Forecasting with AI assistance
AI's four real jobs in forecasting — decomposition analyst, driver-model architect, assumption drafter, and challenger — and the ownership line that never moves.
Now the honest version of 'AI-powered forecasting'. There are two technologies in play and they do different jobs: classical statistical/ML forecasting (exponential smoothing, ARIMA-family, gradient boosting — often hiding inside your planning tool's 'predictive' feature) is genuinely good at high-volume, stable-pattern series; LLM assistants are genuinely good at the analyst work around the forecast. Confusing the two produces either naive overtrust or dismissive undertrust. For a finance team in spreadsheets, the LLM's four real jobs:
- Decomposition analyst — paste the two-year weekly series and ask for level/trend/seasonality/events with the evidence for each claim ('weeks 38-48 average 22% above the annual mean in both years'). Verify against your own eyeballing. This turns your chart-staring into a written, checkable analysis in minutes.
- Driver-model architect — describe the business and let AI draft the driver tree: revenue = loads × revenue/load; loads = contracted lanes + spot; collections = revenue lagged by DSO distribution; disbursements = carrier settlements (weekly) + payroll (biweekly) + fuel (weekly, indexed) + fixed runs (monthly). You'll edit it — that's the point; editing a complete draft beats architecting from blank, and the tree is the forecast's logic, readable by anyone.
- Assumption drafter — for each driver, AI proposes the assumption with its basis: 'DSO 34 days (trailing 6-month average, range 31-38)'. Every assumption lands in an assumptions log — the single artifact that separates professional forecasts from spreadsheets with opinions: when actuals miss, you diagnose which assumption broke instead of relitigating the whole model.
- Challenger — the pre-mortem, forecast edition: 'Here's my 13-week forecast and assumptions. Attack it: which assumption is most fragile? What would week 6 look like if I'm wrong about X? What did comparable situations do?' AI is a tireless red-teamer of forecasts precisely because it holds no attachment to yours.
The ownership line
The line that never moves: the analyst owns the forecast. AI drafts structure, proposes assumptions, stress-tests logic — and then a named human adopts each assumption, signs the forecast, and answers for it in the meeting. This isn't ceremony. A forecast is a commitment device — treasury draws on it, the CFO plans around it — and commitments need owners the way controls need owners. The fastest way to lose forecasting credibility in the AI era is the sentence 'that's what the model said'; the fastest way to build it is 'here's my forecast, here are my assumptions, here's which one I'd watch.'
The paste-and-ask failure mode from every prior course applies double here: an LLM summing 13 weeks of cash flows will be confidently, slightly wrong often enough to matter. The division of labor is absolute — AI proposes structure and assumptions; the spreadsheet computes; you verify. If your assistant runs actual code on the data (an analysis mode), the numbers are computed and the check shifts to logic. Know which mode you're in before quoting anything upward.