For online retailers who have shipped an order they should not have

Fraud Screening and Order Holds: Stopping Suspect Orders Before They Ship

Most ecommerce platforms and payment gateways now score every order for fraud risk. The score is the easy part. The failure happens downstream: the order has already been exported to the ERP, dropped into the warehouse pick list or sent to the 3PL before anyone looks at the score, and by the time someone decides to cancel it the parcel is on a truck. A fraud tool that is not connected to fulfilment is a reporting tool.

This page is about the integration around fraud screening rather than the scoring itself. When orders should be released to fulfilment, how a hold reaches every system that could act on the order, how reviewers get the context to decide quickly, what cancelling actually has to undo, and how to assemble the evidence that wins a chargeback dispute weeks or months after the sale. It also covers the other cost of fraud controls: good customers held, delayed or declined.

Review Your Order Hold Flow (03) 9003 0111

Realistic ROI

0 orders

Released before a decision

The design target: no suspect order reaches picking until it is approved

Up to ~4 months

For a chargeback to arrive

Dispute windows vary by card scheme and reason, so evidence must be kept that long or longer

3 outcomes

Approve, cancel or verify

Every held order needs one, recorded with who decided and why

Same day

Review target for held orders

A practical goal so held customers still get the delivery they paid for

Four Gaps Between a Fraud Score and a Stopped Order

Each gap is a timing or data problem between systems. None of them is solved by buying a better scoring tool.

The order leaves before the score lands

Fraud scores are often calculated asynchronously, seconds or minutes after the order is created. Many ERP and 3PL connectors pick up new orders as soon as they appear. If the export does not wait for the risk decision, high risk orders reach the warehouse in the same batch as everything else. The fix is a release gate: orders are only exported to fulfilment once they have a decision, with low risk orders released automatically within moments.

A hold in one system is not a hold everywhere

Tagging an order as held in the ecommerce platform does nothing if the ERP already has it, or if the 3PL received it an hour ago. A real hold has to be written to every system that could act on the order, and the integration has to know whether it arrived in time. If the order is already picked or dispatched, someone needs to be told immediately so it can be intercepted with the carrier where that is possible.

Reviewers decide without context

A reviewer looking at a fraud score in isolation sees one transaction. The context that makes most decisions easy lives elsewhere: whether this customer has ordered before, returned before, used this address before, contacted support, or matches a loyalty member who shops in store. Bringing that history into the review screen is what turns a ten minute investigation into a thirty second decision.

Chargeback evidence is scattered

When a disputed transaction arrives, winning it depends on evidence from several systems: the order and payment authentication details, delivery tracking and proof of delivery from the carrier, the customer’s account history, and any communications. Assembled manually, it often takes long enough that the response deadline passes. Assembled automatically, it is ready the day the dispute lands.

How a Connected Fraud Flow Works

Six pieces, from the moment an order is placed to the moment a dispute is resolved. Your existing scoring tool stays where it is.

No unscored order exported

Release gate

New orders are held in the ecommerce platform until a risk decision exists. Low risk orders are released to the ERP, warehouse or 3PL automatically, usually within seconds, so the gate adds no visible delay for most customers. High and medium risk orders, and any order the scoring service failed to score, stay in a held state and go to review rather than defaulting to release.

Held in every system

Hold propagation

When an order is placed on hold after release, the hold is written to the ERP sales order, the warehouse or 3PL order and any drop ship supplier order, and each system confirms it. If any system reports the order as already picked, packed or dispatched, an urgent alert goes to the person who can intercept it, with the carrier and tracking reference attached.

Fast, consistent decisions

Review queue with context

Held orders appear in a single queue with the score and reasons, plus customer history pulled from your other systems: order count and value, previous returns and disputes, prior addresses, loyalty status and recent support contact. Reviewers record a reason with every decision so the next reviewer, and the rules, can learn from it.

