AI agents for accounting. Procedure included.
The numbers sit in the general ledger, in payroll, in the billing system, in Excel, and in a system you switched off two years ago. No single report has all of it, the work is manual, and the month-end close deadline does not move.
Countz has an agent for each task: reconcile the bank and the balance sheet accounts, review journal entries and transactions for errors, answer the audit requests. Each agent comes with its procedure written: what to check, how much of it, and what a discrepancy means.
The agents are free. They run on the ChatGPT or Claude subscription your team already pays for. None of your data comes to Countz.
Agents you can run today.
Each agent was built for a task a customer had to get done. If your task is not listed, use countz-analysis skill to plan it out interactively, or describe it to us and we build the agent for it with you.
Five groups of agents, by who runs them.
Reconcile the bank and the balance sheet accounts
Every discrepancy researched, with supporting documentation and the adjustments it needs.
Input: the bank statements, the trial balance, the subledgers and the supporting schedules for each accountReview journal entries and transactions for errors and anomalies
Unusual amounts, wrong coding, uncategorized transactions. Each one flagged for review, with the reason.
Input: the general ledger detail for the month, the journal entry listing, and the prior months to compare againstAudit support: answer the audit requests with supporting documentation
Built from the reports you already produce, with the open items listed.
Input: the audit requests, the reports already produced, and the detail behind each oneSOX and internal controls: assemble the evidence for each control
Including the completeness and accuracy of the reports each control relies on.
Input: the control description, the reports it uses, and the detail behind themPrepare the accrual, prepaid and fixed asset schedules
Rolled forward from last month, depreciation and amortization recalculated, each schedule reconciled to the general ledger.
Input: last month’s schedules, open POs and contracts, the invoices that have landed, and the general ledger balancesReconcile payroll to the general ledger
Wages, payroll taxes and benefits, with every discrepancy explained.
Input: the payroll reports, the payroll journal entries, and the accrual and liability accountsDraft the board and lender package, with every figure tied to its schedule
What moved and why, from the reports the company already pulls. Every figure ties to its supporting schedule or the general ledger, and anything that does not tie is listed.
Input: GL detail on both periods, the budget, the memos behind each movement, and the supporting schedules behind each captionCompute each covenant the way the credit agreement defines it
Add-backs traced to entries and clauses, with the headroom on each.
Input: the credit agreement, the trial balance, and the add-back scheduleRecompute the revenue schedules from the contracts and the billing
Under ASC 606, tied to the general ledger, with each contract that does not tie listed.
Input: the contracts, the billing detail, and the revenue accounts in the general ledgerAfter a system implementation, tie the old system’s balances to the new one’s
Converted balances against the originals, and the breaks in the closes since.
Input: the old system’s trial balance, the new one, and the conversion mapping behind themTest the add-backs behind the adjusted EBITDA
Each one checked: does it recur, and is there a document behind it.
Input: three years of GL detail, the bank statements, the add-back schedule and the contractsGet the same schedules out of every company you own
One procedure, whatever system each company runs.
Input: each company’s trial balance and subledgers, in whatever shape they arriveGet your client’s historical numbers ready for their first audit or an S-1
Statements built from the closed trial balances, each figure traced to its schedule.
Input: three years of your client’s trial balances, the schedules behind each caption, and the prior auditor’s fileRun the diligence on your client’s target
Adjusted EBITDA rebuilt from the entries, each add-back tested. The deal calls stay with the deal team.
Input: three years of the target’s GL detail, the bank statements, the add-back schedule and the contractsRebuild your client’s numbers after a restatement or a comment letter
The corrected figures recomputed from the entries, with what moved and in which period.
Input: your client’s general ledger as filed, the corrected entries, and the schedules on both sidesTest 100% of the transactions in any account
Any account, any period. Every item in the population tested against the document behind it, and every exception listed. The conclusion stays with you.
Input: the general ledger detail you are testing, and the documents behind itRead the contracts and test what was recorded against them
Each contract read in full. Every difference between its terms and what was recorded, with the clause behind it.
Input: the contracts, and the schedules or revenue recorded against themReview the workpapers and say what needs the reviewer’s attention
Every workpaper that is not supported, with each open question tracked.
Input: the workpapers for the engagement, and the review notes already raisedVariance analysis against budget and prior month, with the commentary
Each variance traced to its transactions, or to the memo where it was agreed. The commentary is a draft for you to edit.
Input: the general ledger detail for both months, the budget, and the memos where a movement was agreedAnswer a question about a number in the reports
The figure rebuilt from the general ledger and the documentation under it, with what could not be reached listed.
Input: the report the number sits in, the general ledger detail behind it, and the supporting documentationCheck what each entity submitted and tie the group figures back to them
Mapping and translation recalculated, with every submission that does not tie to the group listed.
Input: each entity’s submitted trial balance, the mapping, the rates used, and the consolidated figures“The seller has eleven add-backs to EBITDA. How many of them are real?”
Files Countz read
- Three years of GL detail and the trial balances behind it
- The bank statements for the same period
- The seller’s add-back schedule and the memos under it
- Contracts, invoices and the receivables aging
Procedure Countz ran
- Rebuilt EBITDA from the entries, month by month, for all three years
- Tested each add-back: does it recur, and is there a document behind it
- Traced revenue to cash and rebuilt working capital by month
Result Countz delivered
- A bridge from reported to adjusted, each step tied to its entries
- Seven add-backs supported, three that recur in all three years, one with nothing behind it
- Customer concentration and retention computed from the invoices
- A list of what could not be tested, and why
Install, upload your reports, review the result.
Common questions.
Why not use ChatGPT or Claude on its own?
You can. Each accounting task needs its own instructions, tests on real reports and fixes for what it gets wrong, and every attempt uses part of your subscription’s usage limits. Countz has those agents already written, each with its procedure, so your time goes to the work and the review.
Nine of fourteen model runs flagged almost every clean financial statement as wrong.FinVerBench, fifteen current models tested on SEC filings from 43 S&P 500 companies, May 2026 · arxiv.org/abs/2605.29586
Where does our data go?
It stays in your AI account. None of it reaches Countz, so we hold nothing of yours to secure or delete.
Can I rely on the numbers?
The result shows which files were read and how much of each, traces every figure back to the source data, and lists what could not be reached. You review it as you would review work your own team prepared.
Will the auditor accept the result?
Auditors test what the company produces, however it was produced. The result gives them an audit trail from each figure to its source and a note of what was not covered.
What does it cost?
The agents are free. Running them uses your ChatGPT or Claude subscription like any other task you give it, and counts toward its usage limits.
Tell us the first task you would give Countz.
Describe the task as you would to a colleague. A person reads every request and replies with install instructions.