The Future of Finance Isn't AI Without Humans. It's AI Doing the Work, Humans Owning the Decision.

For the past two years, most enterprise conversations about AI in finance have circled a capability question: can an AI system draft a variance explanation, benchmark a disclosure against peers, or check a sustainability report for compliance gaps as well as a trained analyst can. In July 2026, Workiva — one of the largest platforms enterprises use to prepare SEC filings, financial close packages, and ESG disclosures — answered that question by shipping three named, specialized AI agents directly into the reporting workflow: a Tie-Out Agent that checks figures for consistency across a report and generates audit-ready explanations for variances, a Benchmarking Agent that analyzes peer companies' 10-K and 10-Q filings to surface disclosure gaps, and a Sustainability Disclosure Agent that drafts and validates ESG disclosures against ESRS and ISSB standards.

The capability question is, for practical purposes, settled. What Workiva's own product design makes clear is that a different question now matters more: not whether the agent can do the work, but who is named as having reviewed and approved it before it goes out the door. Deepak Bharadwaj, Workiva's EVP and Chief Product Officer, put the underlying logic bluntly in a press interview around the launch: an organization cannot send an AI agent to prison. Wherever a reporting task carries that kind of consequence — a signature, an attestation, a regulatory filing — a specific human has to be the one accountable for it, whatever assistance an agent provided along the way.

This is precisely the territory Crest.Digital's Agentic Risk & Continuous Assurance practice was built to address, and it extends directly from two questions the practice has already covered: who is accountable for an AI agent's decision, and why AI agents need audit trails. What Workiva's launch adds is a live, market-tested answer from inside the exact process — financial reporting — where the accountability question has the least room for ambiguity.

AI Agents Enter the Close

Workiva's July 29, 2026 launch is worth reading closely because of how deliberately it's scoped. Rather than a single general-purpose "reporting assistant," the company shipped three narrowly defined agents, each mapped to a specific, well-understood task in the reporting lifecycle. The Tie-Out Agent performs comprehensive cross-checks across a financial report, flags discrepancies between figures that should match, and generates an audit-ready explanation for each variance it finds — work that traditionally consumes significant analyst hours during a tight close window. The Benchmarking Agent analyzes peer companies' SEC filings to identify disclosure gaps and build custom peer groups, turning a research task that used to take days into something closer to hours. The Sustainability Disclosure Agent drafts and validates ESG disclosures against the ESRS and ISSB standards and produces compliance scorecards, addressing a reporting area where requirements are still actively evolving.

Underpinning all three is what Workiva calls Workiva Knowledge, a persistent intelligence layer grounded in an organization's own data, instructions, and reporting content that is meant to compound in usefulness with each successive reporting cycle, plus a Model Context Protocol gateway that lets customers bring their preferred AI models while Workiva's own governance and data-lineage controls stay in place regardless of which model is doing the underlying work. That governance-first framing is not incidental — it's the point. As Bharadwaj described it, finance, sustainability, risk, and compliance teams are moving from execution to orchestration as agents take on more of the work itself, and the platform is explicitly designed to keep every agent action traceable, defensible, and tied to a specific individual, not just technically capable.

For internal audit and finance-controls teams, the significance of this launch isn't really about Workiva as a vendor. It's evidence that the market has already converged on an operating model for agentic AI in regulated financial processes — agents handle the volume and the pattern-matching, humans hold the sign-off — well ahead of any AI-specific regulation mandating it. That convergence is exactly what internal audit, risk, and compliance functions should be testing against, whichever reporting platform or AI tooling their own organization actually runs.

Why "Can It Do the Work" Was Never the Real Question

Bharadwaj's framing of Workiva's design philosophy is worth quoting in full, because it captures the accountability logic more precisely than most formal governance documents manage to: "It's really about the risk profile of our buyers and users. Because of who we serve in financial reporting and GRC and sustainability, what they want is the work that they do can be trusted because ultimately in many of these cases somebody is signing their name to it." The design rule that follows from that is simple to state and hard to game: repeatable, rules-based tasks can run with minimal supervision, but anything that requires business judgment, or where someone needs to look at it and approve and sign off, has to route to a human.

That rule matters more than the agents' individual capabilities, because it answers the accountability question Crest has argued elsewhere is the one that actually determines whether agentic AI is safe to deploy in a regulated process: not "is the agent accurate enough" but "who is the named, accountable party when something goes wrong." An agent's accuracy can improve with every release. Its capacity to be held accountable cannot, because accountability isn't a technical property — it's a legal and professional one, attached to a person who can be examined, sanctioned, or held to a professional standard, none of which apply to software regardless of how capable it becomes.

This is also why "human-in-the-loop" as a phrase has started to feel insufficient for regulated financial reporting. A human glancing at an AI-generated output before it's filed is a weak form of oversight if that review isn't evidenced, isn't tied to a specific named reviewer, and doesn't happen contemporaneously with the judgment being made. The stronger version — the one Workiva's task-boundary rule and COSO's new guidance (covered next) both point toward — is human sign-off as a designed, tested, and evidenced control, not an assumed byproduct of having a person somewhere in the workflow.

