The workflow is one key and a loop. Every client is a book under one account, the agent works book by book, importing, categorizing, flagging, reconciling, and the practitioner reviews only the flagged items before locking each period. That is the whole shape. What changes versus a traditional practice is not the accounting; it is who does the volume. The agent does the mechanical share across every client, and your hours move from data entry to review. Here is the month end in detail, including the part you should never hand over.
What does the setup look like?
On LedgerMCP, a firm is one free account holding unlimited client books. Each book is a separate business: its own chart of accounts, transactions, reports, and lock dates. Your agent, Claude, ChatGPT, Claude Code, or any MCP client, gets a single API key that works across all of them. It can list your clients, open the right book, and take on a new client by creating a fresh book with a seeded chart, no per-client signup or subscription. Keys can also be scoped, read-only or limited to specific books, if you want a tighter leash. The economics and access model behind this are covered on the bookkeeping for CPAs page; the point here is that setup is a one-time cost, and after it, month end is the same loop every month.
What is the month-end workflow, step by step?
- Set up one book per client. Done once. Each client is its own book under your account, and your agent reaches every book with one key.
- Import each client’s month. Hand the agent the bank and card CSVs, or let connected feeds sync. Imports dedupe, so an overlapping statement never double-posts.
- Categorize and flag. The agent categorizes against that client’s chart and the conventions you have recorded, posting balanced double-entries, and flags anything ambiguous as a question instead of guessing.
- Reconcile every account. The agent ties each account to the statement ending balance. A clean tie is recorded as proof; a nonzero difference comes back with suspects.
- Review the flags and lock. You work the flagged queue, make the calls, review the trial balance, and lock the period.
Then the agent moves to the next book and the loop repeats. If you want to see the single-book version of this close narrated in full, our walkthrough of a month-end close with Claude runs it end to end.
How does the agent work each book?
Inside one client book, the agent’s pass is mechanical and checkable. It imports the month, matches transfers between the client’s own accounts so moves are not double-counted, categorizes each transaction into a balanced posting, and splits the ones that need it. Anything it is not sure about, an unfamiliar vendor, a charge that could be owner or business, becomes a flagged question addressed to you rather than a silent guess. Then it reconciles: each bank and card account is tied to the statement ending balance, and the tie-out is persisted, so "March tied out on April 3rd" is a recorded fact, not a memory. When the difference is not zero, the agent gets back the likely suspects to chase. The mechanics are on the bank reconciliation feature page.
What do you actually review?
Not everything, and that is the point. Re-reading every transaction the agent posted would recreate the hours you just saved. The review surface is three things: the flagged queue (the agent’s explicit uncertainty), the reconciliations (does every account tie to its statement), and the trial balance before you lock. Because every posting balances by construction and reconciliations are proven against statements, a clean queue and clean ties mean the mechanical layer is sound, and your attention goes where judgment lives. Practices for keeping the agent’s accuracy high over time, conventions, spot checks, and drift control, are in our post on keeping AI bookkeeping accurate.
What should you never delegate?
Be honest about where the line sits, because the engagement letter has your name on it, not the agent’s. Three things stay with the practitioner:
- Judgment calls. Is this deductible, is this equipment or repair, how should this loan and its interest be structured, does this client’s situation change the treatment. The agent can propose and cite the pattern; you decide.
- Client communication. Questions to the client, uncomfortable findings, scope changes, and advice all come from you. An agent asking your client about a suspicious charge is not a relationship; it is a liability.
- Sign-off. Locking the period, delivering the statements, and standing behind the numbers is professional responsibility, and it does not transfer. The agent prepared; you reviewed and closed.
This is the same division of labor a staff bookkeeper and reviewing partner already use. The agent is the staff role at unusual volume, with the useful property that every write it makes lands in an append-only audit log under its own credential, is reversible in one click, and cannot touch a locked period. Delegation with a paper trail is what makes the model defensible.
Where does this run?
The workflow above assumes software with the right shape: unlimited client books under one account, an agent write surface with flags and dry runs, persisted reconciliations, period locks, and standard double-entry exports. That is the shape LedgerMCP is built to, free, with your own agent doing the work. What to look for when comparing options, and how the mainstream firm ecosystems price the same job, is in the companion post on bookkeeping software for accountants, and the firm-level picture is on bookkeeping for CPAs. The honest caveat: it is cash-basis oriented v1 software without invoicing or payroll, so pick your pilot client accordingly. Start with the messiest backlog book, run one close this way, and compare the hours.
Quick answers
Can a CPA use AI to run client bookkeeping?
Yes, for the mechanical share: importing, categorizing, transfer matching, and reconciling across client books, with anything ambiguous flagged instead of guessed. The practitioner reviews the flags, makes the judgment calls, and signs off. The split only works on a ledger with real double-entry and an audit trail.
How does one API key work across many client books?
On LedgerMCP, every client book lives under one free account, and a single API key works across all of them. The agent lists the books, opens each in turn, and runs the same close loop per client. A new client is one create-business call with a seeded chart of accounts, not a new signup.
What should a CPA never delegate to the agent?
Judgment calls (is this deductible, is this owner or business, how should this loan be structured), client communication, and sign-off. The agent does volume and flags uncertainty; the practitioner decides, talks to the client, and locks the period under their own name.
What does the practitioner actually review each month?
The flagged queue, the reconciliation for each account, and the statements. You are not re-reading every transaction; you are working the exceptions the agent surfaced, spot-checking against the reconciled statement balances, and reviewing the trial balance before locking.
