AI Forecasting Is a Skill Now — Continuous, Driver-Based, and 23 Points More Accurate
July 29, 2026·7 min read
TL;DR
The annual spreadsheet forecast is being replaced by something continuous and driver-based: AI learns the relationships that move the P&L and updates as new signals arrive. The accuracy gap is the headline — 42% of organizations rate their forecasts great or good on average, but that jumps to 65% for teams using AI/ML. The catch: your data readiness, not your tool, sets the ceiling. Here's what an AI forecasting copilot actually is, and how to build one in a single sitting.
The forecast stopped being an annual event
For most finance teams, "the forecast" is still a quarterly ritual — a spreadsheet reopened, rows re-keyed, assumptions re-argued, and a number shipped that's stale by the time the deck is printed.
AI changes the shape of the exercise. Instead of a periodic snapshot, it makes forecasting continuous and driver-based — learning the relationships that move the P&L and updating as new signals arrive, rather than rebuilding from scratch each cycle.
The reason to care isn't elegance. It's accuracy. On average, 42% of organizations rate their forecasts as great or good — but that figure jumps to 65% for teams using AI/ML, per Infosys BPM. That's a 23-point gap between the teams doing it by hand and the teams doing it with a model that learns.
The data
The case for AI forecasting rests on two things: a measurable accuracy lift, and a clear-eyed view of what actually drives it.
| Signal | Figure | Source |
|---|---|---|
| Organizations rating their forecasts great/good, on average | 42% | Infosys BPM |
| The same rating, for teams using AI/ML | 65% | Infosys BPM |
| What predicts the accuracy ceiling more than tool choice | Data readiness | Infosys BPM |
What an AI forecasting copilot actually is
Strip away the jargon and it's a concrete tool that does four things a spreadsheet can't do on its own:
It forecasts from drivers, not from last quarter's number. Rather than trending a total upward, it models the underlying variables that move it — price elasticity, regional win rates, ad-spend decay, renewal probabilities — and rolls those up into the P&L. Change a driver, and the forecast responds the way the business actually would.
It runs scenarios on demand. Best case, base case, worst case aren't three separate weekend rebuilds. Because the forecast is continuous and driver-based, you flex the inputs and read the output in minutes.
It explains the drivers. A number is useless in a board meeting if you can't say why it moved. A copilot attributes the change — this much from softer renewals, that much from ad-spend decay — so the forecast is a story, not a guess.
It's defensible and reviewed. The point isn't to hand the model the keys. It's to produce a forecast a human analyst has checked, understands, and can stand behind — faster than doing it by hand, and more accurate.
That last point matters. The 65% figure belongs to teams who use AI/ML as part of a reviewed process — not teams who outsourced judgment to a black box.
How to get started
You don't need a data-science team to build the first version. You need a real driver tree and a copilot to do the modeling. Here's the loop, built with Claude:
Identify the real drivers. Ask Claude to work backward from your P&L: what handful of variables actually move revenue and cost? For most businesses it's a short list — price elasticity, regional win rates, ad-spend decay, renewal probabilities — not the hundred line items in the model.
Build a driver-based forecast. Wire those drivers to the outputs so the forecast is driven by the relationships that move the P&L, not by copy-pasting last period forward. Now the model reasons the way the business does.
Run best / base / worst scenarios. Flex each driver and let the copilot recompute all three cases side by side. This is where continuous forecasting earns its keep — the scenarios are minutes apart, not rebuilds apart.
Auto-explain what moved the number. Have Claude attribute the change between any two versions of the forecast: which drivers pushed it up, which pulled it down, and by how much. That attribution is what turns a number into a decision.
Add a review step. Bake in a human checkpoint — an analyst who sanity-checks the drivers and the output before it ships. The accuracy lift lives in the reviewed AI forecast, not the unattended one.
Build it live
The distance between "AI forecasting is more accurate" and "I can build an AI forecast" is one working session — because the copilot does the modeling once you know the loop. In Dexity's AI for Forecasting workshop, you build your AI forecasting copilot live in 90 minutes — guided by a practicing FP&A/RevOps analyst — walking away with it working: real drivers, best/base/worst scenarios, auto-explained variances, and a review step you can defend in a board meeting.
Sources: Infosys BPM — AI-driven FP&A forecasting (accuracy figures and data-readiness); Cube — AI forecasting (continuous, driver-based forecasting and the variables that matter).
