Last week, KPMG put out a press release announcing an AI agent — built on Google’s Gemini, embedded inside Workday — that runs the month-end close. You hand it a checklist. It executes the sequence, flags anomalies, analyzes trial balances, and routes exceptions to humans for sign-off.
It’s worth being honest about the press release itself: it’s light on specifics. No named clients, no before-and-after metrics, no demo. It was timed to a conference and reads like most Big Four announcements. But here’s the thing — the underlying story is real, and it’s worth understanding regardless of what KPMG is or isn’t doing.
Because KPMG isn’t alone. HPE’s CFO is in production with a similar agent built with Deloitte. Several startups are purpose-building agentic platforms for the close. Some of my own consulting clients are actively standing up close agents right now. And according to Deloitte’s Q1 2026 CFO Signals survey, more than half of finance chiefs say AI agent integration is their top technology priority for the year.
The month-end close is the most active near-term target for agentic AI in accounting. Here’s what you actually need to know.
What the Close Actually Is — and Why It’s Been Hard to Automate
The month-end close is the process of reconciling all financial activity for a period — journal entries, account reconciliations, intercompany eliminations, variance analysis, trial balance review — so the company can produce accurate financial statements. For a mid-size company, it typically takes five to seven business days. For a large multinational, it can stretch to ten or more.
Most of the work is procedural. The same steps every month. Bank recs, amortization schedules, standard journal entries. This part has been getting automated for decades — ERP systems in the 90s, close management tools like BlackLine and FloQast in the 2010s. Those tools reduced manual data entry and gave teams better visibility. But they didn’t remove the human coordinator. Someone still had to click through the checklist, move files, and decide what to do next.
The hard part has always been the exceptions. The unusual journal entry from a foreign subsidiary. The balance that doesn’t tie. The variance that might be a timing difference or might be a control failure. That edge — where something looks wrong and someone needs to decide what it means — has kept humans essential.
What “Agentic” Actually Changes
Earlier AI tools in accounting were reactive. They answered questions when asked, flagged items in a dashboard, generated summaries. Then they waited for a human to decide what to do next.
Agentic AI initiates actions. It reads your checklist and knows what comes next. It pulls data from the ERP, runs the reconciliation, analyzes the variance, writes up findings, and routes the exception to the right person — all without a human coordinating each step. Think of it like autopilot in aviation. The pilot is still essential, still in the cockpit, still making every call that matters. But autopilot handles the continuous small adjustments that don’t require judgment, so the pilot’s attention is free for the moments that do.
One architectural point that matters a lot here: you don’t want the AI model rewriting the reconciliation logic every time it runs. Models change, outputs vary, and that’s a reliability problem. What you want is a deterministic script — a tool that takes two datasets, does the transformation, and produces a consistent result — and then give the agent the ability to call that tool. The agent orchestrates; the script executes. That distinction is what makes these deployments auditable.
The Hackett Group’s 2026 benchmark found that companies with full agentic workflow deployment are compressing their close cycle from an average of 6.2 days to 1.8 days. That’s not incremental improvement. That’s a different operating model.
Human in the Loop Is Not Optional
Every major close AI deployment today preserves human review and sign-off at key control points. This isn’t a design choice companies are making out of caution — it reflects regulatory expectation, audit standards, and basic risk management. SOX, SEC reporting, and auditor standards all presuppose human accountability for financial statements. You can’t delegate that to an agent.
And AI does fail. According to Infosys, 95% of companies experienced at least one AI incident in 2025, and only 2% reported having adequate guardrails. If you’re deploying one of these systems, that needs to be front of mind. The agent prepares. The controller approves.
What This Means for CPAs
For controllers and accounting analysts, the job shifts from executing the close to governing it. Instead of running the checklist, you’re designing the process so the agent can run it. You’re reviewing exception queues with judgment. You’re owning sign-off decisions. The CPA who understands why a reconciliation is structured a certain way — not just how to run it — becomes more valuable, not less.
For auditors and advisory CPAs, AI-assisted closes create new questions that don’t have settled answers yet. How do you test controls when an agent is executing them? What does your ITGC landscape look like when a third-party LLM is a participant in the financial reporting process? Who’s responsible when the agent makes an error that flows into filed financials? CPAs who can advise on AI governance in the close are going to be in high demand.
Key Takeaways
- The month-end close is the most active near-term target for agentic AI in accounting — real deployments are live, not theoretical.
- Agentic AI differs from prior close automation by initiating multi-step actions on its own — it can run a checklist end-to-end without human coordination at each step.
- The key architectural choice: give the agent deterministic tools to call, not raw prompts to rewrite the process each time.
- Human in the loop is a regulatory and risk management requirement, not a preference. The agent prepares; the controller approves.
- For controllers, the job shifts from executing the close to governing it. For auditors, new ITGC and control-testing questions are emerging without settled answers.
Want the CPE credit? Take the full lesson on EverydayCPE and earn 0.2 CPE credits: [lesson link]
Leave a Reply