Industry Perspective

We Rebuilt Reverse Logistics
from First Principles.

Agentic Engineering — software that reasons, decides, and learns — applied to one of e-commerce's most persistent operational challenges.

User-Centred Design
Systems Thinking
Agentic Engineering
E-commerce returns operations center

LumaCogent exists to close a gap that has widened for years. Ambitious, growing e-commerce companies operate with a fraction of the operational intelligence available to the largest players in the market. Amazon and its peers have spent a decade building proprietary AI systems that make their operations faster, smarter, and self-improving. Mid-market companies largely do not. We are building the systems that change that.

We started with reverse logistics — one of the most pressing problems in e-commerce, and one that reaches into almost every function of the business: customer experience, warehouse operations, finance, inventory, and merchandising.

We asked ourselves: how might we make reverse logistics fundamentally better — not incrementally, but from first principles? Recent advances in Agentic Engineering — software that reasons, decides, and learns — reveal an opportunity to do exactly that. The question is not how to automate the existing process. It is whether the entire function can be redesigned into something structurally different. The answer, we found, is yes — and the impact is substantially larger than optimising the current state.

The same systems architecture applies to other operational functions across e-commerce. Returns is where we started.

What follows covers what we found when we mapped the as-is reality, what we built, the architecture underneath it, the methodology that guided every decision, and a working proof of concept.

The Current State

Returns is among the most costly and complex operational challenges in e-commerce.

The business impact is well understood by anyone operating at scale. What is less visible is why these problems persist — and what is actually driving them.

To understand what was driving these outcomes, we mapped the operation end-to-end — every actor, every system, every handoff — across every stage of the return journey.

As-Is Service Blueprint — Current State Returns Operation
Open full blueprint ↗

The mapping revealed 43 distinct issues. These were not edge cases — they were structural, recurring, and compounding. The most impactful:

