Key takeaways
- Order exceptions include payment holds, out of stock items, address failures, split shipments, delayed fulfillment, return and refund issues, pricing mismatches and wholesale order errors.
- Automation can classify the exception, gather order, stock and customer context from the ecommerce platform, OMS, ERP and payment processor, propose a fix and draft the customer message.
- Refunds above a limit, fraud decisions, goodwill credits and anything that changes a customer's money or terms should stay with a named person.
- Returns are a large enough share of retail volume that the exception queue behind them deserves its own measurement, not just a general resolution time.
- Start with one high volume exception type, such as an out of stock item or a routine payment hold, before expanding to more complex cases.
An order exception is the moment a purchase stops moving through the normal path: the payment gets flagged, the item is gone from the shelf, the address will not validate, or a promotion code prices the order wrong. In a small operation, someone just looks at the order and fixes it. Past a certain volume, exceptions pile up in a queue nobody owns, customers find out their order has a problem before your team does, and the same categories of issue get triaged from scratch every time. The fix is not a rulebook everyone forgets to follow. It is making sure each exception reaches the right person with the order, stock and customer context already gathered.
What counts as an order exception in retail?
Order exceptions cluster into a short list of repeat categories, across a direct to consumer storefront and a wholesale or B2B channel alike.
- Payment holds and fraud review flags, where a processor or a fraud tool stops an order before it ships.
- Out of stock after purchase, when the ecommerce platform sold an item the warehouse says it does not have.
- Address validation failures, where the shipping address cannot be matched to a deliverable location.
- Split shipments, where one order has to leave from more than one location or on more than one date.
- Delayed fulfillment, where an order sits past its committed ship date.
- Returns and refund exceptions, such as a return outside the stated window, a missing item, or a refund that does not match what was paid.
- Pricing or promotion mismatches, where a discount code, a bundle or a price change did not apply the way the customer expected.
- Wholesale or B2B order errors, such as a purchase order quantity or price that does not match the agreed terms.
Ecommerce platforms formalize some of this themselves. Shopify, for example, tracks a fulfillment status of on hold separately from unfulfilled, specifically for orders flagged for fraud risk, short on inventory, or over a cost or quantity threshold the merchant sets, and keeps them out of the normal fulfillment queue until the hold is released (Shopify). The same platform separates order status (open, archived or canceled) from fulfillment status (unfulfilled, in progress, partially fulfilled or fulfilled), a distinction worth building automation rules around (Shopify).
Which systems handle a retail order exception?
An order exception usually needs facts from more systems than the person resolving it has open at once.
| System | Role in an order exception | Typical data it holds |
|---|---|---|
| Ecommerce platform (for example Shopify or Adobe Commerce) | Where the order is placed and where fulfillment status, including holds, is tracked | Order detail, fulfillment status, customer account |
| Order management system (OMS) | Coordinates sourcing and splitting an order across warehouses or stores | Inventory allocation, ship from location, split shipment logic |
| Enterprise resource planning (ERP) such as NetSuite | Holds the financial side: invoicing, credit terms, wholesale pricing | Order and invoice detail, customer terms, credit limits |
| Warehouse management system (WMS) | Confirms what stock actually exists and what shipped | Pick and pack detail, on hand quantity, backorder status |
| Payment processor | Flags fraud risk and manages the hold on funds | Authorization status, fraud score, chargeback history |
| Help desk (for example Zendesk or Gorgias) | Where the customer finds out and where your team resolves the case | Ticket history, customer contact, prior resolutions |
| CRM | Customer history and value, useful for deciding how to resolve a case | Purchase history, loyalty tier, service history |
An exception that looks small, a payment hold on a normal looking order, still needs the payment processor's fraud score, the customer's order history in the CRM or ecommerce platform, and sometimes the ERP's credit terms before anyone can decide whether to release it.
What can automation and AI actually do with an order exception?
The realistic scope for a retail order exception, like a logistics one, is everything before the decision.
A workflow can watch the ecommerce platform, the payment processor and the OMS for the signals that create an exception: an order held for fraud review, an item that sold out after the order was placed, an address the platform could not validate. It classifies the exception, then checks current stock across locations, the customer's order and account history, and any promotion rule involved, so the full picture exists before anyone opens the case. For routine exceptions it can propose the fix: substitute a comparable item, split the shipment, place a backorder with a realistic date, or calculate what a return actually owes. It drafts the customer message in plain language, referencing the real order rather than a generic apology, then routes the case to whoever owns that exception type, a fraud analyst, a fulfillment coordinator or a service lead, for them to approve, adjust or reject.
None of this requires the system to decide whose money is at risk. It requires it to prepare a decision a person can make quickly.
What should stay with your team?
Four categories belong with a person, not a workflow, because they change what a customer is owed or trusted with.
Refunds above a limit. Any refund past a threshold your finance team sets should route to a person, regardless of how confident the workflow is in the amount.
Fraud decisions. Declining or holding a real customer's order over a false fraud signal costs a sale and, often, the relationship, so a person should confirm before an order is cancelled outright.
Goodwill credits. Discretionary credit for a bad experience is a judgment call about the relationship, not a calculation.
