If the team cannot explain the workflow in five minutes on Monday, it is not ready for a model. Most AI workflows fail quietly. The model works in a sandbox; the Monday morning handoff does not. The test I use is deliberately simple, so it survives busy weeks, and it can be run as a fifteen-minute stand-up.
The five questions
1. Who owns it? One named person accountable for the outcome — not “the finance team” and not the vendor. If ownership is shared, name who breaks the tie.
2. What is the metric? One number that already matters: cycle time, rework, aged exceptions, review hours, days to close. If the metric only exists because the pilot created it, be suspicious.
3. What happens to exceptions? Where does an item go when the tool is unsure or wrong, who clears it, and by when? A workflow without an exception path is a workflow that silently drops things.
4. What data may the tool see? Name the data classes. Client identity, unpublished financials, payroll and advice drafts are restricted by default. If the answer is “whatever people paste in”, stop here.
5. When is the review? A recurring slot in the calendar where someone looks at the metric and a sample of outputs. If it is not booked, it will not happen in filing season.
If any answer is blank, pause the rollout. Not cancel — pause. Blank answers are cheap to fix before go-live and expensive to fix after.
A worked example: bank reconciliation in a small practice
Take an illustrative CA practice using AI to suggest matches between bank statements and client books. Owner: the manager for that client group. Metric: unreconciled items older than the review date. Exceptions: anything the tool marks below its confidence threshold, plus any match over a value limit the partner sets, goes to a queue the article assistant clears and the manager reviews. Data: bank statements and ledgers stay inside the firm’s approved environment; nothing goes to consumer chatbots. Review: fifteen minutes every Monday on the queue and a sample of auto-matched items.
Notice what is missing: the model name. The test does not care which tool you use. It cares whether the work around the tool is designed. That is also why the test travels well across functions — collections, vendor onboarding, WhatsApp intake, document extraction.
What failure usually looks like
The metric improves in week one because people are paying attention, then drifts back. Exceptions accumulate because nobody owns the queue. Someone pastes client data into an unapproved tool because the approved one was slower. Each of these is a Monday-test failure that was visible early. The earlier short note on the Monday morning test introduced the idea; this is the version I now run with teams.
Related reading: start with the AI for chartered accountants in India hub, then AI tools that pay back for finance and CA teams, Notion as a finance control layer.
Monday-morning checklist
- Name the owner.
- Name the metric that already matters.
- Name the exception path and who clears it.
- Name the data classes the tool may see.
- Book the weekly review — if any line is blank, pause the rollout.
Sources
Educational commentary from implementation work — not personalised financial, tax, legal or investment advice. Verify vendor pricing, privacy terms and your own policies before you buy.