Bookkeeping with AI is already happening. Agents code transactions, route AP invoices, categorize expenses, draft journal entries, flag variances before a controller has logged in. The work is real. The problem is what happens to it after.
When an agent codes a transaction at 2am, who reviewed it? On what basis? If it flags a $48,000 variance, where does that flag live until the controller sees it Tuesday morning? Most teams running AI bookkeeping today answer with a chat log or a screenshot pasted into Slack. None of that is an audit trail.
Platforms hold the ledger. Dock holds the agent's output.
Here is the architectural claim. QuickBooks Online, Xero, NetSuite, and Sage Intacct remain the system of record for ledger entries. Ramp and Brex remain the system of record for card transactions and reimbursements. None of that changes. Dock is the system of record for the agent's interpretive work around those entries: variance recommendations, journal-entry drafts, partial reconciliations, policy-flag rationales, the "I checked these seventeen line items and two look wrong" reasoning that today vanishes the moment a chat session ends. Dock rows carry pointers back to the source of truth, like a qbo_journal_entry_id or xero_invoice_id field. The agent reads fresh from the platform each time, and Dock's consent gate fires the actual mutation back. The interpretation persists in Dock. The posted entry lives where it always has.
Three workflows show what this looks like in practice.
Month-end close with variance flagging. An agent reads the prior period's P&L from NetSuite or Sage Intacct, pulls the current month, and surfaces every account where the variance exceeds a threshold the controller set. Each flag becomes a row in Dock with reasoning, links to source transactions, and a status column. The controller works the queue. Downstream journal-entry drafts route through a consent gate before posting back to the ledger. The decision row stays in Dock. The posted entry lives in the ledger.
AP invoice routing and approval. Vendor invoices arrive. An agent reads, extracts, matches against POs, and proposes a coding plus an approver. The proposal becomes a row in Dock with vendor_id, GL account, amount, confidence, and a link to the original document. Approvers see what the agent saw and why. On approval the entry posts to QuickBooks Online or Xero. The Dock row keeps the full trail.
Expense categorization and policy enforcement. Ramp and Brex transactions flow into an agent that categorizes against the chart of accounts and checks against policy. Borderline cases become rows in Dock with the agent's interpretation of the policy clause it relied on. A finance reviewer resolves the row. The categorization syncs back to the card platform and the GL.
Why accounting specifically needs this
Accounting is one of the few domains where "an agent did it" is not enough. PCAOB AS 2201 requires evidence that internal controls over financial reporting operate as designed, meaning anyone asking has to trace a posted entry back to the actor that authorized it and the basis for that authorization (PCAOB AS 2201). Thomson Reuters' 2025 reporting finds that 79% of tax and accounting professionals expect AI to have a high or transformational impact within five years, while most firms still lack policies for agent-produced work (Thomson Reuters, The Impact of AI on the Accounting Profession). That gap between adoption and defensibility is what Dock is built to close: every agent action is a row, attributed to a named agent identity, time-stamped, linked to its inputs, reviewable later. The work survives an internal audit, an AICPA peer review, or a regulator inquiry six quarters out.
Open /use-cases/accounting to see the worked example, or fork the ap-invoice-routing-approval template to put a real workflow in your workspace today. Adjacent reading lives in Dock for Finance, which covers FP&A and treasury patterns that pair with the close. For the principles underneath, see Agent Audit and Compliance, Agents Are Principals, The Dangerous Ops Contract, and Why We Kept Flat Pricing. And for the workflow-side counterpart to this essay, AI for Bookkeeping and Month-End Close walks through the close cadence step by step.
FAQ
Does Dock replace QuickBooks, Xero, NetSuite, or Sage Intacct? No. Those platforms remain the system of record for ledger entries, AP, AR, and the chart of accounts. Dock is the system of record for the agent's interpretive work: variance reasoning, draft entries, policy flags. Dock rows reference the ledger entry by ID. Mutations fire back to the platform through a consent gate.
How does Dock handle SOX defensibility? Every agent action is a row with a named agent identity, a timestamp, the inputs it read, the rationale it produced, and the human who reviewed or approved. That trail is what an ICFR audit under PCAOB AS 2201 actually asks for. The posted ledger entry points back to the Dock row that authorized it.
Where does the agent's reasoning get stored? On the row itself. When an agent flags a variance or proposes a journal entry, the reasoning is a first-class field, not a side comment. Reviewers see what the agent saw and why. The reasoning survives chat-session boundaries.
What about audit trail integrity? Rows in Dock are append-only at the event layer. Edits to a row produce a new event with the prior value, the new value, the actor, and the timestamp. Pulling the full history of a single row reconstructs the decision sequence end to end.
How agent-ready are the platforms today? Honestly mixed. QuickBooks Online, Xero, and NetSuite all expose APIs that support read-fresh plus posted-mutation patterns, which is what an agent needs. Sage Intacct supports the same pattern with somewhat heavier auth setup. Ramp and Brex expose card and transaction data well, though policy enforcement still benefits from the agent's interpretive layer sitting outside the card platform. None of these need to change for Dock to fit alongside them.
Can a small firm use this, or is it only for SOX filers? The same trail that satisfies SOX also answers the simpler question a small firm faces: who did this, and why. AICPA peer review, lender due diligence, and tax-return support all benefit from the same persistence. The workflow is the same. The audience for the trail is what changes.