Beneath

In a legacy retail distribution center, the ICQA Manager authors the Weekly Business Review that leadership relies on for site quality. The current rules-based platform deterministically picks the worst-gap process path as the deep-dive focus — and in a meaningful share of weeks, that pick misses the real story. When it does, the manager must do analytical work the tool cannot: spotting a long-running out-of-tolerance pattern, an outsized single-shift anomaly, or a department drifting quietly below the threshold, then rewriting the narrative to match. That edit-required cycle adds up to 2 hours and is the single largest variable cost in the workflow — materially worse as the site consolidates from two ICQA Managers to a solo-coverage model.

Beneath is a quality investigation agent for weekly business reviews. It starts from a deterministic dashboard layer to build trust in an environment that is not yet AI-native, then adds an agentic investigator and composer that generate secondary callouts, speaker's notes, and a ready-made presentation deck. The deterministic worst-gap logic remains the spine of the deep-dive; the agent's editorial authority is deliberately bounded to secondary callouts and speaker's notes that enrich the narrative with cross-week judgment. The workflow includes review, per-callout veracity flags, feedback, regeneration, and a scheduled weekly run — collapsing the 1–2 hour edit cycle into a roughly 5-minute review cycle in most weeks.

Beneath runs a two-phase Claude Opus 4.7 conversation on a weekly cron (Monday 9 AM Central). An investigator phase reads the deterministic deck's payload, then loops through investigation-tier API endpoints — every reason code, workstation, team member, and freight category — building structured candidate observations with confidence scores and citations. A composer phase writes the final callouts and speaker's notes in the voice of an experienced site quality manager. A layered guardrail stack gates every claim: a forbidden-phrase guard applied at three points to keep output observational (never prescriptive), a bare-code regex that rejects unlabeled reason codes, and a confidence floor at 30 that drops weak observations. An advisory veracity layer — deterministic grounding plus a best-effort LLM judge — surfaces per-callout flags in the ReviewPanel, and a run-level eval gate requires all factual-accuracy checks plus 80% of semantic checks to pass; a failed gate triggers graceful degradation to a deterministic-only deck.

Beneath is a B2E internal product inside a Fortune 50 omnichannel mass retailer. The primary user is the ICQA Manager at a single distribution center, who owns site-level quality reporting and manages a team of 30–40. Secondary users are Operations Managers and Senior Operations Managers, who gain self-serve access to current reports, and tertiary consumers are the Site Director and Operations Director. The economic buyers are the DC operations leadership chain — Site Director and Operations Director — who sponsor tooling that reduces reporting burden and improves review consistency. Regional VP and above are explicitly out of scope.

The site is piloting a future network staffing model, consolidating two ICQA Managers into one; going from an edit-required reporting cycle to a 5-minute review cycle is the difference between a workable solo role and one absorbed by reconciliation and authoring. The business value is realized indirectly: manager hours redirected to floor work and coaching, faster leadership decision cycles, and quality issues caught earlier. The North American warehouse operations software market is projected to grow at roughly 10–14% CAGR through 2028, with AI-assisted reporting growing faster. Beneath is now built and published to production, deliberately observational rather than prescriptive — the appropriate posture for a supply chain building trust with AI before any prescriptive surface earns its place in a future phase.