01Reactive operations — no advance return notice, no volume forecasting
02Fragmented data and siloed systems — no single source of truth
03Human actors running blind — no real-time view or operational intelligence
04Inconsistent grading, refund, and disposition rules
05Unreliable fraud detection — refunds issued before inspection
06Subpar shopper return experience — static portals, no visibility, slow refunds
07True cost of returns unquantified — indirect costs invisible on the P&L
View all 43 pain points 14 Priority 1  ·  22 Priority 2  ·  7 Priority 3
Priority 1 — Direct revenue loss, broken feedback loops, or analytics failure  14 pain points
1Reason code accuracy is structurally low — shoppers select the fastest dropdown option; all downstream analytics are corrupted at the sourceInitiationP1
2Inspection findings never reach Merchandising — damage type, quality defects, and sizing issues are not reported to the Buying team; high-return SKUs are reordered without modificationInspection → MerchandisingP1
3Restocked returns never reach Merchandising — Buying team not notified when returned inventory re-enters stock; cannot inform promotions, markdowns, or buying decisionsInventory → MerchandisingP1
4Restock cycle time is 5–10 business days — popular SKUs are off-market during the window; direct margin loss on every day a sellable item is unavailableInventory ManagementP1
5Returns data fragmented across 3–5 disconnected systems — OMS, WMS, CS platform, and accounting each hold a piece; no single source of truth; reporting assembled manually in Excel, always 5–14 days laggingReportingP1
6WMS has no advance return notice — warehouse receives zero pre-notification of inbound returns; receiving is entirely reactive; peak backlogs stretch processing to 1–2 weeksWarehouse ReceivingP1
7Refund-on-receipt fraud window — most retailers refund before inspection; items in poor condition, wrong items, or empty boxes are discovered only after the refund is already processedRefund ProcessingP1
8Total cost of returns is systematically underestimated — direct costs tracked; indirect costs excluded: lost sale during restock window, liquidation discount vs. full-price recovery, customer churnFinance / ReportingP1
9Operations Supervisor has no real-time visibility across Stages 3–7 — no dashboard for in-transit volume, receiving throughput, inspection queue, disposition split, or restock cycle time; all decisions are reactiveOperationsP1
10Grading standards are informal and unenforced — no documented rubric with reference photographs; grades drift across staff, shifts, and days; borderline items decided individually with no documented criteriaInspectionP1
11No photographic documentation at inspection — disputed grades cannot be reviewed; fraud evidence is lost; damage attribution is subjective and costlyInspectionP1
12Disposition rules are informal and not programmatic — similar items in similar condition receive different dispositions depending on which staff member processes them; liquidation is reactive and poorly timedDispositionP1
13OMS reason code data corrupts all downstream analytics — Shopify Analytics and NetSuite reports give a misleading picture of why items are actually returned; no cycle time tracking, no disposition split, no SKU-level return rateAll downstreamP1
14Returns portal absent for smaller retailers ($5M–$30M ARR) — CS agents manually verify eligibility and create RMAs in Shopify; not scalable; no reason code capture; no WMS notificationInitiationP1
Priority 2 — Operational friction with meaningful downstream cost  22 pain points
15Re-tagging and inventory re-entry is entirely manual — every item individually re-tagged with SKU, price, and condition; scales linearly with return volume; bottleneck at peakInventory ManagementP2
16SKU entry errors create ghost inventory — returned items entered under wrong SKU, size, or colour appear available in OMS but are physically misplaced; error propagates until next inventory countInventory / OMSP2
17WMS → OMS sync is not real-time — receiving confirmation and inventory updates batch-sync; shopper's refund is held even though item has physically arrived; creates CS contacts and cash flow lagReceiving / InventoryP2
183PL data handoff is inconsistent and delayed — batch files from 3PL to merchant are sometimes incomplete; receiving confirmation, condition flag, and RMA match data are unreliableWarehouse ReceivingP2
19RMA matching at receiving is manual and error-prone — missing, damaged, or wrong-item returns require manual investigation; no-RMA returns create further bottleneckWarehouse ReceivingP2
20BNPL refunds are a separate, manual workflow — Klarna, Afterpay, and Affirm refunds not automatically triggered by OMS; require manual action in each provider's portal; high error rate and delayRefund ProcessingP2
21Partial refund calculation has no systematic tool — damage-adjusted refund amounts determined by individual judgment with no documented framework; inconsistent outcomes, no audit trailRefund ProcessingP2
22Chargebacks are under-tracked and costly — rate and cost of bank disputes not tracked separately from returns costs in P&L; no proactive chargeback managementFinanceP2
23Liquidation management is reactive and manual — items accumulate in staging for weeks before vendor contact; batches poorly curated; timing driven by storage pressure not price optimisationDispositionP2
24Shrinkage in staging zones goes undetected — items can be moved or mis-routed without WMS record; discovered only at weekly or monthly inventory reconciliationDispositionP2
25Condition codes are freeform or inconsistently defined — "Like New" means different things across WMS configurations and staff members; condition data is structurally unreliableInspectionP2
26Paper-based inspection workflows eliminate data entirely — at less mature operations, condition data captured on paper and never entered into WMS; inspection outcome data does not exist in any systemInspectionP2
27Portal analytics siloed from back-end operations data — Loop/Narvar dashboards cover initiation and reason code distribution but are not connected to WMS inspection outcomes, OMS refund records, or carrier performanceReportingP2
28Return policy is buried and ambiguous — not surfaced at product page, cart, or checkout; vague language generates pre-emptive CS contacts before return is initiatedPre-initiationP2
29No real-time outstanding refund obligation view — OMS refund data is a daily or weekly rollup; Operations Supervisor cannot model total refund exposure for open returns; cash flow pressure at peakFinance / OperationsP2
30High-value / high-fraud-risk escalation is ad hoc — no WMS flag triggers supervisor review for high-value or high-fraud-risk items; inspector decides unilaterally with no formal protocolInspectionP2
31Damage attribution is subjective — determining customer-caused vs. manufacturing-caused damage requires product expertise floor staff often lack; conservative, costly disposition decisions resultInspectionP2
32Exchange processing is double-handling — an exchange must be processed as a return plus a new order; two OMS records, double labour, poor data integrity, no native tooling at mid-market scaleRefund / OMSP2
33OMS not receiving WMS inspection data — return record stays at "Received" even after inspection completes; refund trigger may be delayed or incorrect depending on merchant's refund policyInspection → OMSP2
34WMS holds most granular returns data with no BI connection — inspection outcomes, disposition splits, and cycle times only accessible via manual CSV export; not connected to any BI tool at mid-market scaleReportingP2
35No self-service refund status tracking for shoppers — shopper cannot check refund status between receipt confirmation and bank posting; every inquiry requires a CS contactPost-refundP2
36Refund timing is unpredictable and uncommunicated — 5–14 business day total processing is common with no proactive status update after receipt confirmationRefund ProcessingP2
Priority 3 — Real issues with low cost magnitude or high complexity relative to gain  7 pain points
37Webhook failure risk on tracking events — if OMS tracking webhook fails, return status stays at "Pending" despite item being in transit; shopper receives no confirmation; CS spikeIn-transitP3
38Accounting system sync is batch — QuickBooks/NetSuite do not receive refund events in real time; Finance reconciliation is always lagging; cost of returns always delayed in P&LFinanceP3
39Multi-carrier tracking is fragmented — merchants using multiple carriers have no unified performance view; transit time, loss rate, and damage rate siloed in individual carrier portalsIn-transitP3
40No benchmarking capability — cannot compare return rate, processing time, or recovery rate against segment peers; no way to assess whether performance is good, average, or poorReportingP3
41Packaging burden on shopper — shopper must source own materials; no packaging provided; friction point that increases return abandonmentReturn ShippingP3
42Label delivery friction — PDF label requires home printer; QR code requires working scanner at drop-off; both create friction and abandonment for less tech-savvy shoppersInitiation / ShippingP3
43USPS scan gaps create phantom in-transit periods — packages accepted at USPS post office may not be scanned for 24–48 hours; no tracking event fires; CS contacts spike; Ops Supervisor has no visibilityIn-transitP3

