The AI Controller: Job Description for a Role That Didn't Exist Two Years Ago

No recruiter has posted this job. Search any job board for "AI Controller" and you'll get a handful of listings that bolted "AI" onto a traditional controller req and changed nothing else, or nothing at all. And yet talk to a controller who has spent the last year putting agents into their close, and they will describe a job that has almost nothing to do with the one on their business card.

That gap is worth naming, because a role the market feels but hasn't written down yet is exactly the kind of role people are quietly already doing without a title for it. Here's the job description. Not a forecast of some future org chart. A description of what a controller running an AI-assisted close is actually responsible for right now, written the way a real req would be written, followed by the honest version underneath each line.

---

The job posting nobody's written yet

Title: Controller (AI-assisted close) Reports to: CFO Location: Wherever the close used to happen. Nothing about where you sit changed.

What you'll actually do:

• Review agent-prepared reconciliations, journal entries, and flux commentary before they post, and decide what clears versus what gets kicked back
• Set and adjust the autonomy level for each workflow: which accounts run with a human checking every line, which run with spot review, which have earned the right to run with lighter oversight
• Own the exceptions queue as the primary daily surface, not the trial balance
• Explain a populated number to an auditor, a board member, or your own CFO by tracing it to source, not by re-deriving it from scratch
• Decide what a recurring judgment call becomes policy, and what stays a one-off, so a single resolved exception doesn't quietly become the standing rule
• Keep the sign-off chain intact: nothing posts to the ledger without a name attached to the decision to let it post

Requirements:

• Everything a controller's job already required: technical accounting judgment, the ability to explain a number under pressure, and the willingness to put your name on it
• A new skill nobody trained you on: reviewing a completed artifact for the specific way it fails, which is a different exercise than reviewing a junior's draft for the specific way they fail
• Comfort setting a dial rather than flipping a switch. Nothing here is all-or-nothing anymore

Nobody has to grade themselves against this literally. The point isn't the format. It's that if you read that list and thought "that's just my job now," you already know why this piece exists.

---

Why this job didn't exist two years ago

Two years ago, "controller" meant the person who built the close, checked it, and signed it, largely in that order and largely alone or with a small prep team under them. The job was structured around producing the artifact. Review was a step at the end, not the whole job.

What actually changed isn't that the title got new letters in front of it. It's that the preparation layer of the close, the reconciliations, the mechanical journal entries, the first draft of the flux commentary, started getting produced by something other than a person, and the controller's actual hours moved almost entirely to the review side of the line. A recent look at the labor-market data behind this shift found CFOs surveyed globally expecting a real reshaping of the org chart underneath them: a meaningful share expect fewer purely junior, preparation-only seats, with growth concentrated in midlevel and senior roles built around review and judgment instead. Survey authors have described the resulting shape as a pyramid flattening into a diamond.

A controller sitting at the point where that diamond gets wide is not doing the old job with a new tool bolted on. They're doing a materially different job, and the job description above is what it actually looks like once you write down what fills the hours instead of assuming the title tells you.

---

What changed, item by item

The unit of work is the exception, not the account. The old close moved account by account: open the reconciliation, build it, tie it out, move to the next one. The new close moves exception by exception: open the queue, and everything in it is already something that didn't clear cleanly on its own. The accounts that matched are not where the day goes. This is a genuinely different rhythm to the job, and it took most controllers we've talked to a few close cycles to stop reflexively re-checking work that had already cleared.

Review is a distinct skill from preparation, and it's not the one your training built. Reviewing a junior's draft means checking whether they understood the assignment. Reviewing a completed reconciliation that an agent built means checking for the specific ways that kind of process fails: a matching rule that was too generous, a source document that was misread, an edge case nobody anticipated when the rule was written. It's a real skill, most controllers are building it live on the job right now, and almost nobody has formal training in it yet.

Autonomy is a setting you manage, not a decision you make once. The old job had one lever: automate this step, or don't. The new job has a dial per workflow, and the dial moves as confidence builds and moves back the moment it shouldn't. A stable, well-precedented reconciliation earns lighter review over several clean cycles. A new entity, a new vendor pattern, a first-time position, stays fully gated regardless of how well everything else is running. Managing that dial, workflow by workflow, is a real and recurring part of the job that didn't exist when everything was either manual or not.

Institutional memory becomes something you actively manage instead of something that just accumulates in one person's head. The old version of this was informal: the senior person who'd been there ten years just knew how a specific counterparty's odd statement format got handled, or how last quarter's one-off got resolved. Now that knowledge gets captured deliberately, so the same resolution gets proposed the same way the next time, and the controller's job includes deciding when a one-off judgment call should become a standing rule and when it shouldn't. Letting every resolution quietly become policy is its own failure mode, and catching that is now part of the job.

Explaining a number moved from "walk through my calculation" to "trace this to its source." When a controller built the workpaper by hand, they could explain a number because they remembered building it. When an agent built it, the controller has to be able to trace a populated cell back to where it came from and be confident in what they're seeing, on demand, in front of an auditor or a board member who doesn't care how the number got there, only whether it's right and whether someone can prove it.

---

What didn't change, and why that part matters more than the new stuff

It would be easy to write this piece as "everything about the controller job is different now." That's not true, and the parts that stayed the same are the parts that make the new parts trustworthy.

Sign-off is still a person's name on a decision. Nothing in any of this moves legal responsibility for the number. A controller who signs off on a reconciliation is still the one who signed off on it, whoever or whatever built the first draft.

Judgment calls are still judgment calls. Deciding whether a variance is explainable, whether a position is defensible, whether an exception is a one-off or a pattern worth escalating: none of that got automated, and there's no credible version of this job where it does.

The relationships didn't move. Explaining a number to a nervous board member, negotiating with an auditor about scope, telling a business partner their number is wrong: still entirely a human conversation, still the highest-leverage hour in the job.

That combination, more review hours and fewer preparation hours, sitting on top of judgment and relationships that never moved, is the actual shape of the role. Not less accounting. Different accounting.

---

You might already be the AI Controller if

A quick, honest self-check, the kind worth screenshotting if any of it lands:

• Your daily starting point is an exceptions list, not a blank reconciliation template
• You've caught yourself re-checking work that already passed review, out of habit rather than necessity
• You've had a conversation about whether something should run with lighter review this cycle
• You've explained a number to someone by tracing it to source rather than by remembering how you built it
• You've had to decide whether a one-off fix should become the standing rule
• Your title still says "Controller," full stop, with nothing in front of it

If four or more of those are true, the job posting nobody's written is describing the job you're already doing.

---

<div class="cta-block">

Curious what the exceptions queue actually looks like on your own close? Book a pilot and we'll scope it against a workflow you already run.

Want to see the review layer first, before anything else? Start free.

</div>

---

Related reading

• Will AI Replace Accountants? What the Data Actually Says
• How to Become an AI CFO: A 12-Month Roadmap
• The AI CFO Tech Stack: What Finance Leaders Are Actually Running in 2026

---

This piece describes a role as it's emerging across finance teams generally, not a job description for any specific employer, including ours.