Returns Automation in 2026: AI Triage That Saves Margin

Returns automation in 2026 is no longer a portal and a prepaid label — it is an AI triage layer that decides, for every returned unit, whether the best-margin outcome is a returnless refund, an auto-approved restock, an exchange, a refurb channel, liquidation, or a fraud hold with a human in the loop. With US retail returns at an estimated $849.9 billion in 2025 (NRF / Happy Returns) and processing costs that multiple sources place between a fifth and two-thirds of item value, that routing decision is where the margin lives.

This is the third post in our returns series, and it deliberately covers the layer the other two do not. Our returns-reduction playbook covers prevention — stopping avoidable returns before they happen with PDP, sizing, and root-cause fixes. Our exchange-experience playbook covers retention — turning an inevitable return into a kept customer and preserved revenue. This post is the operations layer: what to do, mechanically and economically, with every return that happens anyway.

We cover the 2026 cost math (three methodologies, kept separate), a disposition decision tree with human-review gates marked, the fraud typology mapped to the AI countermeasure that catches each pattern, the vendor stack as it actually stands after the Returnly shutdown, and a buy-vs-build framing for teams deciding whether to license a platform or wire their own triage logic.

The 2025 numbers, from the NRF / Happy Returns 2025 Retail Returns Landscape report (published October 2025), frame everything else in this playbook. US retail returns totalled an estimated $849.9 billion — a 15.8% return rate across all retail sales, with the online channel running higher at 19.3% of online sales. That is down from $890 billion and a 16.9% rate in 2024 — the first year-over-year dollar decline since 2020 — but the structural picture has not reversed: overall return rates roughly doubled between 2019 and 2024, from about 8.1% to 16.9% of retail sales, a shift both NRF and Shopify attribute to ecommerce growth and free-return-driven bracketing rather than any temporary spike.

Our read of the trend: the dollar decline is not the industry shrinking its returns problem — it is retailers finally managing it. Return rates settling slightly below the 2024 peak while absolute volume stays near historic highs means the winners are not the merchants with the fewest returns; they are the merchants who process each return for the least cost and recover the most value from it. That is an operations capability, and in 2026 it is increasingly an AI capability. NRF’s Katherine Cullen, VP of Industry and Consumer Insights, framed the same shift from the experience side: “Returns are no longer the end point of a transaction. They provide an opportunity for retailers to create positive customer experiences.”

Cost-of-return figures get blended constantly in returns content, and blending them is how bad budgets get built. Three widely cited sources measure three different things. Shopify Enterprise (July 30, 2026) cites a per-item processing-cost range of 20–65% of an item’s original value. Appriss Retail’s 2026 Total Retail Loss Benchmark report puts the industry-wide average near 30% of item value, and totals US processing costs at $211 billion for 2025. Optoro — a returns-processing vendor, so treat its figures as vendor-stated — benchmarks a single return at 20–39% of the item’s original cost and adds that returns take roughly twice the labor of outbound fulfillment.

The table below applies each methodology to one worked example — a $60 apparel item — without merging them. Every dollar figure is computed from the source’s own stated percentage against the $60 base.

Used together — and only as a converging range, each with its own attribution — the three sources point the same direction: processing a return eats roughly a quarter to a third of the item’s value in the typical case, and can consume most of it at the high end of Shopify’s range. On a $60 apparel item, the quarter-to-a-third case is $15 to $20 gone before you have decided what to do with the garment; Shopify’s 20–65% range on its own stretches that to $12 at the low end and $39 at the high end. Which is precisely why the disposition decision — the subject of the next section — is worth automating: every branch of the tree has a different cost profile, and routing even a modest share of volume to a cheaper branch moves real margin.

Disposition is the decision returns-AI actually automates: for each returned unit, which outcome recovers the most value net of cost? The clearest named example in market is Optoro’s SmartDisposition engine, which routes every unit to its “next-best home” — restock (available-to-sell), a resale or liquidation channel, donation, or destruction — based on item condition, category, and recovery economics. Optoro cites 47% of retail executives naming slow time-to-restock a top returns pain point (vendor-stated), which is the practical case for automating this call rather than queueing it behind a warehouse inspector.

At the other end of the tree sits the zero-effort branch: Shopify’s returnless refund — repay the shopper, skip the shipment entirely — recommended for low-value items where return shipping and handling would cost more than the retailer would recover by getting the item back. In between sit exchange-first (the revenue-retention branch — pair it with a post-purchase upsell flow and the refund often becomes a bigger order), refurb or open-box resale, and liquidation or donation. The matrix below assembles the full tree — with the human-review gate column that vendor marketing tends to leave out.

Two things make this tree work in production. First, the exchange-first branch is a revenue decision, not a logistics one — Shopify’s own app description notes Loop “uses your return rules to approve or deny requests automatically” while steering shoppers toward an exchange before a refund. Second, the gated branches are gated on purpose: every scenario in the bottom half carries either fraud risk or enough dollar value that a wrong automated call costs more than the reviewer’s time.

NRF classifies 9% of all returns as fraudulent in its 2025 data. Keep that figure separate from the attitude numbers the same report carries: 45% of shoppers surveyed say it is acceptable to “bend the rules” when returning an item they are unhappy with, and 39% of Gen Z shoppers surveyed admit to having returned a fraudulent decoy item instead of the one they kept. The 9% measures confirmed fraud share; the others measure self-reported attitudes — they are different denominators and different phenomena.

