Signal 01
Where teams are actually applying AI
Across the FP&A teams we've spoken with, the common pattern isn't asking AI to build the forecast — it's asking AI to help explain one that's already built. Once a forecast is assembled in the usual spreadsheet or planning tool, the next step is almost always a narrative: what changed versus last month, why, and what it means going forward. That narrative step is where AI is doing real, adopted work.
This lines up with a broader pattern across Kingston's workflow library: models are most reliably useful on structuring and explaining numbers, not on originating them from scratch.
Signal 02
What this looks like in practice
A typical cycle: FP&A finalises the forecast numbers in their existing model, then feeds the period-over-period changes to a model like GPT or Claude with a prompt asking for a first-draft variance narrative, grouped by driver. An analyst then edits that draft — trimming it, correcting anything the model got wrong about causation, and adding context only a human would know.
Teams report this cuts drafting time on the narrative significantly, without changing who owns the forecast itself or its assumptions.
Signal 03
Where teams ran into trouble
The teams that reported the least success were the ones that tried to have a model touch the forecast model directly — generating formulas or assumptions inside the spreadsheet itself, rather than working from its output. That approach produced more rework, not less, largely because errors in an AI-suggested assumption are much harder to catch than errors in an AI-suggested sentence.
The practical lesson teams converged on: keep AI downstream of the forecast, not inside it.