AI for Controllers:
Owning the Close
With Agents
The close as a review queue rather than a production line, the four controls you own, and an honest account of the first parallel period.
- The controller constraint moves from throughput to review capacity. That is a better constraint and a different job, and it is the largest role change in the function.
- The close tasks do not disappear, the ownership moves. Production becomes review and coordination becomes monitoring, so the close shortens in elapsed days rather than in checklist lines.
- Reviewing prepared output is its own skill: review the basis not the arithmetic, sample the clean population deliberately, and watch the pattern of corrections rather than individual edits.
- Four controls belong to the controller: thresholds, approval routing, service account scope and evidence retention. All four should be reviewed quarterly.
- Expect the first period to cost more than it saves. Parallel running is what buys trust, and the difference log it produces is the configuration backlog for period two.
Most writing about AI in finance is aimed at the person approving the budget. This is aimed at the person who will run it, because the controller role changes more than any other and the change is concrete: you stop managing how much the team can produce and start managing a queue of prepared work that needs judgment.
This guide covers the close stage by stage under that model, how to review prepared output without checking everything, the four controls that belong to you rather than to the CFO, what it does to the team underneath you, and an honest account of the first parallel period. The examples are drawn from ChatFin deployments.
The controller role, before and after
Most descriptions of AI in finance are written for the person approving the budget. This one is written for the person who will run it, because the controller role changes more than any other in the function and the change is specific.
Today a controller distributes work, chases status, reviews what comes back, and absorbs whatever did not get done. The constraint is throughput: how much the team can produce in the days available. With agents preparing the work, the constraint moves to review capacity and judgment, which is a better constraint to have and a different job to do.
ChatFin presents work as a ranked review queue rather than a set of completed reports, because the ranking is what makes a small team able to review a large population.
Running the close as a review queue
Take the close stage by stage. The tasks do not disappear, the ownership moves.
| Close stage | Who does it now | Who does it with agents |
|---|---|---|
| Subledger to general ledger | Analyst matches and investigates | Agent matches and explains each break. Controller judges the breaks |
| Bank reconciliation | Analyst clears and chases | Agent clears and proposes matches for the residue. Controller confirms |
| Accruals | Analyst builds from open purchase orders in a spreadsheet | Agent drafts with the supporting line attached. Controller approves and sets thresholds |
| Recurring journals | Analyst prepares, controller reviews | Agent prepares, controller reviews. The review is the same, the preparation is not |
| Flux and variance | Analyst assembles, controller interprets | Agent decomposes drivers with transactions attached. Controller interprets |
| Status and chasing | Controller chases owners | Agent monitors and surfaces what is blocking. Controller decides what to escalate |
ChatFin covers all six stages as prebuilt workflows, which is why the close compresses in elapsed days rather than in the number of checklist lines.
How to review agent output well
This is the skill that has no established training, so it is worth stating explicitly. Reviewing prepared output is not the same as checking a junior's work.
Review the basis, not the arithmetic. The calculation is almost never wrong. What deserves attention is whether the right rule was applied to the right population, which is why every prepared item should carry the rule and the source line it came from.
Sample the clean population deliberately. Reviewing only the exceptions creates a blind spot, because an item wrongly classified as clean never reaches you. A small random sample of accepted items each period is the control that catches this.
Watch the pattern of corrections. Individual corrections matter less than their shape. Three corrections of the same type in a period is a policy problem, not three mistakes, and it should be fixed in configuration.
Keep an eye on what stopped appearing. If a category of exception that used to surface has gone quiet, find out whether the underlying issue was fixed or whether a threshold is now hiding it.
ChatFin surfaces correction patterns by workflow so a controller sees the shape rather than the individual edits, and supports sampling of the accepted population as a standing control.
The controls a controller owns
The CFO sets the principles. The controller owns the configuration that makes them real, and there are four levers.
In ChatFin all four are configuration a controller changes directly, without a development cycle, which is what makes a quarterly policy review practical.
What this does to the team underneath you
Being straight about this matters, because the team will ask and evasion costs more than candour.
The production half of junior accounting work shrinks. Keying, matching, chasing and assembling are what agents are good at. What grows is exception judgment, policy work and explanation, and those are harder to hire for than the production skills they replace.
The practical consequence for a controller is that training changes shape. A new analyst who never manually built a reconciliation will not develop the instinct for when one looks wrong, and that instinct is exactly what the reviewing job needs. The teams handling this well have new analysts work a period manually before they work it as a reviewer, deliberately, as training rather than as production.