These were not operational inefficiencies to be optimised. They were structural failures — and they pointed to a different question: not how to make the current process faster, but whether it should be replaced entirely.

The Future State

A fundamentally different kind of operation.

Having mapped the as-is state and defined the problem, we shifted to design. We created a To-Be Service Blueprint that reimagined every actor's role, every system interaction, and every handoff from first principles. Alongside it, we wrote experience scripts — one for each human in the operation — describing what their working day should feel like in the redesigned world. These two artefacts became the specification the architecture was built to deliver.

To-Be Service Blueprint — The Redesigned Returns Operation
Open full blueprint ↗

The shopper opens the brand's website and describes their problem in their own words. No dropdown. No form. The conversation agent identifies the order, classifies the return reason accurately behind the scenes, and schedules a home pickup. Done in under two minutes. For the first time, the reason code that enters the system reflects what actually happened.

Returns Portal — AI Conversation Agent (Interactive)
Open full demo ↗

Inside the operation, the item arrives pre-matched to its record — the warehouse knew it was coming the moment the shopper completed their return. No reactive receiving. No manual RMA matching. The inspector reviews a pre-assembled context package: AI condition assessment, reference photographs, SKU return history, and a fraud signal score. They adjudicate rather than guess, and every decision is photographically documented. A disposition recommendation arrives with full reasoning. The inventory manager approves or overrides — starting from knowledge, not from scratch.

The operations supervisor opens a single view: in-transit volume, inspection queue, disposition split, and restock cycle time — live, from one source. Not assembled from five systems in a spreadsheet. The finance team's reconciliation is pre-assembled; BNPL refunds trigger automatically; chargeback evidence assembles itself.

To show what the redesigned operation actually feels like to work inside, we wireframed the future experience of the Operations Supervisor — the role that, in the current state, spends most of its day assembling data that should arrive automatically.

In the redesigned operation, Diana opens a single surface. She sees live signals ranked by severity, her team's workload at a glance, and a queue of decisions that genuinely need her judgment. Policy proposals from the Learning Agent surface automatically — patterns the system has spotted and structured for her approval. The Cold Start indicator in the header is not cosmetic; it communicates exactly where the system stands in its trust-building phase, and what level of autonomous authority it currently holds. She is no longer a data assembler. The interface is designed around that shift.

Operations Supervisor — Future Experience Wireframe
Open wireframe ↗

The merchandising buyer receives what was previously impossible: structured intelligence from the warehouse — defect patterns by SKU, sizing failure distributions, quality alerts tied to the buying cycle. The feedback loop that has compounded return rates for years is now structural and automatic. High-return SKUs get corrected before the next season's purchase order is cut.

