When finance teams talk about adopting AI, the conversation often starts in the wrong place.
Which model should we buy? Which vendor has the best agent? Should we replace the ERP?
Start with a smaller question:
Which reconciliation does your team already know how to do well?
That is where the real architecture begins.
The accountant who understands the source data, control total, exceptions, and review standard already holds the critical design knowledge. The technical work matters, but the method comes first.
A recent Journal of Accountancy article makes this concrete. It walks through an Excel-based agent workflow that cleaned, validated, and reconciled a simulated general-ledger dataset containing 59,157 transaction rows. In that simulation, the full process took about ten minutes on average.[1]
That result is useful, but it is not the main lesson. It is also not a production benchmark. Different data, controls, systems, and review requirements will produce different results.
The more important lesson is how the work was structured.
The accountants did not simply open a chatbot and ask it to “fix the ledger.” They developed a reusable skill, reviewed the agent’s plan, applied guardrails, recorded changes, and kept the output inspectable for a reviewer.[1]
That is what practical agentic AI looks like in finance.
What an AI skill actually is
An AI skill is a set of reusable instructions that teaches AI to complete a task, use the right tools, and check its work.
For a reconciliation, that skill might define:
- the expected input files and required columns;
- how dates, accounts, currencies, and transaction types should be normalized;
- which records must never be changed automatically;
- how the ledger must tie back to the trial balance;
- what counts as an exception;
- what evidence a reviewer needs before accepting the result.
This is not a “magic prompt.” It is the firm’s operating method, written clearly enough for an agent to execute and a person to inspect.
The skill becomes more valuable when it is versioned, tested, and reused. That is how a one-off AI interaction becomes operating IP.
The five controls: METHOD → PLAN → APPROVE → LOG → RECONCILE
A finance team does not need to begin with a fully autonomous agent. It needs a controlled sequence that protects the integrity of the work.
1. METHOD
Write down how a competent accountant performs the task today.
Which file is authoritative? Which transformations are allowed? Which control totals must match? Which exceptions require judgment? What evidence proves completion?
Do not copy the current process blindly. Remove duplicate steps, legacy handoffs, and checks that no longer protect anything. Keep the controls that matter.
This is where the domain expert becomes the architect. The agent cannot infer a firm’s accounting policy, materiality threshold, or review expectation from a generic model.
2. PLAN
Before changing the workbook, the agent should explain what it intends to do.
A useful plan names the inputs, transformations, validations, expected outputs, and stop conditions. If the source schema has changed or a required control file is missing, the plan should say so before execution begins.
This gives the reviewer a chance to catch a bad assumption early. It is cheaper to reject a plan than to unwind thousands of incorrect edits.
3. APPROVE
Approval should match the consequence of the action.
Reading a copy of a ledger and proposing cleaning steps is low risk. Overwriting a source file, posting an adjustment, or changing a system of record is not.
The person approving the run must understand the method and the affected data. “The AI recommended it” is not approval. It is an abdication of responsibility.
For early adoption, keep the agent in a prepared-work lane: it can clean a working copy, flag exceptions, and produce a reconciliation package. A human decides whether the result is accepted or posted.
4. LOG
Every material transformation should leave evidence.
What changed? Which rule caused the change? Which rows were excluded? Where did the agent encounter ambiguity? Which version of the skill ran?
A polished output without a change record is difficult to review and dangerous to reuse. Finance teams need the working trail, not just the final workbook.
The Journal of Accountancy example emphasizes documenting the logic for reviewer inspection and logging modifications.[1] That matters because a reviewer should be able to trace the result without rerunning the entire job from memory.
5. RECONCILE
The agent’s output must tie back to an independent control.
That could be the trial balance, a bank statement, a subledger total, an approved source report, or another authoritative record. If the numbers do not tie, the job is not complete.
This final step is where many impressive demos become real accounting work. Fast transformation is useful. A supported balance is the outcome.
The agent can prepare the reconciliation. The accountant owns the conclusion.
Why Excel is a sensible starting point
Excel is not the future of every finance workflow. It is simply where many teams already understand the data, controls, and review process.
That familiarity lowers the adoption burden. The accountant can compare the agent’s output against a known method instead of learning a new platform while also evaluating a new way of working.
It also creates a bounded first project:
- Use a copy of a recurring reconciliation.
- Define the method and control totals.
- Build the skill.
- Test it against historical cases with known outcomes.
- Run it with human review.
- Record error types, exceptions, reviewer time, and whether the result tied out.
Only then should the team consider broader permissions or live-system actions.
This is the path from AI awareness to adoption: not another chatbot subscription, but one repeatable workflow that produces inspectable work.
What management should measure
Do not lead with the number of agents deployed.
For this kind of finance workflow, measure:
- how often the output reconciles without rework;
- which exception categories still require manual investigation;
- whether reviewers can trace every material change;
- how much preparation time was removed;
- whether the control environment became clearer or weaker;
- whether the skill performs consistently when the input changes.
A faster run with poor evidence is not an improvement. Neither is an automated process that creates a larger exception queue for accountants to clean up later.
The goal is not to remove the accountant from the workflow. The goal is to move the accountant away from repetitive manipulation and toward method design, exception judgment, and accountable review.
Domain experts are the new AI architects
Building agents is becoming a new way to build software. But the people who know the work still define what good looks like.
In accounting, that means the strongest AI architect may not be the person who knows the most about models. It may be the person who can explain exactly how a reconciliation should work, where it can fail, which evidence matters, and when the system must stop.
Give that expert the tools to turn the method into a reusable skill.
Then use a disciplined operating sequence:
METHOD → PLAN → APPROVE → LOG → RECONCILE.
Do it once. Skill it up. Do it again.
Sources
[1] Using an Excel agent to clean, validate, and reconcile data