#title
Cobalt Ridge Automation — Industrial Equipment Warranty Resolution (Sortation-Line Outage to Restored Outcome)

#industry
Manufacturing / Industrial Automation & Material Handling — Aftermarket Service & Warranty

#process
Warranty Resolution: Outage Intake, Coverage Adjudication, Containment, Parts & Field Restoration, Cost Reconciliation, Quality Recurrence


#business goal
Transform process to reduce cost drivers: re-work from repeat site visits and unlinked repeat failures, manual intervention across five disconnected systems of record, and failed restoration SLAs incurring customer credits. Secondary goal of increasing customer satisfaction by restoring the customer outcome rather than closing the claim — target reduction of mean time to resolution from 11.4 days to 3.5 days and P1 line downtime from 96 hours to 24 hours by FY2027.


#covered by KB
true

#target solution (bpmn, case, rpa, other)
case

#complexity
complex

#input quality and type
Full synthetic organizational discovery packet in native formats — two departmental process flow diagrams drawn independently in different tools (PowerPoint from Warranty Administration, Visio SVG export from Field Service), a 90-minute process owner interview, a 2-hour cross-functional workshop transcript with eight participants, a two-slide executive brief with FY2026 baseline and FY2027 targets, a v0.9 draft systems architecture document with an integration gap register and landscape diagram, and a hand-annotated printout of a full legacy warranty claim record (WC-2026-04871). Documents are internally consistent; no version conflicts or contradictions are planted. Quality is realistic-good: each artifact is competent for its own purpose and incomplete for the whole.

#test-failure mode
This evalset tests whether a model performs analysis or merely transcription, across five distinct traps.

(1) LINEAR STITCHING TRAP — the two departmental diagrams, laid end to end, read as a clean sequential flow: intake, coverage, board approval, parts, dispatch, repair, cost posting. A failing model concatenates them and recommends BPMN or a linear Flow. The correct recommendation is Case Management, and the evidence is structural rather than stated: the record must survive 60 days (customer call 2026-03-09 through SAP reconciliation 2026-05-08), evidence re-enters the process non-linearly through repeated email loops, work crosses an organizational boundary that no single manager controls, per-stage SLAs must be tracked and paused, and four conditional exception paths exist that no diagram contains. No document in the packet uses the word "case."

(2) THE MISSING CONTAINMENT STAGE — this is the highest-value signal in the packet. No source artifact contains a containment, interim-mitigation, or workaround step. It is absent from both departmental diagrams, absent from the process owner's walkthrough, and absent from the workshop's step-by-step narration. But the evidence demands it: claim WC-2026-04871 shows 275 hours of customer line downtime during which the claim moved through administrative review and nothing whatsoever was done for the customer's production line. A strong model infers a containment stage that exists in no input document. A weak model reproduces the gap because the sources do.

(3) EXCEPTION PATHS AS EMAIL BEHAVIORS — four conditional paths exist today only as informal email conduct and are described exclusively in pain-point language scattered across the transcript, never as process steps: chasing customer evidence with no clock pause, engineering exceptions sent to a distribution list (7.8-day average), parts substitution agreed by telephone and never documented, and product-quality escalation that reaches only 22% of repeat failures. A correct model surfaces these as distinct conditional stages with their own entry criteria and SLAs, rather than as bullet-point complaints or as inline steps in the happy path.

(4) SOLUTION-SHAPE LEAKAGE — exactly one participant, IT Applications Manager Jon-Erik Lindqvist, articulates the shape of a solution, once, at the 01:38 mark of the workshop transcript. No other document and no other speaker does. A model that anchors its entire recommendation on that single quotation has performed extraction, not analysis; its reasoning should be traceable to the operational evidence (duration, re-entry, cross-boundary coordination, SLA tracking, persistent record) independent of that quote.

(5) AUTOMATION MODE REASONING — the process contains work of genuinely different kinds, and the model should differentiate rather than blanket-assign. Agent-suitable work involves judgment over unstructured or correlated evidence (correlating alarms against service history, summarizing customer uploads, flagging missing and conflicting facts, detecting recurrence across the install base). API-suitable work is deterministic system retrieval and commitment (asset identification, entitlement lookup, parts reservation, delegation-of-authority evaluation, cost reconciliation). Human work is irreducibly judgmental or accountable (approving containment, forming a coverage position, approving technical deviations, authorizing commercial outcomes, confirming safe-service readiness). Process work is orchestration and notification. A failing model assigns everything to one mode or maps modes to org charts rather than to the nature of the work.

Secondary checks: does the model detect that the 88-minute gap between alarm ALM-4471 and the customer's telephone call implies a trigger the business does not currently use? Does it notice that prior claim WC-2025-03318 on the same asset was never linked, and derive a recurrence-detection requirement from it? Does it respect the WT-9 replacement constraint rather than proposing to retire it?

#ba-constraints
1. WarrantyTrack WT-9 is under vendor contract through March 2028 and cannot be replaced within this fiscal year — any solution must operate alongside it, and WT-9 exposes no API
2. Field Service reports through Regional Service Directors, not through Global Warranty Operations — the design must function across an organizational boundary and cannot assume a single management chain or unified prioritization
3. Customer telemetry from Sentinel Historian is governed by per-customer data-sharing addenda and coverage across the install base is incomplete — any telemetry-triggered behavior must degrade gracefully for customers without an executed addendum
4. Delegation of authority is fixed: warranty claims exceeding $10,000 require VP Global Service approval, and claims exceeding $50,000 require CFO co-approval — these thresholds must be modeled, not abstracted away
5. Any action that changes a customer commitment (restoration date, coverage position, service credit) requires sign-off from the VP, Global Service & Aftermarket

#business goal
Transform process to reduce cost drivers: re-work from repeat site visits and unlinked repeat failures, manual intervention across five disconnected systems of record, and failed restoration SLAs incurring customer credits. Secondary goal of increasing customer satisfaction by restoring the customer outcome rather than closing the claim — target reduction of mean time to resolution from 11.4 days to 3.5 days and P1 line downtime from 96 hours to 24 hours by FY2027.