The Transformation, Measured.
Improvements that compound across every dimension of the operation.

None of these improvements is isolated. When reason code accuracy goes from 60% to 90%, the entire downstream picture changes — fraud detection sharpens, merchandising intelligence becomes reliable, and the return rate starts to fall. When restock cycle time halves, sellable inventory returns to market faster, margin recovery accelerates, and the operation handles the same volume with fewer hands. These are not efficiency gains on individual tasks. They are structural changes that reinforce each other.

MetricToday — typical mid-marketRedesigned Operation
Shopper return experienceStatic portal, 10+ minutesConversational, <3 minutes
Restock cycle time5–10 business days<48 hours
Inspection throughput~50 items / person / day (manual)~120+ items / person / day (AI-assisted)
Disposition decision timeHours to days (manual, per-item)Minutes (recommended with reasoning)
Reason code accuracy~40–60% (dropdown selection)90%+ (NLP-derived from conversation)
Cost visibility lag5–14 days (CSV / Excel)Real-time (unified dashboard)
Merchandising feedback loopBroken (manual, ad-hoc)Structural and automatic
True cost of returns visibilityPartial (direct costs only)Fully loaded

Modeled projections based on architectural design and industry benchmarks. Not observed production data.

Role Transformation
The redesign did not eliminate roles. It transformed them.

The Human Actor Transformation Model shows what changes for each of the six roles in the operation — from where they spend their time today, to the role the redesigned system creates for them. Volume moves to the system. Judgment stays human. What each person actually does becomes meaningfully different — and, by design, more valuable.

Human Actor Transformation Model

Six roles, transformed. The system handles volume; humans handle judgment.

RoleTodayTransformedWhat the System Handles
Operations SupervisorData assembler — CSV exports, Excel reports, manual issue identificationStrategic supervisor — reviews prepared alerts, approves decisionsAll data aggregation, all SLA monitoring, all routine reporting, all inbound forecasting
Returns InspectorManual inspector — every item, manual grade, informal criteriaExpert adjudicator — reviews AI recommendations, confirms or overridesAll grading of high-confidence items; all SKU history lookup; all fraud signal matching
Inventory ManagementDisposition decision-maker from scratch — no system supportRecommendation approver — reviews AI reasoning, overrides when context differsAll disposition recommendations with reasoning; all vendor communication on approval
Finance TeamManual reconciler — BNPL portals, chargeback spreadsheets, lagging dataException reviewer — two-minute exception queue, pre-assembled packagesAll BNPL routing; all chargeback evidence assembly; all accounting sync
Merchandising / BuyingData chaser — emails Operations for return rate data that rarely arrivesIntelligence consumer — structured, buying-cycle-aware intelligence automatically routedAll returns-to-merchandising routing; all defect threshold alerting
Warehouse ReceivingManual operator — scans every package, matches RMAs by handException handler — monitors terminal, handles exception lane onlyAll pre-matched receiving; all conveyor routing for in-scope returns
"We didn't build a returns tool. We rebuilt the operation — and the same approach reimagines every function in e-commerce."
How We Built It

We started with the human experience. Then designed the machine around it.

The methodology determined what problems to solve and for whom. The design principles governed what we optimised for in every decision. The architecture principles governed how we built the system that delivered it.

Methodology — double diamond
Diamond 1 · Leg 1
Discover
As-Is Service Blueprint — 43 pain points mapped across every actor, system, and handoff.
Diamond 1 · Leg 2
Define
Prioritisation — narrowed to 14 Priority 1. Root causes identified before any solution is proposed.
Diamond 2 · Leg 1
Develop
Architecture designed to address the 14 highest-leverage problems. Experience scripts written per actor before architecture is drawn.
Diamond 2 · Leg 2
Deliver
Validation against each actor's experience script. Where the design fails the human, the architecture changes.
Design principles — what we were designing for

These principles governed every product and experience decision throughout the design process. They are not aspirations — each one was tested against the experience scripts written for each human actor in the operation. Where the design failed the principle, the design changed.

