Capability Overview of the SCM Assistant Agent
What SCM_AGENT_WITH_RECS is, who it was built for, what it can do across the four SCM domains, how it structures its answers, representative example questions, and its stated limitations — as described in the Signify SCM POC agent handover summary (MathCo, prepared for Signify, 28 Aug 2026, Proof of Concept status).
What the SCM Assistant Agent is
Today, answering a simple supply-chain question at Signify means pulling data from multiple systems, writing a query, and stitching the result together by hand — a process that's slow, technical, and hard to repeat consistently across a planning team. This POC set out to test a different way of working: give planners and managers direct, conversational access to the data itself, so a question that used to take a SQL query, dashboard, or a spreadsheet now takes a sentence typed in plain English.
SCM_AGENT_WITH_RECS, built on Snowflake Cortex, is that SCM Assistant Agent. It sits over Signify's US consumer supply-chain data — spanning purchasing, delivery, demand planning, and inventory — and lets planners ask a question the way they'd ask a colleague. In return, it delivers a plain-language answer built for a business decision, not a data dump: it leads with the number that matters, explains what it means, and, where the result points to a clear next step, names the action and who should own it.
It is especially effective at answering questions about the business's key KPIs — a service-level metric, a forecast number, a stock position — along with other day-to-day monitoring questions such as a ranking or an open-item check within one function (delivery, inventory, purchasing, or demand), which it answers reliably and consistently. It is less reliable on more complex asks — questions that combine more than one domain, track a pattern over time, or ask it to explain the root cause behind a result rather than just report it.
Designed around three planning roles
The SCM Assistant Agent was designed around three planning roles, each using it a different way:
| Role | How they use the SCM Assistant Agent |
|---|---|
| Demand Planner | Asks it to check whether the demand plan is tracking actual sales, spot which products are consistently running over- or under-forecast, and catch stock that's stopped moving before it turns into a write-off. |
| Supply Chain Performance Manager | Asks it what's on hand, what's still inbound, and where safety-stock coverage or vendor lead times are putting supply at risk, turning a periodic stock review into a question they can ask any time. |
| Fulfilment Planner | Asks it how customer deliveries are performing, why orders were missed, and which open orders need attention right now, instead of waiting on a scheduled report. |
What the SCM Assistant Agent can do — by domain
The SCM Assistant Agent is a conversational front end over Signify's US SCM data, covering four domains.
Purchasing (inbound)
- Vendor / supplier on-time delivery performance and lead-time reliability, including how actual lead time compares to the planned lead time.
- Purchase order status at a glance — which POs are open, overdue, or only partially received, and by how much.
- Vendor fill rate and schedule-line fill rate — how much of what was ordered actually arrived, and against which delivery schedule line.
- Goods-receipt tracking.
- Vendor order-confirmation reliability — confirmed vs. reduced vs. still-open confirmed quantities.
- Flags severe lead-time breaches — purchase orders where the vendor came in dramatically later than planned, not just technically late.
- Master-data / planning-parameter completeness — e.g., which material-plant combinations are missing key MRP parameters.
Delivery & fulfilment (outbound)
- Delivery reliability (DRM%) and OTIF (on-time, in-full), overall or sliced by plant, customer, or brand.
- Why customer order lines were missed — no stock, credit block, transport, warehouse, and other reason categories — with a share-of-misses breakdown.
- How many days late a delivery ran once it missed its requested date.
- Order and line fill rates to customers.
- Open and overdue sales orders, and end-to-end order-to-delivery cycle time.
- Which currently open customer orders are at risk of missing because of an inventory or supply shortage, and the projected shortage quantity.
Demand planning
- Demand plan vs. actual sales, and the resulting plan-vs-actual gap — including a longer, multi-month-ahead horizon, not just the current month.
- Forecast accuracy and forecast bias — whether the plan is running over or under actual demand, and by how much, at the standard one-month-ahead horizon planners use.
- Demand variability, to help flag which products are inherently harder to forecast.
- Days-of-supply and stockout-risk flags by material and plant.
- Dead-stock and slow-moving exposure — inventory that has stopped moving and is a write-off risk.
- Whether projected inbound supply (on hand + open POs) is enough to cover the demand plan, and where a shortage is projected and by when.
Inventory
- On-hand and in-transit (goods-in-transit) stock quantity and value, by plant.
- Available vs. restricted stock — e.g., stock held in quality inspection or blocked, versus stock actually usable.
