Forecast exception and bias review scores every item-location each planning cycle, finds where the statistical line or a planner override is off, explains the cause, lets stable items pass through untouched and routes only the exceptions to the demand planner.
It is cycle day, and the scene chapter 3 used for A3. Overnight the forecasting agent, a role built from the blueprint, scored 4,800 item-locations against version 2 of its rules. For 4,737 the refreshed statistical line sat inside the error band versioned with the rule, so it wrote the new baseline to the planning suite and kept each job status. Sixty-three breached the band, and those are L. Tan’s queue at 09:00.
For material 7741-X at plant 1210 the last three overrides raised the statistical line by 18% on average, and actuals came in below the line each time. The reason code on each was “customer upside”, and the Northbay Logistics forecast upload shows flat demand. The agent proposes to withdraw the October override and to mark the family as stable, so that next month it passes through untouched.
L. Tan works the 63: 58 accepted, 3 edited, 2 rejected with a reason. J. Ridley locks the consensus plan at 11:30. Juno writes the key figures to the planning suite, releases the planned independent requirements to SAP S/4HANA, and keeps the job status and the API response against rule version 2.
Gartner points early agentic efforts at high-volume planning activities with a low cost of error and names touchless forecasting for stable SKUs first; most current agentic capability, it notes, improves query interpretation, recommendations and conversational support (Gartner, 20 May 2026) 1. Forecast review is high in volume, measured every cycle and cheap to get wrong on one item, and a demand plan still takes more than five days to build at a typical consumer goods company (McKinsey, 2 March 2022) 12.
The offers are dense. Logility ships four demand agents, one of which detects forecast winners and laggards and reveals bias patterns, each switchable between autonomous and planner-in-the-loop (Logility, 7 April 2026) 45; the first Kinaxis Agent Studio example analyses forecast quality across business units (Kinaxis, 5 February 2026) 18; o9’s decision replay agent diagnoses overrides, pricing assumptions and safety-stock policy as root causes, and AB InBev runs 90% of planning processes in its advanced markets touchless (o9, 9 June 2025) 39.
Daybreak’s demand agent reviewed 1,200 product-site decisions in one cycle, submitted four in five under policy and held 240 for review, and a consumer goods manufacturer reports $7M a month less inventory; an independent review notes that the figures are vendor-published (Daybreak, product page; Lokad, August 2026) 51 52.
The research on overrides is older and independent. Across 60,000 forecasts at four supply chain companies, upward planner adjustments were much less likely to improve accuracy, and small adjustments often damaged it (International Journal of Forecasting, 2009) 59. Across 3.5 million decisions, a light-touch method that asks the planner only about the items that need judgement matched accuracy with 38% less planning time (International Journal of Forecasting, 2025) 60. That design is the one this use case follows.
After the statistical run, the demand planner exports the forecast and the last one to three months of error, accuracy and bias by item-location, to a spreadsheet, sorts by error and volume, and reviews the worst 50 to 100. History is checked by eye for outliers: promotions, one-off orders, stock-outs recorded as low demand.
Sales and key accounts are e-mailed for market input, and the planner waits. Overrides go in cell by cell in the planning suite or the spreadsheet, often with a habitual uplift. The consensus meeting with sales, marketing and finance locks the plan, which is released to the ERP as planned independent requirements.
A demand plan still takes more than five days to build at a typical consumer goods company (McKinsey, 2 March 2022), across thousands of item-locations per planner, on a monthly cycle with a weekly refresh. The time goes to the items that were fine, because the planner has to look at each one to know.
The cycle by hand: more than five days at a typical consumer goods company (McKinsey, 2 March 2022). Illustrative day split.
This is the A3 scene of the ladder, seen from the machinery. The agent decides and executes within versioned rules, the error band being the rule, and L. Tan handles the exceptions only. The estate already holds the sales orders and the planned independent requirements from the first week; the planning suite’s forecast versions and the sell-through feeds would be added.
| Item-location | Band | Miss | Cause | Proposed correction | Decision |
|---|---|---|---|---|---|
| 7741-X · 1210 | ±12% · lag 1 | +18%, three cycles | overrides “customer upside” · customer upload flat | withdraw the October override; mark the family stable | Accepted 09:20 |
| 8802-K · 1210 | ±15% · lag 1 | −22%, last period | a stock-out recorded as low demand | restore history with the outlier flag | Accepted 09:26 |
| 5513-B · 1220 | ±10% · lag 1 | +9%, two cycles | design-win pipeline moved out a quarter | lower the baseline by 6% | Edited to 4% · 09:41 |
| 9021-M · 1220 | ±20% · lag 1 | +31%, last period | launch ramp ahead of its curve | move to the new-product ramp curve | Rejected: keep the ramp · 09:55 |
| 59 more rows | each with its cause | Worked by 10:40 · 58 · 3 · 2 |
L. Tan’s screen on cycle day: one row per breach, with the band, the miss, the cause and the proposed correction. Showcase estate; illustrative rows.
