Demand sensing and replenishment AI agent: a blueprint for supply planning
How a governed demand sensing and replenishment AI agent workforce reads daily sell-through, ranks the divergences that matter, and stages replenishment proposals in your planning system for planners to approve.
- Problem
- Forecasts go stale daily while replenishment runs weekly, so planners pay for late reactions in expedites, stockouts and markdowns.
- Agent design
- Signal intake, sensing, replenishment and forecast feedback agents that propose and never execute.
- Systems
- Planning platform, ERP inventory and orders, retailer and syndicated data feeds, promotion calendar.
- Hard limits
- Agents write only to planning staging areas, never to execution, and never decide customer allocation in a shortage.
- Value levers
- Expedites avoided, lost sales recovered, inventory closer to real demand, planner time released.
- First release
- One category with clean signals, sensing in observation mode, then proposals with full planner review.
Why weekly replenishment cycles miss daily consumer demand
Most consumer goods planning runs weekly. The statistical forecast is refreshed, demand planners apply overrides, supply planning nets it against inventory and open orders, and orders and stock transfers are released. Point-of-sale data, retailer inventory and the promotion calendar arrive in different systems on different schedules. A promotion selling at twice its planned lift on Tuesday is often invisible to supply until the next cycle, and by then the fix is an expedite booked before anyone has checked whether the spike was real.
Planners spend hours each week pulling retailer files into spreadsheets, deciding whether a jump is demand or a data error, and choosing which exceptions to work first. Exceptions are ranked by percentage error, so a large miss on a slow item can outrank a small miss on a top seller.
| Divergence type | Typical signal | What the agents do | May act alone? |
|---|---|---|---|
| Base demand shift | Non-promoted sell-through above or below forecast for several days | Sizes the gap in units and value, proposes a transfer between distribution centers before a new order | Transfers inside an approved band |
| Promotion lift miss | Promoted volume well off planned lift in the first days of the event | Compares lift with similar past events, proposes pull-forward or a cut to the follow-on order | Planner approves |
| Channel inventory build | Retailer orders ahead of retailer sell-through, weeks of cover rising | Flags the sell-in and sell-through gap, proposes slower replenishment to that account | Planner and customer team |
| Seasonal sell-through risk | Seasonal item selling slower in some regions than others | Proposes reallocation between regions while full-price time remains | Planner approves |
| Feed anomaly | Missing store files, late syndicated data, store count or unit errors | Holds every divergence that depends on the feed and alerts the data owner | Never acts on suspect data |
| Constrained supply | Demand exceeds available supply and customers must share it | Prepares the allocation picture by customer and commitment for the decision makers | Never |
| New item or launch | No sales history, forecast built from analogs | Reports early sell-through against the analog curve, proposes nothing | No |
The demand sensing and replenishment AI agent workforce: four roles that propose
The live design names a sensing agent and a replenishment agent. This blueprint adds a signal intake agent in front, because bad data causes more bad replenishment calls than bad models, and a forecast feedback agent behind, so the operating forecast learns from what sensing finds. None of the four can release an order to execution.
| Agent | Job | Reads | Can write | Never does |
|---|---|---|---|---|
| Signal intake agent | Loads point-of-sale, retailer inventory and syndicated feeds and checks completeness, timeliness and units | Retailer and syndicated feeds, store and item master, feed schedules | Harmonized daily signal set and a feed quality status per source | Fill gaps by guessing or pass suspect data to the sensing agent |
| Sensing agent | Compares signals with the operating forecast and ranks divergences by value and service risk | Signal set, operating forecast, promotion calendar, inventory positions | Ranked divergence list with confidence and the evidence behind each flag | Change the forecast or create any proposal |
| Replenishment agent | Turns confirmed divergences into proposed orders, transfers and reallocations using the planning system's logic | Divergences, inventory, open orders, in-transit stock, lead times, approval bands | Proposals in the planning staging area, each with reasoning and its service and cost effect | Release orders to execution, change safety stock or touch customer allocation |
| Forecast feedback agent | Reports persistent forecast bias and which flags planners rejected, and why | Audit log, proposal dispositions, forecast versions | Weekly bias and acceptance report to demand planning | Act on any open proposal or edit forecast models |
Architecture: demand sensing agents beside the planning system, read-only to execution
The planning platform stays the system of record. Agents read forecast, inventory and orders through governed connectors, read external signals through the data platform, and write only to a staging area planners already use, such as a proposal table or scenario. Execution systems receive nothing from the agents.
