Forecasting is dead, long live forecasting

Producing 15 tons of chocolate next month, ordering 200 pallets of frozen vegetables, hedging a contract for so many cubic meters of rubber. Decisions that are based on a forecast that is a running average or a more complicated way of pinpointing expectations. But does this make sense?

Forecasts in the form of a single best guess number are losing their relevance - and if we are honest that is not a new thing at all. The problem is that the number is almost always wrong, and almost always wrong in ways that matter. One customer was expected to order 2,000 tons with a 50% probability. The system spits out an average of 1,000. You produce 1,000. If the customer orders, you cannot deliver. If the customer does not, you have 1,000 tons of inventory you cannot use. Either way, the forecast made you take the wrong decision.

The first fix is well-known: switch from point forecasts to stochastic forecasts. Instead of a single number, the system produces a prediction interval. There is a 95% probability that demand lies somewhere between X and Y. This is genuinely better, and almost everyone working in supply chain agrees that it is.

The improvement that nobody uses

The bigger issue is that the availability of higher quality information often does not translate into higher quality decisions. The reason is simple. Analysts do not have the time to read a prediction interval product by product, work out what it means operationally, and adjust. There are too many SKUs, too many constraints, and not enough hours in the day.

So what happens? People revert to what they know. Simple rules of thumb. A bit on top for the critical customer. A safety buffer here, a safety buffer there. The advanced system is bypassed in the name of pragmatism and experience. The new forecast is technically more accurate, but the decisions are the same as before, and often worse, because the buffers are now stacked on top of a system that already accounted for variability.

In short, the value of stochastic forecasting was supposed to be that you could quantify the trade-off. In practice, the value gets eaten by the time it takes to extract it.

The bottleneck is not the forecast

Once you see this clearly, the question changes. The problem is not that stochastic forecasts are too complicated. The problem is that the planner does not have the bandwidth to consume them, product by product, every week.

This is a familiar shape: a more capable tool, an organization that cannot use it, value left on the table. The instinct of most companies has been to put another tool on top — typically a chatbot or assistant that wraps the forecasting system. The planner can now "ask questions" of the system. Voilà, agentic supply chain.

I do not think this is the answer. A chatbot still requires the planner to be the one asking the questions, which means the planner is still the bottleneck. The bottleneck has just moved one layer up.

Push, not pull

The interesting shift is towards systems where the agents do the asking. A small group of agents continuously scans the data, the forecasts, and the surrounding signals — sales-rep notes, customer communications, news on input commodities, weather — and pushes back to the planner only when there is something worth deciding on. The planner stops being the orchestration layer for the data and becomes the decision layer at the end.

Concretely, this changes what a planner sees. Instead of "the forecast for next month is 1,000 tons", the agent surfaces something like:

Demand is bimodal. Either we produce 2,000 tons and risk excess if the large account does not order, or we produce nothing and risk stocking out if they do. One five-minute call to the account's sales rep will collapse the probability to 99% or 1%, after which the choice is trivial.

That is a very different kind of work. It is also the kind of work humans are actually good at. Supply chain is applied economics under uncertainty, and people are typically very bad at reasoning intuitively about uncertainty. Pre-hashed scenarios with probabilities attached close that gap.

The untapped data is mostly text

A second observation, worth flagging on its own. A large share of the information that would improve operational decisions does not live in the ERP or the data warehouse. It lives in inboxes. Sales-rep notes after a customer call. WhatsApp messages with the account manager. Negotiation drafts. RFQ replies.

In most organisations, this material is structurally invisible to the planning systems. Nobody reads it systematically; nobody links it to the SKUs it affects. With agents and decent language models, it suddenly becomes accessible. The gold mine that has been sitting under the building for a decade can finally be excavated.

Building this for real

All of which sounds inspiring, and like every inspiring technology pitch it is mostly true at the demo and mostly broken in production. The reason most agentic supply chain projects will fail is the same reason most software projects fail: the data plumbing is not there, the integrations with the operational tools are missing, and the users were not part of the design.

The actual work is dull and specific. Connect the agents to the right systems: ERP, planning tool, CRM, the inbox of the sales team. Decide which decisions are pushed and which are not. Co-design the interface with the planners who will use it, not in a workshop afterwards but during the build. Get something that works for one product family before pretending it works for the company.

The good news, and this is a real shift, is that the cost of building this kind of tailored tooling has dropped sharply. Custom development used to be the reason you bought generic SaaS. That equation is changing fast. Tailored, fine-tuned, organization-specific tools, the kind that used to be reserved for the largest enterprises, are now reachable for the mid-market.

The forecast was never the point

Forecasting is not dead in the sense that demand modeling does not matter. It still does. But the forecast was never the point; the decision always was. For two decades we have been pretending that better forecasts would automatically yield better decisions. They do not. They just yield more numbers.

The interesting work is downstream of the forecast: pushing the right scenarios to the right people at the moment they need them. The forecast is the input. The decision is the work.

The team at Crunch has spent the past months working out what actual implementations of such systems should look like. If the above sounds like something you would want to build - get in touch.