The growth is concentrated in specific schemes. Among retailers who track these categories, NRF / Happy Returns report the following year-over-year increases in 2025:

Alongside outright fraud sit the two consumer behaviors Shopify explicitly names as cost drivers in its July 30, 2026 guide: wardrobing (buying an item for temporary use — an outfit for one event — then returning it) and bracketing (buying multiple sizes or colors intending to keep one and return the rest). Each pattern has a distinct tell, and each tell already has a named AI countermeasure in market:

• Wardrobing — tell: wear signs, removed tags. Countermeasure: condition rules at intake (Shopify’s stated defense is to decline clothing returned without tags) plus camera-based item audits. Happy Returns’ Return Vision is an AI camera-plus-software system that audits items for incorrect logos, altered tags, material mismatches, and product swaps before they reach the warehouse.
• Bracketing and serial abuse — tell: multi-SKU purchase patterns and return cadence. Countermeasure: behavioral intelligence. Happy Returns’ Risk Behavior Scoring models return frequency, timing, geography, and history; Loop’s named Fraud Detection feature likewise flags abusive behavior and high-risk returns before approval (vendor-described).
• Empty-box and false-tracking scams — tell: weight or content mismatch at first scan. Countermeasure: in-person verification. Happy Returns states its Return Bar drop-off network cuts fraud by at least 85% versus mail-in returns — a vendor-stated figure from its internal data, with no independent replication we could find, but the mechanism is plain: empty boxes get caught at first physical touch, before entering the reverse-logistics stream.

Returns fraud is one lane of a wider problem — if you are building the full defense, our layered fraud-detection stack playbook covers the payment and chargeback side that this post’s returns-specific screening plugs into.

The returns-automation market consolidated meaningfully over the last three years, and a lot of published content has not caught up. Here is the stack as it actually stands in July 2026, with each vendor’s named AI capability.

Vendor marketing sells autonomous decisioning. The honest operator framing is narrower: full autonomy for low-stakes, low-value branches — returnless refunds, auto-restock of tagged, unworn items — and mandatory human sign-off for anything crossing a fraud-risk or dollar-value threshold. That is not our invention; it is the pattern Shopify itself recommends for fraud triage via Shopify Flow: apply thresholds on order value or item count, auto-tag the shopper for review, notify customer service by email or Slack, route the case to manual approval before any refund, and exclude flagged accounts from free-shipping or full-refund offers. Rules flag; humans approve. It is the credible middle path between fully manual (where 65% of Shopify-merchant refunds still sit) and fully autonomous (where a wrong fraud call becomes a chargeback, a lost customer, or both).

The remaining question is whether to buy that triage layer or build it. The same logic we laid out in the buy-vs-build decision for agentic tooling applies here, with one returns-specific twist: the highest-value AI components — behavioral fraud scoring across hundreds of brands, camera-based item auditing, a 10,000-location physical verification network — depend on pooled data and physical infrastructure a single merchant cannot replicate. The routing logic between those signals, though, is exactly the kind of thin orchestration layer a small team can own.

Where agents fit: the 2026-credible version of “agentic returns” is an agent that drafts the disposition decision with evidence attached — behavior score, condition grade, item value, policy citation — and executes it only on the autonomous branches, queueing everything gated for a one-click human approve or override. Designing that division of labor is the core of our AI transformation engagements: the goal is not removing people from returns, it is removing returns from people — except the four gated rows in the disposition table where judgment pays for itself.

The sequencing below assumes a Shopify-stack merchant with meaningful return volume and no dedicated returns platform — the situation the 65%-manual figure says most merchants are in.

Days 1–30: instrument and codify

Pull ninety days of returns and compute your own cost-per-return using the table in Section 02 — your number, not the industry average, decides your thresholds. Codify existing policy into Shopify Return Rules (eligibility windows, restocking fees, return-shipping terms) so auto-approval has something to approve against. Set the returnless-refund threshold: any item where reverse shipping plus inspection exceeds expected recovery.

Days 31–60: gate before you automate

Build the Flow-style fraud gates first — order-value and item-count thresholds that tag, notify, and route to manual approval — so that when auto-approval turns on, the risky tail is already fenced off. Then enable auto-approval for the clean branch: unworn, tagged, in-window, clean history. Track override rate: if reviewers overturn the automation more than occasionally, the thresholds are wrong.

Days 61–90: add disposition and decide the platform

Introduce disposition routing beyond restock — open-box resale and liquidation channels for graded returns — and run the buy-vs-build matrix from Section 06 against your fraud exposure and exchange share. This is also the point to evaluate a drop-off network if empty-box claims show up in your data. If you want a second set of hands on the margin math or the platform evaluation, this is exactly the shape of our ecommerce engagements.

Looking forward: NRF released an updated Retail Fraud Taxonomy — a shared classification framework developed with Target and the Chertoff Group — on July 14, 2026. Standardized fraud categories are what returns-triage models train against, so expect the vendor stacks above to converge on common risk labels over the next few cycles. Merchants who tag their returns data against a consistent taxonomy now will get more out of every model they buy or build later; the ones who keep free-text return reasons will keep absorbing the fraudulent share of their returns without ever being able to see the pattern.