Every system updated

Clean approve, cancel or verify

Approve releases the order downstream in one action. Cancel voids or refunds the payment, cancels the order everywhere, releases reserved stock and records the reason, without leaving a phantom sales order in the ERP. Verify sends the customer a request, such as confirming details or presenting identification at click and collect, and keeps the order held until they respond or a deadline passes.

Dispute response ready

Evidence pack

For every shipped order the integration keeps the facts a dispute needs: authentication result, IP and device signals your tools capture, billing and delivery details, tracking events and proof of delivery, and communications. When a chargeback notification arrives, those facts are assembled into a draft response in the format your gateway or acquirer expects, for a person to check and submit.

Rules that improve

Outcome feedback

Chargeback outcomes, confirmed fraud, and orders cancelled that turned out to be genuine are fed back against the original score and decision. Over a few months that shows which rules catch fraud, which only catch good customers, and where the thresholds should move. Most scoring tools can use this feedback directly if it is sent back to them.

Common Situations, Before and After

Task	Traditional	Connected Properly	Notes
High risk order at 9pm	Picked overnight by the 3PL	Held until reviewed	The release gate means the 3PL never received it, so there is nothing to intercept.
Hold placed after export	Warehouse not told	Hold written everywhere, confirmed	If the order already shipped, the alert arrives in time to try a carrier intercept.
Long standing customer flagged	Declined or delayed	Approved in seconds with history	False declines are a real cost. Context protects good customers as much as it catches bad ones.
Suspicious click and collect order	Collected by anyone with the email	ID check required at pickup	The store sees the verification requirement on the pickup screen, not in an email nobody reads.
Cancelled fraud order	Ghost order left in the ERP	Cancelled and stock released everywhere	Also keeps sales reporting and stock availability honest.
Chargeback notice arrives	Evidence gathered by hand	Draft response assembled	A person still reviews and submits. The time saved is in gathering, not judging.
Card testing attack	Noticed when declines spike	Alert on unusual small orders	Gateways have their own protections. An alert on order patterns adds an early warning.
Threshold tuning	Changed after a bad week	Changed on recorded outcomes	Needs a few months of outcomes linked back to decisions before it is reliable.

What to Get Right

Declining good customers

Every fraud control has a false positive rate, and the lost sales from blocking or delaying genuine customers can exceed the fraud prevented, though the balance varies widely by business. Measure both. A review step for medium risk orders, rather than automatic cancellation, is usually the better default for retailers whose customers buy repeatedly.

Holding money without a decision

Orders cannot sit in review indefinitely. Set a deadline after which a held order is either approved or cancelled and refunded, and tell the customer what is happening. Your terms should say you may cancel orders for suspected fraud, and refunds for cancelled orders should be prompt. Australian Consumer Law also applies to how you describe delivery times to held customers.

Treating fraud signals as free data

IP addresses, device information, identity checks and fraud scores about an identifiable person are personal information under the Privacy Act 1988 where it applies. Your privacy policy should cover the collection and any disclosure to fraud service providers, access to review data should be limited, and retention should match your dispute and legal needs rather than forever.

Moving card data through the integration

The integration should never handle full card numbers. Work with tokens, last four digits and the authentication results your gateway provides, so your systems stay outside the most demanding parts of PCI DSS scope. If a design needs raw card data to work, the design is wrong.

Blunt rules that discriminate

Blocking whole postcodes, regions or name patterns is easy to configure and catches large numbers of genuine customers, often in ways that look unfair. Prefer signals about the transaction and its history over signals about who someone appears to be, and review rule outcomes for patterns you would not want to defend.

Ignoring the release gate during outages

When the scoring service is slow or down, the easiest fix is to release everything. Decide in advance what happens: perhaps release low value orders from known customers and hold the rest. Write it down and build it, because the decision made at 2am during an outage is rarely the right one.

How Yes AI Approaches Fraud Integration

Map where orders can escape

We trace every route an order takes from checkout to dispatch, including drop ship suppliers, marketplaces and click and collect, and show exactly where a suspect order can get past the current controls.

