Ask a CPA to review a month-end close and they will do something software vendors keep ignoring: they will read the financial statements. Line by line, this month against last month, with a pen. The checklist matters, the reconciliation reports matter — but the statements are where a reviewer actually forms the judgment that the period is right.
Close software mostly refuses to work this way. It gives you dashboards about the close: progress rings, task lists, status chips, a score. The statements — the thing being reviewed — sit somewhere else, and the reviewer plays a matching game between the findings in one window and the numbers in another.
We just rebuilt DeepLedger's month-end close around the opposite idea, and it went live this week. It's called the Close Sheet: the statements are the page, and everything the AI agent found is pinned to the line it affects.
What the agent does before you ever look
DeepLedger's agent runs the close the way a staff accountant would, through six steps: reconcile all bank and credit-card accounts, review AP and AR, record accruals and deferrals, record depreciation, review the statements, and run a final trial balance. Along the way it scores the period against a 16-point checklist — bank reconciliation, recurring transactions posted, revenue in the right period, prepaid amortization, cutoff, suspense-account cleanup, and so on.
Two design rules govern all of it:
Proposed, never posted. When the agent believes an accrual or depreciation entry is needed, it drafts the entry and attaches it to the close as a proposal — with its reasoning written in first person ("May shows $4,800 in rent; June shows $0 because the bill arrived July 3"), a source reference, and an auto-reverse date where one applies. No adjusting entry a reviewer hasn't seen ever reaches the ledger. This is the same write-safety model that governs every DeepLedger tool.
Completing is not closing. When the agent finishes, the run doesn't close — it moves to needs review and lands in your task list. The agent literally cannot sign its own work; the tool that completes a close refuses to run until all 16 checks are scored, and only a human sign-off locks the period.
In our runs the agent completes 13 of the 16 checks unattended, in roughly twenty minutes of agent time per month. The other three are the ones that should belong to a person: the judgment calls it flagged, the proposed entries, and the signature.
The Close Sheet: review on the paper, not next to it
Here is what you see when you open the close in the portal: a comparative income statement and balance sheet, rendered as a document. Not a summary of them — the actual statements, this month against last.
In the margin, the agent's findings sit exactly where they belong:
- A question about a rent line that doubled? It's pinned to Rent expense, with the agent's reasoning and the evidence it pulled.
- A proposed accrual? Pinned to both the P&L line and the balance-sheet liability it touches, with approve and reject buttons on the note itself.
- An FYI observation — a variance the agent explains but that needs no action — sits in gray and never blocks your sign-off.
Below the statements, the tie-outs run as a band: bank reconciliation, payroll, sales tax, suspense, cutoff, AP/AR, period lock. Each one either ties or carries a note explaining why it doesn't yet.
Every margin note has three exits: decide it yourself, hand it to the AI agent as a task, or fix it yourself with a deep link into the books. When a note is resolved, it becomes a numbered footnote on the document — so the finished close reads the way a reviewed statement should: numbers, with footnotes, signed.
And the signature is real. Sign-off freezes a permanent close record — the statements, the checks, the entries, the footnotes, who signed and when. That's the artifact you keep for the auditor, not a screenshot of a dashboard. (Why that audit trail matters more than raw speed is a post of its own.)
Why documents beat dashboards for close review
The dashboard model optimizes for the person running the close. The document model optimizes for the person responsible for it. Those are different jobs, and with an AI agent doing the running, the second job is the one your time actually goes to.
A reviewer's questions are positional: why is this line what it is? A margin note anchored to the line answers that question where it's asked. A finding in a separate panel — "Check 11: prepaid amortization — action needed" — forces the reviewer to go find the line, re-derive the context, and come back. Multiply that by a dozen findings across a dozen clients and you've rebuilt the exact drudgery the agent was supposed to remove.
There's a quieter benefit too: anchoring keeps the agent honest. Every check the agent raises must name the statement line or tie-out it affects. Nothing vague survives — "something looks off in expenses" isn't a valid finding; "Meals & entertainment is 4× the trailing median, here's the transaction" is.
Where this fits in the product
If you're new to DeepLedger: it's an AI bookkeeping platform for QuickBooks Online — a remote MCP server with 23 tools that works with Claude, ChatGPT, or any MCP-compatible client, plus a portal where the humans review, approve, and now sign. The agent handles daily work like bank-feed categorization all month; the close is where that work gets proven. A month of well-categorized, well-documented transactions makes the Close Sheet mostly footnote-free — which is the point.
The month-end close workflow is built in — nothing extra to buy or enable. Connect QuickBooks, ask your agent to start the close for June, and open the Close Sheet in the portal to watch it draft live.
Try DeepLedger with your QuickBooks account — first month free, no credit card. Or start with the complete guide to what QuickBooks AI can automate today.