Reviewing Workiva AI output before it reaches a report

Workiva AI can now write variance commentary, explain why two figures differ, and run parts of a control test. An early access tie-out feature also checks 10-⁠Ks and 10-⁠Qs for internal consistency. Each of those outputs can end up in a filing, a board pack, or an audit committee report. The review question is the same one you ask of a new preparer’s work: what did they rely on, what did they change, and who checked it?

Workiva’s FAQ says Workiva AI is not always accurate and may produce content that is incorrect, including on recent laws, regulations, and accounting guidance. Workiva’s August 2026 survey of executives found that 84% were at least somewhat confident in AI output without human review, while 26% said internal audits had found AI errors that reached external audiences or board members. Only 11% said their data quality was sufficient for AI use.

Close the gap between that confidence and those errors by deciding, for each task, what the reviewer checks and what record the review leaves.

Figures and narrative fail in different ways, so check them differently.

Figures should come from linked, reconciled sources. If a number in a report was typed by AI rather than linked to a source cell, treat that as an exception to fix. Run your normal tie-outs, link checks, and reconciliation to the trial balance. Where your workspace has the early access Tie-out Assistant, a tie-out run helps here: it groups related figures and flags possible mismatches for a reviewer to verify, flag, or annotate. It does not replace the reconciliation that proves a figure is right.

Narrative needs a reader who knows the business. Check that commentary agrees with the linked figures it describes, that explanations match what actually happened in the period, and that nothing reads as an accounting conclusion your team has not reached. If a sentence cites a source, open the source. A citation tells you where the AI looked; you still have to check what it took from that source.

For SOX teams, the same split applies to AI-assisted testing. A testing agent can check a population and sample, propose test attributes, and suggest results with annotated evidence. The control tester and reviewer still own the conclusion, and the control testing worked example shows the review steps that stay with people.

Workiva’s AI built for trust article separates what is available now from what is coming. As of October 5, 2026, the available items include:

• Citations. AI responses show inline citation markers and a list of the sources they drew on.
• Permissions. Workiva AI only retrieves content the user can already open, down to the section level.
• Confirmation. Actions that change your content require confirmation, and a workflow runs with the permissions of the person who starts it.
• Activity log. AI interactions are recorded with the user, the interaction, and the model provider. Workspace admins can view these records.
• Workflow steps. Within a session, you can select Show thinking to review the reasoning, tools, and actions a workflow took. Capture it if you need it later.

Workiva lists several items as coming, with no committed dates: AI authorship in document history, citations to the exact cell or span of text, exportable citation sets, evidence packages for auditors, and read-only auditor access. Until those exist, your own process has to show where AI was used and who approved the result.

Starting a new Tie-out Assistant run clears its earlier notes and resolved items. If those notes are part of your review record, capture them before you run it again.

Decide these points for each AI-assisted task, write them down, and share them with the people who review the process:

1. Which Workiva AI feature the task uses, who may run it, and in which workspace.
2. Which figures must come from linked sources, and which checks prove them.
3. How AI-drafted or AI-edited content is identified in the file until it is approved.
4. What the reviewer checks, and where the approval is recorded.
5. Which records you keep, such as activity log entries, tie-out notes, citations, and approvals, and where they are stored.
6. What happens when the output is wrong: who corrects it, and whether the prompt, workflow, or source needs to change.

Talk to your external auditor before AI-assisted work becomes part of a process they rely on. There is no PCAOB standard specific to AI. A 2024 PCAOB staff Spotlight reported that preparers using generative AI in financial reporting stress human supervision and review of its output, and some emphasized controls over the reliability and accuracy of the data used to train the models. COSO’s 2026 publication, Achieving Effective Internal Control Over Generative AI, applies the Internal Control–Integrated Framework to generative AI and is a useful reference when you document these controls.

If your controls already live in Workiva, add the AI-assisted steps to the same risk and control matrix rather than keeping a separate list. The Controls Management guide explains how risks, controls, testing, and issues connect.

Pick one recurring task, such as commentary for a monthly report or the test of one control, and run it beside the prior cycle you already trust. Compare the AI-assisted result with the known answer, record what the reviewer changed, and decide whether the task is ready for the next cycle. The Workiva AI readiness checklist covers the settings, access, and source data to settle first.