The market’s usual level is inform and draft-and-approve for the exceptions, with execute within rules for stable items: of the vendors cited above, one sets each agent to autonomous or planner-in-the-loop with a switch, one keeps a person in the loop throughout, and two submit decisions under policy. The blueprint designs statistical baseline forecasting and bias analytics at A3, consensus orchestration at A2, and the plan sign-off and lock at A0. The two agree, and the split is the point: the stable items run delegated, the lock stays human.
| Element | Setting | Set by | Changes |
|---|---|---|---|
| Objective | Keep forecast error inside the band for every item-location at plant 1210, and put only the breaches in front of the planner | L. Tan, planner | New version |
| Data scope | Forecast versions, actuals, overrides and reason codes, sell-through, customer forecasts, lifecycle status; controller boards at plants 1210 and 1220 | A. Novak, data lead | New version |
| Rules | The review policy and its error band by segment, version 2: ±12% at lag one for stable items, ±20% for new items | L. Tan with the Planner Agent | New version, simulated first |
| Trigger | When the statistical run completes each cycle, at period close, and on a sell-through or customer forecast upload beyond its band | L. Tan, planner | At once |
| Actions permitted | Write the baseline · propose a correction · set the segment flag · request sales input · release the consensus plan on the lock | L. Tan, owner | New version |
| Limits | A correction above 15% of volume or $50,000 of revenue waits for the manager · confidence floor 0.80 · the lock is a person’s | L. Tan and J. Harrison, finance | New version |
| Escalation | Breaches to L. Tan · items flagged as sales bias to the account manager with the evidence · above the threshold to the demand planning manager | L. Tan, owner | At once |
| Level | A3 Delegated for the baseline inside the band; the lock at A0 with J. Ridley | J. Ridley, director | Up on evidence, down in one step |
| Feedback | Every accept, edit and reject, with its reason code, is recorded and used in the next cycle | Always on | Fixed |
Every row is a setting or a versioned rule; the level and the limits are the ones a manager signs.
Lima’s agent blueprint is a design of 960 agent roles across five end-to-end chains. It is the map of what Juno can be configured to do, and the first agents are built from it one at a time, the way the PO cancellation agent was.
Additional capability in Plan-to-Produce lists the rest of the map for this chain.
| Measure |
|---|
| Forecast accuracy and bias by item-location and lag (case baseline: SKU-level accuracy up 10 to 12%, McKinsey) |
| Forecast value added of overrides against the statistical line |
| Share of item-locations touchless each cycle |
| Planner hours per cycle (baseline more than five days per plan, McKinsey; 38% less time with light-touch review) |
| Expedite events caused by forecast error (baseline 49%, APQC via Logility) |
| Fill rate and finished goods inventory (case: fill rate up 3 to 5%, finished goods down 6 to 8%, McKinsey) |
Baseline forecast value added first: the accuracy of the final plan against the statistical line, by item-location and lag, because it shows in one number whether the overrides the agent withdraws were helping. The share of item-locations touchless each cycle comes next, read from the run record, so the 4,737 of 4,800 on cycle day has a history behind it. Planner hours per cycle is the measure the desk feels, and the five days of the manual path is its starting point.
Because the research says small and upward adjustments often damage accuracy, and their time comes from the items that need judgement. The band is versioned: a team that wants it narrower sets it and simulates the last three cycles first.
A moved signal is a trigger: a sell-through or customer forecast upload beyond its band brings the item to the planner with the evidence. A launch or an end of life takes the item out of the stable segment.
The demand planning manager, as a vocabulary the agent counts. A code that keeps appearing on overrides that ran above the line is what the quarterly value-added review is for.
The plan is right more often. Between the plan and the shop floor sit the parameters that turn it into orders: safety stock, reorder points, lead times. Who keeps those current?
The agent would recompute each parameter against the service target and the measured variability every cycle, write the small changes and hand the large ones to L. Tan.
Bring two systems and one decision your planners make every week. In the first week we register the data, confirm the mapping with your team, answer your questions in ChatLima and run the first agent at the level you choose.