How one demand signal becomes a replenishment proposal
Two questions route every divergence. Does it rest on suspect data, a supply shortage or a new item? Those always go to a person. If not, is the action inside the approval band set for that item and action type? In early releases no bands exist, so every proposal goes to a planner.
Integrating planning platforms, ERP and retailer data for replenishment
External signals land in the data platform, plan and inventory data come from the planning system and ERP, and proposals return to the planning system. As the live page says, the model matters less than signal latency: sensing on store data three days late is just a slower weekly cycle.
| System | What agents need from it | How agents read | How agents write | Design note |
|---|---|---|---|---|
| Planning platform (SAP IBP, Kinaxis Maestro, o9, Blue Yonder) | Operating forecast, replenishment parameters, network and lead times | Platform APIs or scheduled extracts under a service identity | Proposals into a dedicated planning scenario or proposal table, never the released plan | If a demand sensing module exists, agents consume its output and add ranking, reasoning and the proposal |
| ERP (SAP S/4HANA, Oracle Fusion SCM) | On-hand stock, open sales and purchase orders, stock in transit | APIs or replicated tables in the data platform | None; approved proposals reach the ERP through the planning system's normal release | One path to execution, so agents cannot bypass planner release or ERP controls |
| Retailer and syndicated data (retailer portals, NielsenIQ, Circana) | Store or account level sell-through, retailer inventory, store counts | Governed connectors into the data platform | None | Data use rights and refresh schedules differ by retailer; the intake agent tracks each feed's expected arrival |
| Data platform (Snowflake, Databricks, hyperscaler stack) | Harmonized signal history and the agents' working tables | SQL and platform-native agent tooling | Signal set, divergence list and feed status tables | Most build effort lands here: item and store mapping, unit conversion, history |
| Trade promotion management | Event dates, planned lift, mechanics and funding | API or export of the promotion calendar | None | Late-loaded promotions show up as false base demand shifts |
Build with planning-platform agent tooling, data platform agents, Microsoft Copilot Studio or a custom layer, following where planners already work.
Human oversight in demand sensing and supply planning
Planning oversight is about commitments: inventory bought, capacity used, customers served. Decisions that move money or touch a customer sit furthest from the agents. Approval bands are versioned, owned by a named person, and start empty.
| Decision | Who decides | Agent's role | Control |
|---|---|---|---|
| Ranking and filing divergences | Agent | Ranks by value and service risk | Suppressed flags logged and sampled weekly |
| Stock transfer inside an approval band | Agent, once bands exist | Stages the transfer with reasoning | 5% sampled; override rate above the agreed limit suspends the band |
| Replenishment order changes | Supply planner | Drafts the change with service and cost effect | Planner approves in the planning system before release |
| Promotion forecast adjustment | Demand planner | Shows actual lift against similar past events | Override recorded against the forecast version |
| Production or supplier changes in the frozen horizon | Supply planning lead | Builds the scenario and knock-on effects | Outside agent bands at any size |
| Customer allocation in a shortage | Customer team and S&OP | Prepares the allocation picture | No agent proposal is staged |
| Approval band changes | VP planning | Proposes changes from the feedback report | Versioned; each auto-approval records the band version it used |
Illustrative economics of demand sensing and replenishment agents
Example inputs for one consumer goods business unit with $500 million of revenue in scope; change them to your own. Released planner time is cash only if redeployed, while avoided expedite premiums, recovered margin and lower carrying cost are cash either way. Working capital released is shown as a one-time balance sheet effect.
