The shift to AI-enabled work

For generations the monthly close has been done by hand, late into the night. Artificial intelligence is starting to change that. It isn’t taking the job; it’s taking the keystrokes.

Scroll to begin

Every month, in accounting teams around the world, the same ritual repeats. Accounts are pulled and tied out. Variances are explained. Journals are booked, reconciliations cleared, sign-offs chased down the hall and across time zones. For most of accounting’s history there was only one way to do it: by hand, by people, often well past midnight.

That arrangement is changing, and the shift is subtler than the headlines suggest: a redistribution of which part of the work belongs to whom. The repetitive work shifts to the machine, freeing accountants to spend their day on the work only they can do: the analysis, the calls, the decisions.

Chapter one

The anatomy of a task

Look closely at a single close task and you can watch the change unfolding. Traditionally, one person owns it from the first keystroke to the final sign-off. Now the work is beginning to split between person and machine. Following one task through, step by step, shows what each side takes on — and where the accountant’s role gets harder to replace, not easier.

01 — Memory

It no longer starts from scratch

Done by hand, a close begins cold — the method living in someone’s head, or a folder of last month’s files. Numeric moves the starting line: it holds the tasks and requirements, but also a memory of how each was handled before, so the work can compound instead of resetting every month.

02 — The trigger

The work can start itself

Here the work breaks from tradition. A task doesn’t have to wait for a person to begin it: someone can kick one off by running a skill, or an agent can start it on a schedule, before anyone is at their desk.

03 — Human in the loop

It keeps a human in the loop

Routine work runs straight through — the AI prepares it and no one waits. But the moments that call for judgment are handed back to a person on purpose: the material variance, the one-off accrual, the final sign-off. The human stays in the loop exactly where their experience matters most — and stays out of the busywork everywhere else.

04 — Accountability

The accountability stays with the team

The labor is moving to the machine; the responsibility isn’t. A named person still owns each result and answers for it to the auditor, the CFO, the board. AI changes who does the work, not who signs their name to it.

05 — The output

The output shows its work

What comes out isn’t just a final number — it carries the whole trail. Every step is laid out: the supporting documents it pulled, the calculation under each figure. A reviewer isn’t handed a black box to take on faith; they can follow exactly how the work was built, line by line.

Numeric Source of truth Manual trigger You run a skill Agent trigger An agent fires Task opens In Numeric Requirements Tasks Requirements Async Checkpoint Claude prepares Runs the skill Claude starts Initial prep run Preparer input Claude waits For preparer Claude continues Resumes work Preparer reviews Submits + records Workpapers + memory Workpapers memory · Treatments · Preferences · Changes · History Reviewer reviews AI work + edit trail Sign-offs Sign-offs Approved Sign-off logged Returned Notes attached ↺ Back to the preparer · feeds AI refinement
The keystrokes leave. The judgment stays.

Chapter two

The infrastructure underneath

Putting AI to work means thinking about a few new pieces of infrastructure — the scaffolding that skills and agents stand on. None of it has to be perfect on day one. Teams start simple, then tune and extend the setup close after close.

01 — The foundation

New pieces to put in place

A real close runs on infrastructure — scheduling, configuration, monitoring, and the connectors (MCPs) that reach your tools and data. These are new pieces most teams haven’t had to think about before. They don’t all arrive at once: you lay the foundation first, then build on it as you go.

02 — The automation

On demand, or in the background

The same workflow can run two ways, depending on the risk it carries. Lower-risk, routine work can run as an agent in the background — preparing the month-end overnight, staged and ready for you the moment you sit down. Higher-stakes work you run yourself as a skill, in the moment, with a closer hand on the wheel.

03 — Operating it

Managed as shared infrastructure

As automation spreads across the team, new questions follow: how to centralize the skills everyone depends on, how to version-control changes so improvements are tracked and shared, how to run the work in a distributed way across people and entities. The focus moves toward operating the automation as shared infrastructure — not a scatter of one-off scripts.

04 — The loop

It improves every close

The setup you start with isn’t the setup you keep. Every edit a person makes and every error the monitor catches flows back as signal, and the skills sharpen with each close.

Shared infrastructure The shared layer every skill and agent runs on Scheduler Monitoring Centralization Context File access Skills You run them on demand Build a skill Write, test, ship Skill library Saved and versioned Agents They run in the background Set up an agent Define and schedule Agent registry Always running Skill designer Designs, tests + maintains the skills and agents — the build role Executor Runs skills on tasks · oversees agents · owns the output — the run role ↻ Improves every close
Start simple. Sharpen it every close.

The catch

Why it has to be earned

There’s a reason none of this moves quickly. In most fields, AI that’s right almost all the time is good enough. Accounting isn’t most fields. A number that’s only almost right isn’t a small problem — it flows into the next reconciliation, the next statement, the audit. A cut-off booked a week early, a related-party entry coded to the wrong account, an impairment resting on a stale assumption: each can surface months later as a restatement. Quiet errors compound.

So the person’s job was never to rubber-stamp the machine. It’s to catch what the machine can’t be trusted to see — the transaction that doesn’t fit the pattern, the assumption that no longer holds, the call that turns on context the model never had. That is the work that earns the trust to let more run on its own.

Chapter three

How trust is earned

Trust here isn’t granted; it’s earned — one workflow at a time, on evidence. The same loop runs every close, and it’s what decides how much can safely run on its own.

01 — Propose

The machine proposes; it doesn’t decide

AI does the work and puts forward a result, with the full trail attached — every figure, every source. Until a person signs off, it’s a proposal, not a posting. Nothing reaches the books on the model’s say-so.

02 — Correct

People catch what the machine misses — and every fix is logged

A reviewer corrects what’s off: the cut-off booked a week early, the related-party entry coded to the wrong account. Each correction is recorded against the workflow. Over time that record is two things at once — an audit trail, and a signal of where the machine is reliable and where it isn’t.

03 — Earn

The record decides what runs alone

A workflow that runs clean across many closes earns more autonomy; one that keeps needing fixes keeps a person in the path. Trust isn’t a feeling or a launch date — it’s a track record. That’s why it widens slowly, and only where the evidence backs it.

AI proposes Result + full trail Reviews & corrects Every fix logged Track record Error rate per workflow A clean record earns more autonomy
Autonomy isn’t a switch you flip. It’s earned — one task at a time, only as fast as trust allows.

Coda

What stays human

The change is quieter than the headlines, and it’s the routine work that goes. The work that can be specified, repeated, and checked moves to the machine. The work that can’t — whether a number is right, whether a story holds, whether a balance can be trusted — is exactly the work that makes someone an accountant.

So the role doesn’t shrink; it concentrates. Less time spent producing the numbers, more spent deciding what they mean and what to do about them — the judgment, the context, the call.

More moves to the machine as trust grows, and the interesting work is always whatever sits just beyond its reach — the judgment it hasn’t earned yet. Which is to say: the judgment was always the job. The keystrokes were just in the way.