Supply chain planning.
Past the optimiser.

A planning system runs on master data: lead times, minimum order quantities, yields, capacity, supplier calendars. In most plants that data was entered years ago and has drifted from reality. A supplier whose quoted lead time is four weeks and whose actual delivered lead time has been seven for eighteen months is producing plans that are wrong before any solver touches them.

The highest-value first project in this area is almost never optimisation. It is comparing master data against actual observed performance and surfacing the differences — a task that is straightforward, unfashionable, and reliably produces findings that change plans immediately.

Where language models genuinely add capability is in the unstructured half of the supply chain: supplier emails, order confirmations, shipping documents and advance shipping notices, which carry commitments and changes that never reach the planning system because they arrived as prose. Extracting those into structured signals is a real gap and one nobody had a good way to close before.

Replace a working optimisation engine with a language model. Constrained optimisation is a mature class of problem with excellent classical tooling, and a model that produces a plausible-looking plan without a feasibility guarantee is a downgrade dressed as an upgrade.

Automate purchase order commitments. A system that places orders without human confirmation converts a data error into a financial one. The agent prepares; a person commits.

Promise disruption prediction. Detecting that something has happened and propagating its consequences quickly is achievable and valuable. Forecasting the next unprecedented shock is not, and any vendor claiming it should be asked what they predicted last time.