What to automate first: the Tier 1, 2, 3 rule

Every finance team that has been told to do something about AI faces the same first decision: which process. Pick wrong and the project dies in a compliance review or, worse, quietly succeeds at something nobody trusts. There is a rule for picking, and it takes ten minutes to apply.

Three tiers

1. 01Tier 1: high frequency, rule-bound, low sensitivity. The output is a check or a report; nothing lands in a ledger. These are the first automations. Sub-ledger to GL reconciliations. The balance sheet lead-sheet pack with flux commentary. Actuals load validation, this load versus the last. The forecasting workbook, one tab per cost centre. The audit log that nobody reads, summarized every week. Fixed-asset roll-forward checks.
2. 02Tier 2: the long tail. Recurring, painful, and no vendor ships an agent for it because it is specific to how your company works. Lease budgeting from the lease sub-ledger, two person-weeks a year. Capital carry-forward with the exclusion rules. Vendor insurance and WCB expiry tracking. The close calendar turned into a morning status note. Budget book narrative. This is where a trained team earns its keep, because nobody else is coming to do it.
3. 03Tier 3: touches the book of record or needs judgment. Journal entry preparation and posting. Recurring reclassification entries. Budget write-back to the ERP. Scenario modelling. These are demonstrations of the pattern, and they come last, if at all, and always with a person as the trigger.

The ranking

List the team's recurring processes. For each, write hours per cycle, cycles per year, and a one-to-five score for how rule-bound it is. Multiply. Then strike anything whose output lands in a ledger. What is left at the top of the list is your first two. In the flagship course each table does this in thirty minutes and the sponsor leaves with the company's list.

A pair worth teaching together

The sub-ledger reconciliation is the archetypal first automation: monthly, rule-bound, the approval email already exists, and the worst outcome of a bad run is an unexplained difference that a reviewer catches. The annual regulatory payment report in the extractive sector is the whole pattern in one process, extraction, classification where AI helps, deterministic thresholds and FX in code, and a hard officer attestation at the end. It is a great teaching case and the wrong first automation, because it is annual and attested. Automate the rec first. Automate the report once the team trusts the pattern.

What leadership has to protect

Bottom-up training without protected time produces nothing. The roadmap has to land with a sponsor, someone has to own the first two processes, and the hours to build them have to be real. That is most of what the leadership briefing is about.