From Reporting-Grade Data To Action-Grade Data

Sunil Padiyar is the Chief Technology Officer (CTO) of Trintech, a global leader in AI Financial Close solutions.

getty​For years, enterprises have invested heavily in improving the quality of their data. The objective was relatively straightforward: give people accurate information so they could make better decisions. AI is changing that equation.

As we move from copilots that summarize and recommend to agents that can actually take action, the standard for enterprise data has to change with them. Data that is good enough to inform a human decision may not be good enough for a machine to execute one.

I think of this as the difference between reporting-grade data and action-grade data. Reporting-grade data helps a person understand what happened and decides what to do next. Action-grade data must carry enough accuracy, context and control for a system to safely act on it.

That may sound like a subtle distinction, but in practice, it is a significant one.

Consider what happens when an experienced finance and accounting professional reviews a reconciliation, journal entry or exception.

The data may tell most of the story, but the person often supplies the rest. They know that a particular variance is normal at quarter-end. They recognize that an account requires additional scrutiny because of its materiality. They understand a company policy that may not be explicitly represented in the underlying data.

In other words, people have historically compensated for what the data does not say. That works when technology helps a human make a decision. It becomes much more problematic when AI is making or executing the decision itself.

An AI agent does not automatically possess the institutional context accumulated by a controller over years of experience. If we expect that agent to act autonomously, the information surrounding the transaction has to carry much more of that context.

The transition from copilots to agents, therefore, changes the minimum acceptable standard for enterprise data.

When organizations talk about preparing data for AI, the conversation often centers on cleanliness and accuracy. Those things remain fundamental, but action-grade data requires more.

Imagine an AI system evaluating a financial exception. The underlying amount might be perfectly accurate. But does the system know whether the transaction has been reconciled? Does it understand the applicable accounting policy? Does it know the materiality threshold? Can it identify where the data originated, what transformations have occurred and whether a human approval is required before the next action?

Those are not simply data-quality questions; they are questions of context. For autonomous systems, context becomes part of data quality.

This is particularly important in finance because a technically correct action can still be an inappropriate action. The numbers may be right while the decision violates a policy, bypasses a control or occurs at the wrong point in the process.

That is why action-grade data must include more than the number itself. It needs the information necessary to determine whether an action is appropriate, authorized and defensible.

Enterprises have spent years building data lineage so they can understand where information originated and how it moved through their systems.

Agentic AI introduces another requirement: decision lineage.

If an AI system recommends or takes an action, organizations need to understand not only where the underlying data came from, but what information the system considered, what rules or policies applied, why the action was selected and what happened afterward.

An auditor or controller will not be satisfied simply knowing that an AI agent created a journal entry. They will need to understand why it was created, what evidence supported it, which controls were applied and whether the appropriate approval occurred.

The data, decision and resulting action increasingly become part of the same audit trail.

This is where governance stops being something applied around AI and becomes part of the architecture that enables AI.

There is a danger in taking this argument too far.

If organizations conclude that every piece of enterprise data must be perfect before they can deploy agentic AI, very little will ever get deployed. Enterprise environments are complex, with data spanning multiple systems and inevitable inconsistencies. Perfectly clean data is neither realistic nor necessary.​

The goal is knowing enough about the state and context of the data to determine what the system can safely do with it. That leads to a more practical model for autonomy.

When confidence is high, the data is reconciled, the policy is clear and the action falls within defined thresholds, an AI system may be allowed to proceed autonomously.

When confidence is lower, the transaction is unusual or the potential impact is material, the system should escalate the decision to a person.

This is where human-in-the-loop becomes much more useful than simply requiring a person to approve everything. Human judgment can be concentrated where it actually adds value.

For CTOs and technology leaders, the question should no longer simply be, “Is our data ready for AI?” That question is too broad.

A better question is, “What are we allowing AI to do with this data?”

The required standard should rise with the consequence of the action.​

That means understanding lineage, reconciliation status, business context, policies, permissions and control requirements at the point where AI acts.

The first generation of enterprise AI largely asked whether organizations had enough data for an LLM or GenAI solution to produce useful and accurate answers.

The next generation will ask a harder question: Do organizations have enough context around that data to safely allow agentic AI to act and execute specific business processes?

That distinction will become increasingly important as agentic systems move deeper into mission-critical enterprise workflows.

For years, we designed enterprise data so people could confidently read it. Now we need to design it to understand where autonomy is appropriate so agentic AI can responsibly act on it, what information is required to support agent activity and where human judgment still belongs.​

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?