Relief through conversation
Every interaction — for shoppers and internal actors — feels like talking to a knowledgeable human.
Humans decide, the system does
Orchestration is automated. All judgment stays human. The system never acts without a clear authority boundary.
Proactive intelligence, routed right
Right information, right role, right format — before it is asked for.
Predicted truth, not reported opinion
AI derives ground truth from evidence. The shopper's stated reason is one input, not the answer.
Full cost visibility, always current
Fully-loaded cost of returns — direct and indirect — is the headline metric. Updated continuously, not weekly.
Closed feedback loops, human-approved
The system routes intelligence to the people who can act on it. Humans approve. Loops close automatically.
Architecture principles — how we built it

These principles governed structural decisions — what it means to be AI-native rather than AI-decorated. Each principle has a direct implication for how the system learns, scales, and remains trustworthy as autonomous authority expands over time.

Separate knowing, deciding, and doing
Six independent layers — each upgrades without touching the others. The system never has to be rebuilt to be made smarter.
Bind authority at the code level
Agents reason freely; they can only act within an enforced tool registry. Structural, not a guideline.
Self-learning by design
Every agent decision and human override is training data, captured automatically in full context.
Self-healing by runbook
The system attempts remediation before escalating. Resolved novel failures are added to the runbook automatically.
MVP is a clean subset of the North Star
Every component is designed to evolve — no throwaway work, no migration cost between phases.
The Architecture

A six-layer AI-native system.

The returns function runs on six independent layers. Each has a single responsibility. Each upgrades without touching the others. The system improves continuously without being rebuilt.

Most enterprise AI deployments add models on top of unchanged processes. The model reasons; the underlying system does not. Our architecture is structurally different. Knowing, deciding, and doing are separated into six independent layers — each with a single responsibility, each upgradeable without the others knowing. The agents never query databases directly. All context is assembled by a purpose-built service before any reasoning begins. Every decision is logged to memory — and every log is training data. The result: a system that learns from every activation, expands its autonomous authority as it earns trust, and does not require a rebuild to become smarter. This is what AI-native means in practice.

Layer 5 — Human Interface
Where the system meets its people. Proactive role-aware assistants, the shopper-facing conversation agent, and the governance dashboard.
Layer 4 — Capability Layer
Deterministic execution. RMAs, labels, inspections, refunds, inventory sync, notifications. No reasoning — just reliable, idempotent action on every agent instruction.
Layer 3 — Agent Mesh
The reasoning layer. Ten agents at full scale — four at MVP, six more across Phase 2 and Phase 3. Each perceives events, reasons, and acts within a strictly enforced authority boundary.
Layer 2 — Memory Architecture
Institutional memory. Every agent decision and human override is logged in full context. The system assembles the richest possible context for every activation.
Layer 1 — Knowledge Core
The operational intelligence substrate. SKU return patterns, fraud signals, merchant policy, demand signals — continuously updated from every event in the system.
Layer 0 — Data Fabric
The foundation. An ordered, replayable event backbone that everything else runs on. Every return event — initiation to resolution — flows through it.
Full Architecture — Six Layers
LumaCogent Returns Suite — System Architecture
View interactive diagram ↗
LumaCogent Returns Suite — six-layer system architecture
Six-layer architecture — the full system. Each layer is independently upgradable; internals change, interfaces do not.
MVP Architecture — Phase 1 Build
LumaCogent Returns Suite — MVP Architecture
LumaCogent Returns Suite — MVP architecture
MVP architecture — four agents, full Closed-Loop SLA, and the learning loop operational from day one.
The Agent Mesh — ten agents at full scale

Layer 3 is where the system thinks. Each agent perceives relevant events, assembles a context package via the Context Assembly Service, reasons to a structured output, and acts within a strictly enforced tool registry. Every decision is logged to episodic memory — training data for the next activation. The mesh starts with four agents at MVP and expands to ten as the system earns trust and accumulates episodic memory.

