financial close vendors disclose what context-encoding work customers completed before agents were allowed to execute.
- Whether finance teams running agents report entries that were numerically correct but broke policy or skipped a control.