Stakeholder Gates
Active Workflows
Approval Chains
FinanceBot cost overrun & accuracy gap
4 participants · 2 hours ago
FinanceBot cost per run is $1.20 which is 2.4x our target. Reconciliation accuracy is only 50%. At this rate we're burning $36/day for half the value we expected. @Lisa Park flagging this — do we pause or remediate?
Moving to Amendment stage. The cost profile is unsustainable and accuracy is too low for finance-critical decisions. @Michael Torres can you scope an optimisation sprint? Key targets: cost per run < $0.50, accuracy > 85%.
Scoped the optimisation sprint. Two main levers: (1) Switch from GPT-4o to GPT-4o-mini for the reconciliation check pass — saves ~60% of LLM cost with minimal accuracy impact on structured data. (2) Reduce the document retrieval context window from 8k to 4k tokens — the financial data we're processing rarely exceeds 2k anyway. Estimated new cost per run: $0.38–$0.44.
Ops team notified. I've added a note to the weekly ops report. @Michael Torres — do you need any infra changes for the model switch or is that a config update?
Finance deficit flag — 3 consecutive periods
3 participants · 2 hours ago
FinanceBot flagged a deficit in the Q1 close reconciliation — the automated variance analysis found a $42K gap between the GL balance and the sub-ledger for cost centre 4210. This is the 3rd consecutive period with a variance in this cost centre. Escalating for manual review.
Acknowledged. This cost centre variance is a known issue — it's related to the timing difference on the quarterly accrual entries. Finance are aware and are manually adjusting. @Carol Zhang can you add a note to the FinanceBot config to exclude cost centre 4210 from the variance alert threshold for Q1 close?