Choose a workflow, not an aspiration.
“Use AI across the firm” is not an implementable scope. A stronger starting point is a repeated sequence such as checking which client requests remain open, preparing reminder drafts, and routing unusual cases to a client-service owner.
A promising workflow happens frequently enough to matter, begins with recognizable inputs, produces a reviewable output, and has a person who already owns the result. It should also have a practical rule for cases that do not fit. If the team cannot describe where the work starts, who decides, and when it ends, map it before automating it.
A selection rubric
- Repeatability: the core steps recur, even if some cases require judgment.
- Boundaries: the allowed sources, actions, and end state can be named.
- Ownership: one firm role can approve rules and handle escalations.
- Reviewability: a person can inspect the prepared output before a consequential action.
- Testability: normal, incomplete, conflicting, and unusual examples are available.
Map the controls into the process.
A current-state map should show more than the happy path. Record each input, transformation, system access, decision, output, and handoff. Then mark what can be prepared automatically and what must remain a human action.
Name approved sources
Specify the folders, records, inboxes, or reports the workflow may read. Do not treat general access as a design.
Assign firm roles
Name the workflow owner, approver, and exception recipient so responsibility does not disappear into the automation.
Hold consequential actions
Client-facing messages and work involving professional judgment remain behind human review and approval.
Define when to stop
Missing data, conflicting instructions, unusual clients, or failed access should pause or route—not trigger a forced answer.
For a concrete version of this design, see the client document collection workflow.
Test behavior, permissions, and handoffs.
Testing should use representative scenarios rather than a polished demonstration. Run a complete case, an incomplete case, a duplicate or conflicting input, a permission failure, and an exception that requires judgment. Confirm that the workflow records what happened and gives the next person enough context to act.
Review the prepared output for factual grounding and appropriate tone. Confirm that client-facing or professional actions cannot bypass the named approver. A launch decision should be based on observed behavior against agreed checks, not the novelty of the automation.
Automation becomes operations after launch.
Workflows meet changing inputs, permissions, system behavior, and firm priorities. Monthly management therefore covers monitoring, issue handling, permission review, activity reporting, and agreed bounded improvements. The firm continues to own priorities, approvals, and professional decisions.
This is the difference between a recurring managed service and a one-off assessment or self-serve tool. Firm AI Ops remains responsible for the operating layer of the two launched workflows; the accounting firm remains responsible for the business and professional decisions around them.
For the role of AI inside those boundaries, continue with AI for accountants, or return to the accounting automation guides.