- Aging analysis — how much stock has been sitting 0–6 months, 7–12 months, or over 12 months.
- Dead-stock and slow-moving value at risk of write-off, and where it's concentrated (plant, business unit).
- Safety-stock and reorder-point coverage — which materials are below their safety-stock threshold and need replenishment attention.
- Replenishment lead times and order parameters (minimum/maximum order quantity) feeding into supply planning.
How it answers
Every answer is built for a planner or manager who wants a decision, not a data dump. Answers follow the same fixed structure every time, so they're quick to scan:
Scope → Headline → Breakdown → Insights → Next → Recommended Action
| Element | What it contains |
|---|---|
| Scope | One line stating the filter applied, and the interpretation chosen if the question was ambiguous. |
| Headline | Exactly one sentence with the single most decision-relevant number and its unit. |
| Breakdown | A ranked table (worst / most-at-risk first, with a % of total column), shown only for a ranking or breakdown question; skipped for a single specific number. |
| Insights | Up to 3 short bullets connecting the number to what it means for the business. |
| Next | Exactly one suggested follow-up question. |
| Recommended Action | Included only when the result surfaces something a human can act on. Omitted entirely for definitional answers or a single healthy on-target number. |
- Plain business English throughout — no SQL or technical detail unless specifically asked for.
- Cross-domain questions are handled by querying more than one area and combining the results (e.g. tracing a delivery miss back to a vendor lead-time issue) before answering.
- Answers are grounded strictly in what the underlying data returns — it does not estimate or invent a number it doesn't have.
- Access can be scoped by role: a persona login (e.g. demand planner, inventory manager) automatically sees only the data relevant to that role.
Example questions it answers well
Representative examples by domain, as given in the handover — questions can be phrased naturally, in the user's own words.
| Domain | Sample question |
|---|---|
| Purchasing | “Which vendors have the worst on-time delivery this quarter?” |
| “Which purchase orders are open and overdue past their planned delivery date?” | |
| “What's our vendor fill rate and lead-time reliability by supplier?” | |
| Delivery & fulfilment | “What is our on-time delivery performance (DRM% / OTIF) by plant this quarter?” |
| “Why were customer orders missed last month — no stock, credit block, transport, or warehouse?” | |
| “Which customer orders are open and overdue right now?” | |
| Demand planning | “Show me forecast accuracy for the Hue Connected business unit.” |
| “Which materials are at risk of stockout in the next period?” | |
| “Are we over- or under-forecasting, and by how much?” | |
| Inventory | “What is our current on-hand and in-transit stock position by plant?” |
| “Which materials are below safety stock?” | |
| “What's our dead-stock / slow-moving exposure by business unit?” |
Also cited in the Access/onboarding walkthrough as sample in-chat questions to try:
- “What is our on-time delivery performance by plant this quarter?”
- “Which materials are at risk of stockout?”
- “Show me forecast accuracy for the Hue Connected business unit.”
The SCM Assistant Agent automatically routes the question to the right underlying data (purchasing, delivery, demand planning, or inventory) and returns a natural-language answer. If signed in under a persona role, the data returned is scoped accordingly.
What it cannot do — known limitations
| Limitation | Detail |
|---|---|
| US market scope only | This POC covers the US consumer market and US plants only — it does not cover other regions or business lines. |
| Static snapshot, not real-time | The SCM Assistant Agent works off a fixed data extract, not a live feed. Every answer is stated as of a snapshot date, never "as of today." |
| No product description search | Materials are identified by material code and brand only — there is no free-text product description to search or display in this POC's data. |
| Some reason codes are incomplete | A share of delivery misses and safety-stock/reorder-point records don't carry a specific cause or parameter in the source data. The SCM Assistant Agent will say "uncategorized" or "not available" rather than guess. |
| Won't fabricate an answer | If a question can't be answered from the available data, the SCM Assistant Agent says so and offers the closest available alternative — it does not estimate, extrapolate, or invent a figure. |
| Scoped to SCM | Questions outside purchasing, delivery, demand planning, and inventory — e.g. finance, HR, or general Snowflake/SAP administration — are out of scope for this SCM Assistant Agent. |
| Persona logins see less by design | Signing in under a persona role (demand planner, inventory manager) intentionally restricts what data comes back. This can look like missing data if you expect the full picture — switch to the default role for unrestricted access. |
| This is a proof of concept | Treat outputs as directional decision support while the POC is validated, not yet as an audited system of record. |