Meta cut teams citing AI efficiency gains. The decision puts CFOs in a similar position: if a board asks them to do the same, what data would they show? Most finance teams cannot point to workflow output numbers, review burden data or error rate comparisons before and after AI tools went live.

The Signal  — Cutting headcount based on AI productivity claims without documented proof is the next CFO liability. If the AI tool slows, breaks or produces errors after the cuts, there is no recovery plan and no audit trail. The damage is both operational and reputational. What to watch: CFOs should build the measurement layer before the pressure to cut arrives. Four numbers matter: task completion rate, review time per output, exceptions found and actual cost savings, not projected ones.

OpenAI added bank account connections to ChatGPT through Plaid. Users can now ask questions based on their real account balances, transactions and spending. Clients may have already linked their accounts and acted on AI-generated financial analysis before talking to their CPA.

The Signal  — CPAs will see clients who made tax, investment or business decisions based on AI answers pulled from live bank data. Firms need to know which accounts clients connected, whether source documents were reviewed and how advice from AI tools should be handled. The risk is not ChatGPT itself. It is the unreviewed decision. What to watch: The short-term risk is clients acting on incorrect AI output. The longer shift is clients expecting their CPA to work alongside AI tools, not instead of them.

SaaS-Bench put the leading AI agent models through real multi-step tasks inside standard business software. Even the top performers finished fewer than 4% of full tasks without errors. Most agents stalled at the point where judgment, navigation or exception handling was required.

The Signal  — Finance teams evaluating AI agents on demo performance will get a very different picture than a 30-day pilot on real work. A 4% task completion rate means human review is not a backup layer. It is the entire product. Buying more licenses does not fix an accuracy floor. What to watch: Vendors are racing to improve real-task completion rates. Before expanding any AI agent contract, run a 30-day pilot on actual workflows and measure completion, exceptions and review hours, not speed.