TL;DR: AI agents are starting to change financial plans, and someone still has to own every decision they act on. Abacum CEO Julio Martínez argues that before agents touch the plan, finance teams should build an approval map: for each decision type, who approves it, what evidence they need, and when approval becomes action.
The audit standard didn't move. The work underneath it did.
"An audit is what it is." That's Shawn Panson, PwC's US assurance and transformation leader, talking to CFO Dive about what AI is changing inside his own firm's audits. PwC now uses AI to read supporting documents and flag missing information, and he says the time savings are real. But "what comes out the back end is still comfort on the financial statements."
I found that line reassuring, and I suspect a lot of finance leaders will read it the wrong way. One reading is that the audit hasn't changed, so finance doesn't have to change either. The other is that AI needs a whole new rulebook before it can touch anything that matters. Panson is saying something simpler. The standard didn't move, but the work underneath it did.
Why AI raises the bar on accountability in finance
That distinction matters because agents are starting to do more than draft. They move an assumption, reallocate a line, kick off a process. You can't hold an algorithm accountable in front of an auditor or a regulator, so someone with the right incentives still has to own every one of those decisions. The biggest misconception I hear about AI in finance is that it removes the need for human accountability. It does the opposite, and the CFOs who see that early are the ones who'll get to use agents hardest. We made a related argument about autonomy, accountability and provenance for AI agents in finance a few weeks ago.
Finance leaders already feel this. In the last three months, more than 20 of the finance teams we spoke with raised the same worry about AI forecasts: they can't take a number to the board or to investors unless they can trace it back to its source. And IBM's new CFO study, a survey of 1,500 finance leaders, found that only 6% allow AI to recommend or execute reallocations within defined guardrails. I don't read that as a technology gap. In most companies, nobody has written those guardrails down for the humans either.
Where approvals already break down: a new hire
Take a decision as ordinary as a new hire. Marketing and the sales development team agree they need someone. That conversation feels like a decision, but maybe the chief sales officer has to sign off, or someone else on the team needs to be involved. In most companies I know, headcount approvals already go wrong in both directions. An approved role kicks off a month late because nobody turned the approval into action. Or the leadership team looks at open roles, asks who approved that one, and nobody can say.
Now put an agent in the middle, reading the Slack thread and updating the plan. If approval rights were fuzzy for people, they'll be fuzzier for the agent, and it will act on the fuzz a lot faster than any of us would.
How to build an approval map before agents act on the plan
The most useful thing you can do before agents start acting on the plan is to build an approval map. For every kind of decision that moves the plan (headcount, spend, revenue targets, pipeline targets), write down three things:
Who can approve it. A named role, not "leadership".
What evidence they need to see first. The cost, the budget line it hits, the assumption it changes.
The moment approval turns into action. When the role opens, the spend is committed, or the target changes in the plan of record.
That's also exactly what an auditor will ask of an agent-driven process: who approved, what context the decision was based on, and when it became real. Once that's explicit, you can let the agents run hard on everything beneath it, because you'll know exactly where a human has to step in.
Continuous planning needs continuous approval
This gets more important as planning stops being an annual event. When the plan updates as the business changes, approval can't happen once a year in a budget meeting and then be forgotten. It has to happen continuously, at the speed the agents work, and it has to leave a record. That changes finance's job. The team stops being the one that produces the plan and becomes the one that designs how decisions get made, approved and traced inside it. I think that's a bigger and better job than the one we have now.
Here's a question to try on your own team: if an agent changed one assumption in your plan tomorrow, could you say who would have had to approve it, and could you prove they did?
Join the conversation on Wednesday
I'll be digging into exactly this with Stephen Hedlund from Rillet and Jeremy Budzian from PwC at Finance, AI Agents, and the Future of Assurance, a Rillet and Abacum webinar on Wednesday, October 14 at 12pm ET. Register here if you'd like to join us.
How Abacum helps
Abacum is an AI-native FP&A platform for finance teams that want to use AI hard and still be able to show who approved what. Abacum Intelligence works within the same permission structure as your team, with role-based access controls, version history, and approval workflows. Every AI-generated result is traceable: right-click any cell to see exactly which data sources, formulas, and assumptions drove it.
For the decisions that move the plan, budget owners contribute in one governed workspace while finance keeps control of definitions, timing, and approvals. Open roles, personnel costs, budgets, and approvals sit alongside the forecast. And Abacum keeps a detailed history of edit events across models, spaces, and dimension mappings, so "who changed this number, and when?" has an answer.