Fixed assumptions: 52 planning weeks a year; the build cost excludes internal team time and any new retailer data licences. Example inputs only. They describe no client and are not a forecast.
Expedite avoidance usually carries the cash case, so the baseline needs a clean count of expedites by cause. Lost sales recovery is the least certain lever because stockout loss is estimated; compare sensed items with matched control items to measure it. If the case clears only on planner capacity, start with the category that books the most expedites.
Deployment roadmap for demand sensing and replenishment agents
The path follows the live page: one category with clean signals, observation mode until flags prove reliable, proposals with full planner review, then expansion as acceptance supports it. The planning lead signs each gate. Timing depends mostly on how fast retailer feeds map to your item and location master.
- Weeks 0 to 4
Baseline
Map a year of sell-through for one category to the item and location master, measure reaction time, expedites and service today, and agree matched control items.Gate: feeds mapped, baseline and control set agreed - Weeks 4 to 12
Observation mode
The sensing agent ranks divergences daily; planners mark each top flag as real, noise or already known. No proposals yet. Ranking rules and feed checks are tuned.Gate: planners rate 80% or more of top flags as real - Weeks 12 to 20
Reviewed proposals
Proposals are staged with reasoning and planners approve, edit or reject each one. Sensed items are compared with control items, and the pause switch is tested.Gate: acceptance rate at agreed level, pause tested - Week 20 onward
Bands and scale
Planning leadership sets a first narrow band, usually transfers between distribution centers, then adds categories one at a time on the evidence.Gate: per band, per category
How demand sensing and replenishment performance is measured
The first four measures come from the live page; the rest tell planning leadership whether to trust the agents. All are baselined from your own data and this page claims no values.
| Metric | Definition | Why it matters |
|---|---|---|
| Forecast reaction time | Days from a divergence appearing in the signal to an approved plan change | The core purpose of sensing |
| Expedite volume and cost | Expedited shipments and premiums paid, with cause coded | The most direct cash lever |
| Service level, sensed against control | Case fill rate or on-shelf availability on sensed items compared with matched control items | Separates the agents' effect from the season |
| Inventory positioned against demand | Share of inventory in the locations where it later sold | Shows whether transfers and reallocations land in the right place |
| Proposal acceptance rate | Proposals approved unchanged, edited or rejected, by divergence type | Evidence for new categories and bands |
| Auto-approval override rate | Share of auto-approved or sampled proposals a planner reverses | The control metric; above the agreed limit the band is suspended |
| Signal latency | Hours from point-of-sale event to signal available to the sensing agent, by feed | Late data limits everything after it |
Questions supply planners ask about demand sensing agents
Our planning platform already has a demand sensing module. What do the agents add?
A sensing module adjusts the short-term forecast. The agents act on it: they check feeds, rank divergences by value and risk, draft the transfer or order change, and log the planner's decision.
Can the agents place orders with suppliers or release stock transfers?
No. They write only to the planning staging area. Orders and transfers reach the ERP through normal planner release, so existing controls still apply.
What data do we need before starting?
Daily or near-daily sell-through for one category mapped to your item and location master, plus inventory, open orders and the promotion calendar. The baseline measures feed latency per source.
What happens during a supply shortage?
Constrained items leave the automated path. The agents prepare the allocation picture by customer and commitment; the customer team and S&OP decide.
How long before any proposal is approved without a planner?
Not before week 20 for one category, and only if planning leadership chooses to set a band. Running with full planner review for longer works too.
Working with WGA on demand sensing and replenishment
The engagement follows WGA's AI Workforce Solution method and opens with a readiness assessment of your signal data, planning process definition and change capacity. WGA picks the category with the cleanest feeds and the most expedites, maps retailer data to your masters, and agrees control items with your planning lead. Targets come from your own baseline.
Fees are fixed against the outcome and senior WGA consultants lead the build. The first release proves value quickly and tests the controls end to end, including the pause. Your planning and data teams inherit the feed checks, band records, runbooks and operating rhythm to extend the agents without us.
Discuss demand sensing and replenishment in your operation
Talk to our AI practice