McKinsey published One Year of Agentic AI: Six Lessons from the People Doing the Work by Lareina Yee and colleagues — based on 50+ agentic AI builds. The full piece is worth reading. Here's the FP&A-specific read on each lesson.
FP&A teams are deploying agents on top of broken processes and wondering why results are underwhelming. If your close requires three analysts reconciling data across four systems, an agent automates the chaos. Fix the process first.
Not every FP&A task should be automated. Budget conversation with a business leader. Read on whether a headcount ask makes sense. Judgment call on how to frame a tough number for the board. Those belong to a person.
Bad output that makes it into a model or a board deck doesn't fail quietly. Finance credibility is hard to rebuild. Invest in output quality and evaluation before you scale anything.
This is table stakes in finance anyway — every number needs to be auditable. Build monitoring into the workflow from the start, not as an afterthought.
Agents handle tasks. Finance business partners handle relationships, judgment, and context. The risk is automating the wrong half.
The systems that improve over time are the ones where output quality is being measured and fed back. In FP&A, that means comparing agent outputs to actuals, to human-reviewed versions, to what the business actually found useful.