The architecture beneath the briefing.

Three eras of commerce. Grocery skipped the last one.

Traditional → Conversational → Autonomous. The technical reason grocery stalled at conversational: a generic LLM has no binding to a live catalog, no household memory, and no price truth. In most verticals that's a weak answer. In grocery it's a hallucinated allergen — a liability the brand owns. Run the same prompt down both paths:

And every digital cart pays a middleman tax.

62% of digital discovery is hijacked before a click lands on the retailer's domain; aggregators and 3rd-party retail-media networks siphon 60–70% of the most profitable ad revenue the SAP-powered shelf generates. But a retailer-owned agentic surface layers the margin back — without replatforming off SAP Commerce. The tax on the left; the answer on the right:

An intelligence layer — not a replatform.

Delectable AXP sits on top of SAP Commerce Cloud and S/4HANA via API. No rip-and-replace, no 18-month systems integration. Grocers move from RFP to live pilot in 90 days — RISE customers faster. Four properties make it an operating system for grocery, not another LLM wrapper:

Three ways to answer the same intent — only one is a stack.

A generic agentic layer (the kind a hyperscaler ships) is an LLM wrapper: it generates. Delectable AXP adds the layers underneath generation — multi-model routing, household memory, and non-generative grounding against SAP Commerce + S/4HANA. The difference isn't the model; it's everything below it.

The shopper journey — and the data & protocols beneath every stage.

CX thinks in a funnel, so follow one intent down it — Discover → Plan → Cart → Checkout → Reorder. At each stage you see the omnichannel surfaces live there (web · SAP Emarsys email · social · in-store · agentic) and the exact systems, data, and protocols on the wire — OCC REST, S/4HANA OData, ACP/A2A, MCP, A2UI, UCP — that carry it. Click a stage to drill in.

And the per-product integration surface beneath the flow.

The same journey, mapped to each SAP product — the concrete APIs, data objects, and events, and which direction they move. The rows tagged expansion are where the technical conversation goes next: the S/4HANA grounding stack and the identity & data-residency layer your CISO will ask about.

Intent in. Cart out. 500 milliseconds.

"Friday family dinner, gluten-free, $80 budget, the kids will eat." → the Food HyperGraph plans a recipe grounded in shopper history → live SKU lookup against SAP Commerce → priced from S/4HANA → basket assembled, pre-substituted, under budget. +18% basket lift · 3× conversion · 0 hallucinations. Here's the assembled cart — pre-substituted, under budget, with the glass-box reasoning overlay — then three live surfaces to walk the full journey:

Two proprietary graphs Joule can't build alone.

The Food HyperGraph (142k ingredients · 4.2M edges · 9 dietary vectors · refreshed daily) and the Shopper HyperGraph (847 signals/household · 94% prediction accuracy · fed by SAP Emarsys + S/4HANA). Here's how they're built from the SAP integration — and how they project into HANA Cloud's native Knowledge Graph + Vector engines.

Enrichment is where the catalog becomes intelligence.

A raw SAP article master is a price and a description. Enrichment maps every SKU across ingredient × flavor × nutrition × dietary × cultural context — the step that turns SAP Commerce data into something an agent can reason over without hallucinating. It starts with master data management: many source records become one golden, agent-ready record. Step through it:

And the behaviour it unlocks.

Enriched, golden product data + shopper signals surface as live insights — the production analytics the merchandising and marketing teams act on.

The grocer's CMS — agent-ready content & media, by construction.

This is where it clicks for the CX & Commerce team. Recipes, articles, UGC social video, food-research data and the grocer's own content all become agent-ready — schema-aware, tenant-scoped, content-addressable — and surface through placements the agent composes. Author once; the agent uses it on every surface. Media generation (recipe creation, short-form UGC video) feeds the same pipeline — and SAP Emarsys activates it.

Grounded, routed, evaluated — not generated.

The pipeline behind 0 hallucinations at 490ms: context resolved from the Shopper HyperGraph → the candidate set collapsed to the store's live assortment → constraints (dietary, allergen, budget) applied before generation, not after → health claims evidence-graded → cart returned. Two deep dives written for your architects:

Who steers the agent — operator, merchandising & marketing studios.

The agent is autonomous, not unsupervised. Merchandisers and marketers steer it through control studios — promotion, substitution, audience and brand-safety levers; an A/B test harness with holdouts; and multi-touch attribution. The marketing studio activates through SAP Emarsys; the merchandising studio grounds in SAP Commerce + S/4HANA. Every change is auditable.

And the agentic control surfaces themselves — rendered SAP-branded (your tenant theme, the same product, in sync). These open the live consoles:

Who can steer what — multi-tenant entitlement.

The studios are multi-tenant by default. An organization is entitled to a set of modules; roles grant read / write / admin per module; and every grant is capped by the tenant's entitlement — scoped through SAP identity (IAS / XSUAA). A role only sees the studios its tenant has switched on.

Zero hallucinations is an architecture, not a promise.

Outputs are constrained to the retailer's verified SKU catalog and to real-time inventory + price from S/4HANA — liability mitigated by design. Hard rules (allergen, dietary, recall), regulatory compliance pre-validated at emit, a full audit trail traceable in under 5 seconds, and model provenance bounded to the tenant's live catalog. The SAP IT-security review becomes a checklist, not a research project — the governance substrate below sits in the same execution path as the agent.

Hyperscaler performance. Cloud-software margins.

Not theoretical. Giant Eagle — a $9.3B regional grocery leader — validated Delectable AXP across 216 stores in Q1 2026. The figures below are measured on 1.2M production requests, not projections.

From handshake to go-live in 150 days.

A pragmatic, milestone-based plan. Every phase has a named owner, a measurable exit, and a kill-switch if the joint KPIs aren't met. No 18-month vapor commitments.