A financial agent must preserve the state of work a human has already reviewed instead of silently overwriting it.
2
Recorded execution steps, versioned evidence citations, and auditable evaluation data let reviewers explain how an agent reached a decision.
3
Agents should stop for human review when they change reviewed work, disagree with an earlier run, or use evidence that fails the required rules.
Summary
Anirudh Veeraragavan uses a bank reconciliation audit to explain why financial agents need defensible decision processes. A correct answer is insufficient when an auditor later asks how the agent reached it. He proposes four controls: authorization that accounts for the object's state and prevents changes to human-reviewed work, execution records that capture inputs and outputs for each step, evidence provenance tied to an exact document version and content hash, and evaluation integrity that checks whether the gold data is still correct. A reported 97% score can conceal failures caused by bad labels, changed business rules, or updated audit methods. The agent should escalate unusual cases to a person, with enough context to verify the issue without encouraging automatic approval. Veeraragavan applies the same pattern to credit decisions and anti-money-laundering alerts.
A financial agent needs a defensible process behind its answer
Veeraragavan begins with the audit opinion. The opinion is only as good as the basis behind it, so correctness alone does not settle whether an agent made a reliable decision. In the bank reconciliation example, the agent receives the client's books, reconciling items, evidence, audit methodology, prior audits, and previously reviewed evaluations. A human reviewer still signs off. That workflow can look successful until an audit is selected for inspection and someone asks how the agent reached a decision for one specific procedure and one specific piece of evidence.
Authorization must account for what a human has already reviewed
A human reviewer flags a check, reviews it, and signs it. Later, new information triggers an agent run, and the agent changes the row to no exception. The agent had permission, evidence, and the right conclusion, but it still should not overwrite the reviewed work because the content behind the review changed. Authorization therefore depends on the object, its current state, and the point in time. The agent can draft unreviewed rows as its own suggestion, but it cannot replace a human's authoritative work.
Execution records should make individual agent steps replayable
A later rerun can produce a different result from the original run. Logs showing that the agent fetched evidence, reasoned, and produced an answer do not explain where the difference came from. Veeraragavan proposes a chain of recorded steps, each with an ID, inputs, and outputs. The record should distinguish deterministic code, such as fetching or matching, from model work, such as extracting a field from a PDF or reaching a conclusion. Reviewers can rerun one step, compare it with the earlier run, and identify the exact point where behavior changed.
A citation such as January bank statement, page two does not identify a reliable source. A large engagement may contain several documents with the same name, and the citation may omit the version and uploader. The document can be real without providing enough context to defend the decision. Veeraragavan says a citation is an identity with a specific ID, document version, and content hash. That identity also lets a team enforce evidence rules, such as accepting material sourced directly from the bank rather than a document uploaded by the client.
Previously reviewed audits are not automatically correct gold data. A reviewed example may later be superseded, rely on bad evidence, or reflect an older methodology. Those errors can punish a correct agent or reward an incorrect one. Veeraragavan gives a 97% evaluation result that contains several different problems: some misses come from incorrect evaluation data, while other cases became wrong after business rules or audit methods changed. The headline score cannot show whether the agent is performing better or worse than it appears until the evaluation set itself is checked.
Escalations should be rare and easy to investigate
The agent should stop when it tries to change human-reviewed work, when a new run disagrees with an earlier run, or when evidence violates the allowed rules. Escalations should make the discrepancy and the required investigation clear. They should be easy to verify, though not necessarily easy to approve. Veeraragavan warns that frequent prompts can create automation bias, with people approving everything. In his example, a deposit cleared in 11 business days, and the system calls out what needs investigation. The agent handles most cases, while people focus on exceptions with the needed context.
The controls work together. Without authorization, a later run can change work someone already reviewed. Without execution control, the system may reach the right answer but cannot show how. Without evidence provenance, reviewers cannot tie a wrong answer to a concrete source. Without evaluation integrity, a score such as 97% provides no dependable view of reality. Veeraragavan presents defensibility as an AND condition, because any missing control leaves a different way for an unsupported decision to pass through.
Veeraragavan extends the four questions to other financial decisions. In credit, a team should be able to question why a loan was approved or declined. In anti-money-laundering work, it should be able to inspect the evidence behind closing or escalating an alert. The nouns change across domains, but the controls remain authorization, execution control, evidence provenance, and evaluation integrity. For high-impact financial decisions, saying that the model produced the answer is not enough.