Finance keeps asking AI the wrong questions

Walk into almost any finance org’s AI strategy meeting right now and you’ll hear some version of the same debate: which tasks should AI handle, and which should stay human? Reconciliations and probably variance commentary can be done by AI. Revenue recognition judgment calls need to stay with humans, obviously. Everyone’s sorting a task list into two columns.

It’s a reasonable instinct, but it’s also the wrong axis. It’s also why so many finance AI rollouts stall somewhere between the pilot and the second year.

Task type doesn’t predict where the risk is

The problem with sorting workloads this way is that task type doesn’t actually tell you where the danger is. Two examples:

Invoice coding looks like the safest possible candidate for full automation—the work is rules-based, high-volume, and tedious. But a miscoded invoice on a related-party transaction or a capitalized expense isn’t a rounding error; it’s a control failure with audit and tax consequences that can take months to unwind. Same task category, wildly different stakes depending on the specific instance.

Also consider an AI-drafted first pass at variance commentary for an internal management report. It looks like exactly the kind of ambiguous, judgment-heavy work everyone agrees needs a human. But if a human reviews and signs off on it before it goes anywhere, the actual cost of the AI getting it wrong is that someone has to edit a draft. Low stakes, dressed up as high stakes because it involves language and interpretation.

What actually matters is something most operating-model conversations skip past. Risk.

The variables that matter: can you undo it, and how far does it reach?

Instead of building your finance AI operating model around what kind of task it is, consider how reversible is the outcome and how far is the blast radius extends if it’s wrong.

Apply these two questions to any piece of work before you decide who or what owns it:

1. If this is wrong, how long does it take to find out, and how hard is it to unwind?

   A miscategorized expense gets caught at close and corrected in an hour. A number that’s already gone into an external filing or a board deck is a different animal entirely. The cost of being wrong isn’t the error; it’s the unwind.

2. If this is wrong, who else does it touch?

   An internal draft that stays inside one team’s review loop has a blast radius of zero until a human signs off on it. A number that automatically triggers a downstream system such as a payment run, a customer-facing statement, or a regulatory submission, has a blast radius the moment AI produces it, whether or not a human ever looks at it first.

Plot any piece of finance work on those two axes and you get a genuinely useful answer, instead of a task-category guess.

Reversible + contained

AI can act with a light-touch review. Speed matters more than caution here, and the cost of an occasional miss is genuinely small.

Reversible + wide blast radius

AI can draft or execute, but a human needs to be in the loop before it reaches the next system, not after, because the review has to happen while it’s still contained.

Hard to reverse, regardless of blast radius

This is where humans own the decision outright. AI’s role here isn’t to decide; it’s to compress the time it takes a human to get to a well-informed decision, surfacing the relevant history, flagging the anomaly, and drafting the options.

Notice what this framework helps you accomplish that a task list doesn’t. You’re able to explain why the same task—say, an AI agent adjusting a reserve estimate—might be perfectly fine to automate in one context (an internal working file, easily corrected) and completely inappropriate in another (feeding directly into a public disclosure). The task didn’t change. The reversibility and the blast radius did.

Why most governance conversations get this backwards

Most of the governance debate happening in finance right now is organized around gating categories of work. AI can touch reconciliations. AI cannot touch anything involving revenue recognition. It’s an understandable starting point, and not wrong exactly, but it’s harsh. It either over-restricts AI in low-stakes corners of a sensitive category, or under-restricts it in high-stakes corners of a category that got waved through as safe.

A reversibility-and-blast-radius lens is more work to set up because it requires mapping specific decision points instead of drawing a line around a job function. But it’s the difference between a governance policy that reads well in a slide and one that actually holds up when an agent does something unexpected during a live close.

This tracks with something finance leaders have been telling researchers all year, independent of any one framework: the conversation has moved from whether AI belongs in finance to specific, operational questions about what an agent can decide on its own versus what needs a human, and how confidence level should route work toward or away from a person. That’s the same underlying instinct, just described in terms of task and authority rather than reversibility. 

What this looks like in practice

If you’re designing (or auditing) an AI operating model for a finance function, these three moves will help:

1. Stop gating by task category and start gating by decision point. Pick a workflow like month-end close, AP, or revenue recognition and, instead of asking can AI touch this process, break it into its individual decision points and ask the two questions above at each one. You’ll usually find the same process has three or four genuinely different risk profiles hiding inside it.
2. Put the review checkpoint where the blast radius starts, not where the task ends. If an AI-drafted output automatically triggers a downstream system, the human checkpoint must occur before that trigger, not as a post hoc audit. A lot of “AI went rogue” stories in finance are really “the review happened after the blast radius had already expanded.”
3. Build the escalation logic around confidence and consequence together, not confidence alone. A common failure mode is routing to a human purely based on the AI’s confidence score. But a low-consequence decision with low confidence can often just proceed and get corrected later, while a high-consequence decision deserves human review even when the AI is quite confident, because “confident” and “right” aren’t the same thing when the cost of being wrong is high.

The question you should be asking

Which tasks should AI do? It’s a question that feels productive but mostly isn’t. Instead, it produces long debates and short-lived policies, because the categories keep needing exceptions. The more durable question is narrower and less comfortable. 

For this specific decision, if it’s wrong, how long until we know, and how far do the consequences travel before we find out?

Answer that consistently, and the AI-versus-human split stops being a debate about trust in the technology and becomes a design decision about risk, made explicitly instead of by default.