This is why ChatFin keeps the source line and the rule visible on every prepared item. A reviewer who can see how the answer was reached can learn from it, and one who only sees the answer cannot.
The first period, realistically
Expect the first period to cost more effort than it saves. That is not a failure, it is parallel running, and skipping it is the most common cause of a deployment that never earns trust.
Run both, compare deliberately
Do the work as usual and compare against the prepared output. Log every difference with its cause rather than just correcting it.
Sort differences into three buckets
Agent wrong, policy wrong, or the manual process was wrong. The third bucket is usually larger than anyone expects and it is the most valuable output of the period.
Fix policy, not individual items
Correcting the same class of item repeatedly means the configuration is wrong. Change the threshold or the rule and the class stops recurring.
Cut over one task at a time
Move tasks off parallel running individually as each earns confidence, rather than switching the whole close at once.
ChatFin deployments plan for one full parallel period and treat the difference log from it as the configuration backlog for the second.
Deciding what to escalate, and when
A review queue only works if there is a clear rule for what leaves it and goes to you or to the CFO. Without one, everything either gets escalated or nothing does.
Escalate on materiality and on novelty, not on volume. Fifty small variances of a known type are a configuration conversation. One variance of a type nobody has seen before is worth ten minutes of your attention even if the value is trivial, because it is usually the first instance of something structural.
Escalate anything that touches revenue recognition or a payment release immediately. These two carry consequences that are hard to unwind, and the cost of an unnecessary escalation is far lower than the cost of a late one.
Escalate a pattern of corrections, not the corrections. If the same class of item has been adjusted three periods running, that belongs in front of the CFO as a policy question rather than being quietly fixed a fourth time.
Do not escalate what the trail already answers. If the prepared item shows the rule, the source line and the approver, an anomaly can usually be resolved by reading it rather than by convening a meeting about it.
ChatFin flags items that sit far from the normal pattern separately from items that simply breach a threshold, which is the distinction that makes a novelty based escalation rule workable.
A reviewer who can see how the answer was reached learns from it. One who only sees the answer cannot, and that is the whole argument for showing the basis.
Ashok ManthenaWhere to go next
This page covers the role. The operating detail sits one level down, by the process you are fixing or the system you run it on.
For the executive framing, see AI for CFOs. For a single workflow end to end, see reconciliation automation.
Questions controllers ask first
What does a controller actually do differently with AI agents?
The controller stops distributing production work and starts managing a review queue. Instead of assigning reconciliations and chasing status, the day begins with prepared reconciliations, drafted accruals and a list of exceptions ranked by materiality. The skill that matters shifts from knowing how to produce an entry quickly to knowing which of two hundred prepared items deserves a second look.
How does a controller review agent output without checking everything?
By reviewing the exceptions and sampling the rest, which is how audit has always worked. A prepared item arrives with its source line, the rule applied and the reasoning, so a review is a check of the basis rather than a rebuild of the work. In ChatFin the items are ranked by materiality and by how far they sit from the normal pattern, so the population that needs a person is the first thing on the screen.
Does using AI agents change the close checklist?
The tasks stay, the ownership shifts. Tasks that were production, such as building schedules, matching subledgers and preparing recurring journals, become review tasks. Tasks that were coordination, such as chasing entities for status, become monitoring. The checklist gets shorter in elapsed days rather than in number of lines.
A ranked review queue, not a pile of finished reports.
ChatFin prepares the close work across reconciliation, accruals, journals, intercompany and flux, then presents it as a queue ranked by materiality and by distance from the normal pattern. Every item carries the rule applied and the source line it came from, so a review checks the basis rather than rebuilding the work.
Thresholds, approval routing, service account scope and evidence retention are configuration a controller changes directly. Nothing posts without a named approver. See reconciliation automation for one workflow end to end, or AI for CFOs for the case your executive team will be working from.
Judgment stays with you. The production does not.
Book a demoDiagram: the six stage monthly close cycle, original ChatFin artwork.
Diagram: rule based automation compared with agents across the close workflow, original ChatFin artwork.
All images are hosted on ChatFin infrastructure or served from the Pexels content network. No third party product screenshots, vendor logos or licensed media are used on this page.
Bring the reconciliation you dread most.
We will run it on your data, show the prepared output with its basis attached, and walk the review queue exactly as your team would work it during a close.