The AI did not break this firm's books. A rushed correction did, and the agent learned it overnight.
A four-person bookkeeping firm I have been trading numbers with closes the books each month for 40 small-business clients. This spring they dropped an AI categorization agent into QuickBooks Online and cut their monthly close from eight business days to three. Then the same tool quietly taught itself to miscategorize every contractor payment for one client. How that happened is the most useful thing I can hand you this week.
Before the agent, the close was manual. Two staff hand-categorized transactions across 40 client files, about two and a half hours per client a month, with a roughly 2% miscategorization rate caught at review or, worse, by the client. Eight business days a month, gone to the same keystrokes.
The agent sits inside QBO on a review-and-approve flow. It cleared about 95% of transactions on its own and routed the rest to a human. Categorization dropped to roughly 45 minutes per client, the close ran in three days, and the owner stopped staffing the first week of every month around it. On the numbers, a clear win.
Here is the part nobody demos. One staffer, moving fast, recategorized a single subcontractor payment to the wrong account to clear an exception. The agent treats every correction as training. It learned the wrong rule and applied it to that client's next 60 contractor payments, confidently and silently, at machine speed. The error stayed invisible until the 1099 review flagged a vendor total that made no sense.
The tool that categorizes is not the risk. The tool that learns from your worst correction is. A human bookkeeper who fat-fingers one entry makes one error. An AI agent that learns from that entry makes sixty. Speed cuts both ways, and the review process most firms built for human-paced mistakes does not catch machine-paced ones.
The firm's fix was not to drop the agent. They added two controls. First, a locked exception queue for transfers, owner draws, and contractor payments: lines the agent flags but never auto-posts and never learns from without a sign-off. Second, a weekly check of the rules the agent created, not just the transactions it touched. Reviewing output is table stakes now. Reviewing what the AI taught itself is the new job.
Xero now lets bookkeepers build their own natural-language AI agents inside the platform on an OpenAI-backed engine (CPA Practice Advisor, May 13 2026). JAX already forecasts when a client will actually pay and chases the invoice on that prediction. If your clients run on Xero, your reconciliation and AR-chasing busywork is now a build-it-yourself agent.
Claude for Small Business reads from and writes to QuickBooks Online directly: no CSV exports, no copy-paste between tabs. For a solo bookkeeper that means running a categorization pass or a client recon note against live data instead of a stale export. Useful. Still needs your review before anything posts.
Booke AI now makes the agent show its reasoning per transaction before it categorizes, claiming up to 98% accuracy inside QBO and Xero and clearing about 95% of transactions without a human. The 5% it routes to you is the 5% that was always the actual job. Watch the owner-draw and transfer lines.
The AI-native platform categorizes and reconciles with no manual rules, pitched as a QuickBooks replacement rather than an add-on. Tempting for a brand-new practice. Migrating an existing client book onto it mid-year is a riskier call: see this week's Claude Column.
Adoption is no longer the story. When 98% of firms report using AI, most of them daily, the line between a sharp bookkeeping practice and a sloppy one is not whether you use AI but how well you review what it produces. The doer's edge is gone. The reviewer's edge is everything.
What it is: An AI bookkeeping agent that works inside QuickBooks Online and Xero. It reviews bank feeds, categorizes transactions, matches documents, drafts reconciliations, and routes exceptions to a human. Built for bookkeepers and firms, priced from about $20 per client per month.
What it does well: First-pass categorization at scale. The RPA-plus-LLM setup clears roughly 95% of transactions on its own and now shows its reasoning per line, which speeds your review because you can see why it chose a category. For a practice closing 40 books a month, that first pass is the hours.
What it doesn't do well: Judgment on the messy 5%. Owner draws, inter-account transfers, loan proceeds, and personal-on-business charges are exactly where it guesses wrong with confidence. It also learns from your corrections, which is a strength until you make a sloppy one. Keep a separate review queue for transfers and equity accounts.
Pricing: From about $20 per client per month, scaling by volume and tier. Confirm current plans at booke.ai before you budget: the per-client price that fits a 40-client book is not the one a 200-client firm negotiates.
This week's In Practice comes from a reader in Oregon who runs a two-person bookkeeping practice for 25 clients, all on QuickBooks Online. Her question: "Everyone's pushing AI-native platforms like Digits. Do I migrate my client books over, or layer an AI agent onto QBO?"
Layer first. Do not migrate a live book mid-engagement to chase a better AI. The categorization gains are real on either path, but migration carries conversion risk: historical balances, locked periods, prior reconciliations, 1099 history. You absorb every conversion error while still owning the client relationship. The accuracy gap between an AI-native ledger and a good agent sitting on top of QBO is not worth re-keying a year of someone else's books.
When does AI-native make sense? A brand-new client with no history, or a startup you are onboarding from a shoebox. No legacy data means no conversion risk, and the autonomous-ledger approach shines on a clean slate. That is the narrow case.
The deciding question is not which AI is most accurate. It is which AI you can review fastest, because you sign the financials. Pick the tool that shows its work and keeps you in QuickBooks where your muscle memory lives, and revisit the platform question only when you take on net-new clients, not when you convert existing ones.
Pick your three messiest clients and pull last month's transaction register. Count the lines that are owner draws, transfers between accounts, or loan and equity entries. That count is your AI exception queue: the lines no agent should auto-post and you must always review. Write those account numbers down as a standing review rule before you turn any categorization agent loose. The firms that get burned are the ones that let the AI learn on these lines unsupervised.
P.S. Bookkeepers: reply with the one transaction type your AI keeps getting wrong. I'm collecting the list for a future Security Desk on what should never post unreviewed.