Anything that changes a customer's money or terms. A price adjustment, a wholesale contract term, or a payment plan change needs the same approval a manual change would get.
| Exception type | What automation can do | What stays with a person | Data it needs |
|---|---|---|---|
| Payment hold or fraud flag | Classify the flag reason, check order and customer history for context | Deciding to release, cancel or manually review the order | Fraud score, order value, customer history |
| Out of stock after purchase | Detect the mismatch, check other locations, propose a substitute or backorder date | Approving a substitute that changes price, or a backorder past a promised date | Inventory by location, substitute options, promised ship date |
| Address validation failure | Flag the mismatch, suggest a correction from past orders or address data | Confirming a correction that changes the delivery commitment | Order address, validation error, customer's past addresses |
| Split shipment | Determine the sourcing split and draft the customer notice | Approving extra shipping cost the split creates | Inventory by location, shipping cost, order value |
| Delayed fulfillment | Detect the missed date, draft an update with a new estimate | Approving a discount or credit offered for the delay | Committed ship date, actual status, customer tier |
| Return or refund exception | Match the return to the order, calculate what is owed | Approving a refund above the standard policy or limit | Return reason, order and payment detail, policy terms |
| Pricing or promotion mismatch | Detect the mismatch against the promotion rule, calculate the correct price | Deciding whether to honor the price the customer expected | Promotion rule, order total, customer communication |
| Wholesale or B2B order error | Compare the order to agreed price and quantity terms, flag the variance | Deciding how to resolve a variance with an account | Contract terms, purchase order, account history |
Catalog checks, verifying supplier product data before it goes live, are a related workflow worth mentioning: a wrong weight, price or dimension often surfaces later as an order exception. Store replenishment is the other workflow most retail operations tackle alongside these two.
Illustrative example: Consider a direct to consumer brand selling through its own ecommerce platform and a wholesale channel. A payment processor flags an order for fraud review because the shipping and billing addresses are in different states, a pattern that is often just a gift order. A workflow checks the customer's order history, finds three prior orders shipped to the same address with no chargebacks, and drafts a note recommending release. A fraud analyst agrees and releases the order in under a minute instead of researching the account from scratch. The same week, an item sells out on twelve orders between checkout and fulfillment. The workflow proposes the closest in stock substitute for each and drafts a customer message, and the fulfillment coordinator approves ten as drafted and adjusts two where the substitute was not close enough.
How do you know it is working?
Four measures matter more than a general resolution time, because they describe what changes for the customer and the business.
Time to first action shows how long an exception sits before anyone, human or workflow, does something with it, usually the easiest time saving available today.
Exceptions resolved without a customer contact counts how many issues get fixed, or the customer gets told, before they have to reach out, which is the clearest sign the workflow caught it early enough.
Order accuracy after resolution tracks whether the fix, a substitute, a split or a refund, actually matched what the customer expected, not just whether a case was closed.
Refund and credit leakage measures how much is given away in refunds, credits or discounts against what your policy actually allows, since an inconsistent process usually costs more than a slow one.
Returns deserve their own attention: retailers project total returns will reach 849.9 billion dollars in 2025, with 19.3 percent of online sales expected to come back (National Retail Federation). Online sales made up 17.1 percent of total US retail sales in the second quarter of 2026, up 12.2 percent year over year (US Census Bureau), so exceptions tied to online orders are a growing share of the total, not a shrinking one.
How do you start with one exception type?
- Pick the highest volume, clearest exception type, usually an out of stock item or a routine payment hold rather than a wholesale contract dispute.
- Map exactly where the signal appears today, the ecommerce platform, the payment processor or the help desk, and agree which system is authoritative when they disagree.
- Agree what a proposed fix needs to include before the fulfillment or service team will trust it, with the person who owns that decision today.
- Set the approval routing and a deadline, naming the role each exception type lands with rather than a shared queue.
- Run it alongside the current process for a defined period and compare it against the baseline.
- Add the next exception type only once the first is running with minimal correction.
Which business processes to automate first applies the same scoring approach to any exception type you are considering, and AI document processing for invoices, contracts and forms covers reading the purchase orders and packing slips a wholesale exception often depends on.
How Kastling approaches retail order exception automation
Kastling treats order exception handling as one workflow to understand before anything is built, as part of the AI Integration and Automation service. That starts with a discovery call, and for retail operations a separately scoped paid audit is common next, reviewing the ecommerce platform, OMS, ERP and payment processor involved, and where exceptions are actually costing time today.
Order context gets gathered and routed to the service team the way the underlying work already runs, rather than replaced by a new queue nobody trusts. Refunds above a limit, fraud decisions and goodwill credits stay with a named person, and we agree with the business owner up front how success will be measured, whether that is time to first action, resolution without a customer contact, or refund and credit leakage. Retail covers where order exceptions fit alongside catalog checks and store replenishment.
Human approval planner
Describe one AI task and see the level of human control it needs, with a control pattern your team can set up.
Questions
What is the difference between an order exception and a return?
A return is a customer choosing to send an item back, usually within your stated policy. An order exception is anything that stops an order moving through its normal path before that point, such as a payment hold, an out of stock item or an address that will not validate. Returns can create their own exceptions, for example when the refund does not match what was paid, but the two are not the same problem.
How do we automate fraud review without declining good customers?
Give the workflow the customer’s order and account history alongside the fraud signal, not just the signal on its own, so a flagged order with a clean history is recommended for release rather than treated the same as a first time order with no history. Keep the actual decision to cancel or hold an order with a person, since the cost of losing a real customer is usually higher than the cost of a short review.
Which order exception should we automate first?
Start with the exception type that happens most often and has the clearest resolution path, which for most retailers is an item selling out after checkout or a routine payment hold. Wholesale contract errors and complex refund disputes are usually better candidates once the team trusts how the simpler cases are handled.
Can small retailers use this, or is it only for large operations?
The approach works at any volume, though the payback shows up sooner once exceptions are frequent enough that a person cannot review each one individually. A smaller retailer often gets more value from automating the data gathering, pulling stock, order and customer history into one place, than from full automation of the decision.
How much should stay automated versus reviewed by a person?
That depends on the exception type and how much money or trust is at stake. As a rule, anything that changes what a customer is owed or that could end a customer relationship, a fraud decline, a refund above a limit, a goodwill credit, should go to a person, while gathering context and drafting the first response is where automation carries the most weight.