Keep the scoring you already have

Built in fraud analysis from your platform or gateway is often good enough. If it is, we connect it properly rather than sell you another tool. If it is not, we will say so and help you compare options on evidence.

Build the gate and the queue

Release gate, hold propagation, review queue with customer context and clean cancellation, running on a managed cloud automation layer we operate with alerting on anything that fails.

Privacy and card data handled properly

No raw card data in the integration, fraud signals stored with limited access and defined retention, and privacy policy wording checked against what is actually collected and shared.

From Score Only to Stopped Orders

Five steps. A release gate and hold propagation for a single store are usually live in four to six weeks.

Trace every order path

Checkout to ERP, warehouse, 3PL, drop ship suppliers and stores, with the timing of each hand off and where the fraud decision currently lands.

Agree decision rules

Which risk levels release automatically, which go to review, what verify means for click and collect and delivery, review deadlines, and outage behaviour.

Build gate and propagation

Orders held until decided, holds written to every downstream system with confirmation, and alerts when an order has already moved too far.

Add the queue and the evidence

Review queue with customer history, one step approve and cancel, and evidence retained for every shipped order ready for disputes.

Close the loop

Outcomes linked back to decisions and rules, with a monthly review of false positives, confirmed fraud and dispute win rates.

FAQ

Is the fraud protection built into our platform enough?

For many Australian online retailers the built in fraud analysis from the ecommerce platform or payment gateway is a reasonable starting point. The more common weakness is not the score but what happens after it: orders reaching fulfilment before anyone acts on the score. Fix the release gate and the hold propagation first. Then judge whether the scoring itself is the problem, using recorded outcomes rather than impressions.

Will holding orders for review slow down delivery?

For most orders, no. Low risk orders are released automatically within seconds of being scored. Only the held minority waits, and a same day review target keeps the delay to hours rather than days. Tell held customers honestly if their order needs verification, and set a deadline after which a held order is decided one way or the other.

How long after a sale can a chargeback arrive?

Dispute windows depend on the card scheme and the reason for the dispute, and are commonly up to around four months from the transaction or expected delivery date, sometimes longer. Keep order, authentication and delivery evidence for at least that long, and in practice longer because your financial record keeping obligations will usually require it anyway.

What evidence helps win a chargeback?

It depends on the dispute reason, but typically: the authentication result for the payment, proof that the goods were delivered to the address given, including tracking events and any signature or photo, evidence the customer has a history of genuine orders with the same details, and records of any communication. Your gateway or acquirer specifies the format and deadline. The integration’s job is to have it ready rather than gather it under pressure.

What has to happen when we cancel a fraudulent order?

More than most people expect. The payment is voided or refunded, the order is cancelled in the ecommerce platform, the ERP, the warehouse or 3PL and any drop ship supplier, reserved stock is released, and the reason is recorded. If any of those steps is missed you get phantom orders in reporting, stock that cannot be sold, or a parcel shipped after the refund.

Does fraud screening raise privacy issues?

It can. Device data, IP addresses, identity checks and risk scores relating to an identifiable person are personal information where the Privacy Act 1988 applies to you. Make sure your privacy policy describes the collection and any sharing with fraud service providers, limit who can see review data, and keep it only as long as you need it for disputes and legal obligations.

How long does this take to set up?

A release gate and hold propagation for a single store with one fulfilment path is typically live in four to six weeks. Multiple warehouses, 3PLs, drop ship suppliers and stores add time because each is another place an order can escape. The review queue and evidence pack can follow in a second phase once holds are working reliably.

Make Sure a Held Order Actually Stops

Book a call. We trace how orders move from checkout to dispatch, show where a suspect order can slip through today, and give you a priced plan to close the gaps. The findings are yours either way.

Book a Fraud Flow Review (03) 9003 0111

All discussions held in confidence. Australian-based consultants.