Agents Meet Demand Forecasting

Your forecast only knows the numbers. Your planners know the rest. A new class of AI systems is learning to read what they read. What's real, what's early, and where Lyric places its bet.

For years, demand planning ran on a split. A data scientist opened a notebook, called .fit(), then .predict(), read the error metric, pushed the forecast downstream, and moved on to the next SKU. A planner then decided whether to override it, because the model never knew what the planner knew. One tuned the algorithms, the other carried what history never recorded.

The data scientist's half ended on a benchmark. AutoML won the argument: no single hand-picked model, however carefully tuned, beats a system that searches and combines across many. The planner's half is next. A wave of  capabilities - are arriving to give the model what the planner knows. Demand forecasting in Lyric Studio makes the same bet.

How we got here

The notebook was the first stage: a person picked the model, one series at a time. AutoML automated the picking. Libraries like auto-sklearn made model selection, tuning, and ensembling mainstream in the early 2010s, and by 2020 Amazon's AutoGluon was combining everything from ETS to DeepAR into weighted ensembles that beat any single model across dozens of datasets.

Time series foundation models then automated the training. Pretrained across millions of series, they forecast a series they have never seen. Folded into AutoML frameworks they closed the discussion, and deterministic ensembles of them now top the public leaderboards. Two, Chronos and TimesFM in their current versions, descend from models we profiled when generative AI first met forecasting; what was new then is the floor now.

None of those stages required reasoning. TimeCopilot ships an LLM agent too, but its benchmark standing belongs to the fixed ensemble. TimeRouter routes across foundation models with no LLM at all. Reasoning has since overtaken them, and the newest GIFT-Eval entries are LLM-based. The nature of that win matters: a benchmark series carries no promo calendar and no launch plan, so what reasoning wins there is judgment about the pipeline. In demand, the numbers never arrive alone.

What the numbers can't understand

None of these ensemble-oriented systems can see context. Shift the time periods, change the data frequency, rename the covariates, rename the products and the forecast comes out the same.

Every demand series is entangled with events the numbers never show: a seasonal item that never comes back quite the same, a promotion that lifts everything around the item on deal, a stockout the data records as zero demand when the shelf was empty. Teams have always handled the anticipated part with features: promo flags, price, holiday calendars, engineered in ahead of time. What none of it handles is why this launch will succeed beyond expectation, or that a merchandiser has noted a SKU is going away. That context stays in writing, and someone has to read it and understand it’s impact. Today that someone is a planner, and the override is where that understanding shows up.

None of this is news to a planner. The change is that the reading is being automated. A position paper published this February reframes forecasting as an agentic process of perception, planning, action, reflection, and memory rather than a one-shot model call, and newer systems position themselves inside it. Demand planning is the sharpest case for that reframe. Demand answers to events and decisions, which leave documents: promo calendars, assortment plans, launch briefs, delisting notes. The context a demand model needs already exists in writing, unread by any of the old or existing ways of forecasting.

Two approaches are competing for the job. One puts an agent in charge of the pipeline, choosing models, transformations, and post-processing based on context.¹ The other approach wires reasoning directly into the forecast,reading text about what is happening alongside the numbers. ²

Where Lyric places its bet

We take the agentic innovation at face value. The forecasting core is everything the field already proved, automated search over the same foundation models that top the benchmarks. The agent is the new part, with one job no numeric model can do: include the customer context, a stockout, a competitors price cut and augment what they say into the forecast.

A planner can't trust a number they can't check. An agent that understands context and hands its judgment to a proven engine inherits the discipline instead. The old split between data scientists and planners gets redrawn, rather than erased.The ensemble keeps the work it already won, the agent takes the reading, and the planner keeps the calls no algorithm covers.

The fit-predict era ended because AutoML made its case with numbers. The next one belongs to whoever teaches the forecast to understanding context, without giving up the arithmetic that made automation worth trusting. That's where we're building.

Ganesh Ramakrishna and Vishvesh Oza went further into this in New Math, New Moves, covering where the effort actually goes, what the agentic layer is already returning, and what stays with the planner. Watch the conversation.