Course
AI for Accounting Practices: Automation and Review
Twenty client files, two bookkeepers, and a month end that runs as a system.
Start with
- AI at Work: FoundationsStart from zero and finish with three things you can actually use.
- AI for Accounting and BookkeepingCategorise, reconcile and close a client month without the weekend.
Who this is for
Practice owners, senior bookkeepers and accounting managers responsible for many sets of books and a team. Every example follows Farida, who owns a bookkeeping and accounting practice with twenty clients across e-commerce, trades and professional services, two bookkeepers and a part-time CPA for tax.
You will be able to
- Connect QuickBooks Online through MCP and audit any client in ten minutes without posting anything
- Run the monthly categorisation pass across twenty clients as a Cowork delegation with review
- Build the review checklist Claude runs on every close before a human signs it
- Sweep every client for anomalies and produce the questions list by client
- Turn a closed month into an advisory pack the client reads and acts on
- Write the guardrails: what Claude may propose, what only a human posts
Stops
- Connectors and the read-only auditWire QuickBooks Online to Claude, then get a ranked findings table on any client file in ten minutes, with nothing posted. Set once
- GuardrailsThe written rules that decide what Claude may propose and what only a human posts, with the approval message, the log and the practice Project instructions. Set once
- The categorisation pass with CoworkOne delegation across twenty client files that proposes an account for every uncategorised row, and the review protocol that decides which proposals a bookkeeper actually reads. Weekly
- The review checklistFifteen checks Claude runs on a closed month before a human looks, the exceptions report a reviewer can sign, and the four things the checklist cannot see. Weekly
- The anomaly sweepFive signals swept across twenty files every week, the questions list by client, and the escalation path for the one that stops looking like an error. Weekly
- Client messages at scaleTwenty portal messages drafted in one pass, each in that client's own voice, plus the triage that decides what happens to the replies. Weekly
- The advisory packClosed books into three pages the owner reads: cash, margin, and the two decisions, with a meeting script and a hard line at the edge of the practice licence. Weekly
- Staff, quality and capacityThe review sample chosen by risk rather than convenience, the capacity model built from time data, and the training note that stops an error class repeating. Daily
- Farida's month and the capstoneThe three scheduled routines that carry twenty client files through a month, then the graded capstone: one client from audit to advisory pack with the guardrail log. Weekly

Written for Claude. The ideas transfer to ChatGPT and Gemini; the menu paths don't. Current as of September 2026; where a feature may change we link to support.claude.com.
Claude helped Dasow draft and structure every course. A human at Dasow reviewed every stop.