The spreadsheet problem

In most companies, demand planning still runs on a monthly ritual: pull a statistical forecast, override it line by line with what planners know, then reconcile the numbers across several meetings. Planners are not the problem. They are filling gaps the system cannot see.

The cost shows up in the warehouse. Much of the stock on the shelves is there because nobody trusts the forecast enough to carry less. That makes forecast quality a working-capital issue, not just a planning metric.

Two kinds of AI, two different jobs

Machine learning and AI agents solve different halves of the problem. Machine learning produces a better forecast by reading more signals, such as competitor prices, freight costs and changes in customer behaviour. Agents handle the work around the forecast.

  • Gather and structure the inputs planners used to collect by hand.
  • Investigate unusual results and explain what changed and why.
  • Prepare scenarios and recommendations before the review meeting.
  • Write approved decisions back into planning systems, within agreed limits.

What better forecasts are worth

McKinsey's 2026 research links each percentage point of forecast accuracy to roughly 2.7 percentage points less finished-goods inventory, though this varies by industry. In chemicals businesses it tracked, accuracy rose 15 to 20 points and on-hand inventory fell 20 to 30 percent in the first wave.

Results vary widely. Across about 100 companies with advanced planning systems, the median inventory improvement was 11 to 15 percent, yet around 42 percent of organisations report no clear return from AI in planning. The difference comes from what the forecast can see, the level of detail it works at, how far ahead it looks and what the business does with it.

From a monthly meeting to continuous planning

When agents prepare the data, checks and explanations in advance, the planning meeting changes purpose. Instead of assembling numbers, people spend their time on decisions that need judgement: accepting or challenging a recommendation, responding to an unusual market signal, or trading off commercial and operational goals.

In one life sciences example McKinsey describes, a redesigned process cut forecast error by about half and reduced the people-hours needed by 83 percent.

Seven questions to ask your team

  • Where would a better forecast create the most value: inventory, service, production or purchasing?
  • Can our forecast see anything beyond our own sales history?
  • Do we forecast at the same level of detail we make decisions at?
  • Are our forecast horizons set by real lead times or by the reporting calendar?
  • How much planner time goes to collecting data rather than using judgement?
  • If accuracy rose ten points tomorrow, who would change the targets and stock levels?
  • When a planner overrides the forecast, do we measure whether it helped?

Source

Figures in this guide come from McKinsey & Company, "Reinventing business planning with agentic AI" (2026). The guidance is Aarvik Digital's own.