What COSO's New Guidance Requires

Workiva's product design isn't happening in a regulatory vacuum. In February 2026, COSO — the standard-setting body behind the Internal Control-Integrated Framework that underpins SOX internal control programmes globally — published Achieving Effective Internal Control Over Generative AI, extending its classic control framework directly to AI-assisted decision-making. The guidance's central move is to shift control expectations upstream: rather than only validating the final recorded outcome — the filed number, the approved entry — controls now have to reach the point where an AI system actually influenced the judgment behind it. As the guidance frames it, control can no longer focus only on recorded outcomes; it has to extend to how the decision was formed in the first place.

Three requirements from COSO's guidance map directly onto how organizations should be testing AI-assisted financial reporting today. First, contemporaneous evidence: organizations have to document AI-assisted judgments during the reporting period when the decision actually occurs, not reconstruct that documentation after the fact once an auditor asks for it. Second, full traceability of the decision path — the AI system's input, its output, the human review that followed, any exceptions raised, and any changes made — captured as a connected record rather than scattered across email threads and personal notes. Third, board-level accountability: boards are expected to define acceptable parameters and tolerance levels for AI use in financial processes and to receive reporting that enables timely escalation as those boundaries are approached, not discovered after the fact in an annual review.

This guidance lands alongside, not instead of, the audit-evidence standards already reshaping how external auditors treat AI-assisted work product. The PCAOB's amended AS 1105 audit-evidence standard, effective for audits of fiscal years ending on or after December 15, 2026, specifically sharpens how externally sourced and electronically generated evidence has to be evaluated — language that reaches directly into AI-generated variance explanations and benchmarking analysis of the kind Workiva's agents now produce. The IIA's Global Internal Audit Standards, effective since January 2025, already require internal audit to evaluate the adequacy of governance and risk-management processes, which squarely covers whether an organization's AI-assisted reporting controls are actually operating as designed rather than merely documented. None of this is a jurisdiction-specific requirement: a US issuer filing 10-Ks, a European group reporting under ESRS, and an India-based GCC supporting a global parent's close process all sit inside the same structural expectation — that AI-assisted judgment in financial reporting has to be evidenced, traceable, and reviewed by a named person, wherever the entity is domiciled.

An 8-Point Framework for Testing AI-Assisted Reporting Controls

None of the eight points below require an organization to have already adopted agentic AI in its reporting process. They're a testing sequence for whatever stage an organization is at — from a first pilot to the kind of production deployment Workiva's launch represents — that establishes, with evidence rather than assurance from a vendor or an IT team, exactly where human sign-off is real and where it's assumed.

Point one and point two do most of the structural work. An organization that has correctly classified every task and named a real reviewer for each judgment-requiring one has already closed most of the gap between "we use AI responsibly" as a stated policy and "we can prove it" as a tested, evidenced control.

Closing the Gap: A Six-Step Delivery Playbook

Crest.Digital delivers this as a configurable "Risk Automation Pod" rather than a bespoke software build — a reusable stack of ERP and reporting-platform connectors, a task-classification and risk-scoring engine, a workflow and approval-routing layer, and an evidence repository, customized around which reporting tasks, systems, and control points a given finance or audit function needs to bring under continuous testing first.

Starting with the reporting tasks that carry the highest materiality and the most direct line to a filed disclosure — the close, the SEC filing, the sustainability report — gives the validate step the clearest comparison against manual review, and builds the evidence base internal audit and external auditors will look for first.

Where the Agent Stops and Judgment Stays Human

Continuous testing of AI-assisted financial reporting is strong at three things: confirming every agent-touched task was correctly classified as rules-based rather than judgment-requiring; verifying that a named human actually reviewed and approved each judgment-requiring output, contemporaneously and with evidence; and maintaining the audit-ready trail that turns "someone reviewed this" from an assumption into something demonstrable to an external auditor, a regulator, or an audit committee. It is not a substitute for the judgment itself — deciding whether a variance explanation is genuinely adequate, whether a disclosure gap identified by a benchmarking agent actually matters for this filing, or whether a novel exception represents an acceptable risk. Those calls stay with the named reviewer, the controller, and the audit committee; continuous testing's job is to make sure they have complete, current, contemporaneous evidence when they make them.

This practice connects directly to two adjacent pieces of Crest's Agentic Risk & Continuous Assurance work. The task-boundary discipline described here — agents for rules-based work, named humans for judgment — is the practical, market-tested version of the accountability question already covered in who is accountable for an AI agent's decision. And once continuous testing surfaces a control gap or an unevidenced sign-off, it feeds directly into the same evidence and remediation lifecycle covered in why AI agents need audit trails and Crest's work on continuous evidence versus continuous monitoring. Detection, testing, and evidence are stages of one practice, not three separate initiatives competing for the same budget.

Bharadwaj's line is worth returning to as the closing test for any organization deploying AI agents in financial reporting today: you cannot send an agent to prison. Whatever an agent drafts, checks, or flags, the reporting process only stays defensible if a specific, named human can be shown to have owned the decision — and internal audit's job, starting now, is to test whether that's actually true, task by task, rather than take the org chart's word for it.

Frequently Asked Questions