01MVPOrchestration AgentOwns the RMA lifecycle end-to-end. Monitors every open return, detects stalls before a human notices, and advances the process without being asked. Autonomous within configured thresholds; human approval for policy exceptions and high-value refunds.
02MVPDisposition AgentMulti-signal reasoning across condition grade, SKU demand signal, days in staging, and liquidation pricing — recommending the highest-value outcome for each returned item. Advisory at MVP; autonomous below configured thresholds at Phase 2.
03MVPInspection Guidance AgentAssembles the full context package for the inspector — AI condition assessment, reference photographs, SKU history, fraud signal score — so they adjudicate rather than guess. All grades confirmed by inspector at MVP; Phase 2 moves high-confidence items to autonomous.
04MVPOps AgentMonitors system health continuously and runs a structured runbook of known failure modes before involving a human. Resolved novel failures are added to the runbook automatically — the escalation rate decreases over time.
05Phase 2Anomaly Detection AgentContinuous pattern monitoring across merchants, SKUs, and return cohorts. Surfaces coordinated fraud signals, seasonal anomalies, and SKU-level return rate spikes before they compound — routed to the Operations Supervisor with full context.
06Phase 2Merchant Insight AgentProactive intelligence delivery to each human actor's assistant — the right signal, for the right role, at the right moment. Buying-cycle-aware alerts to Merchandising; refund obligation summaries to Finance; SLA risk warnings to Operations.
07Phase 2Exception AgentHandles return cases that fall outside the designed flows — items with no RMA match, policy edge cases, multi-carrier failures, or returns arriving at unregistered facilities. Determines the safest resolution path before escalating.
08Phase 2Learning AgentStructures episodic memory signals for model retraining. Identifies systematic patterns in human overrides, flags drift between agent recommendations and human outcomes, and surfaces policy proposals to the Operations Supervisor for approval.
09Phase 3Calibration AgentSurfaces systematic override patterns across the full merchant population. Triggers targeted model retraining when drift crosses defined thresholds. The primary mechanism through which the mesh improves at scale rather than per-merchant.
10Phase 3Explainability AgentProduces human-readable rationale for any agent decision — including known unknowns and confidence gaps. Enables merchants to audit the system's reasoning on demand, and supports regulatory and compliance use cases as autonomous authority expands.
It runs.

A working two-agent prototype demonstrates the inspection and resolution pipeline on real return records — the same agent design, the same authority boundaries, the same reasoning pattern as the production architecture.

Agent 1
Inspection Agent
Receives a return record, reviews item details and stated reason, and produces a structured assessment — condition grade, defect flags, fraud risk score, and a recommended disposition — with full reasoning.
Agent 2
Resolution Agent
Receives the Inspection Agent's output alongside disposition rules and inventory context. Reasons to a final recommendation — restock, resell-discounted, liquidate, or discard — with a confidence score and rationale ready for human review.
Two-Agent Pipeline — Inspection & Resolution (Interactive)
Open full demo ↗
What Else This Builds

The pattern applies wherever decisions, knowledge, and execution are bundled into manual roles.

The architecture we built for returns is domain-agnostic. The same six layers, the same agent mesh, the same self-learning loop — applied to other operational functions in e-commerce.

Function 01
Customer Support
  • Automated context assembly before the agent acts
  • First-contact resolution optimisation
  • Escalation routing and prioritisation
  • Proactive issue detection and outreach
  • Self-service resolution for repeating patterns
  • Agent decision support in real time
Function 02
Accounting & Finance
  • Automated real-time reconciliation
  • BNPL routing — Klarna, Afterpay, Affirm
  • Chargeback evidence assembly and response
  • Fully-loaded cost visibility
  • Fraud loss quantification
  • Cash flow forecasting from open returns
Function 03
Marketing Operations
  • Dynamic customer segmentation
  • Churn prediction and intervention
  • ROAS optimisation
  • Post-purchase sequence automation
  • Retention trigger firing
  • Customer lifetime value modelling
Function 04
Inventory Optimisation
  • Returns-adjusted demand forecasting
  • Reorder accuracy optimisation
  • Defect-driven supplier quality management
  • Overstock prevention
  • Markdown and liquidation timing
  • Real-time restocked inventory visibility

What we built is not a returns company. It's a way of reimagining e-commerce companies — one operational function at a time.

An Honest Closing

We are looking for the right first operators.

LumaCogent is an early-stage company. The architecture described in this piece is designed and specified. The methodology is proven through the work documented here. We are now building the first production version of this system, and we are in early conversations with a small number of mid-market apparel retailers who want to be part of that build.

We are not pitching a finished product. We are looking for merchants who understand that their returns problem is both a cost problem and a data problem, and are ready to approach it that way.

If that's your situation, we'd like to talk.

Start the conversation →

No deck. No proposal. A direct conversation about your operation.