Proactive outreach for unpaid orders

#Offer help before an unpaid order is abandoned

Proactive outreach for unpaid orders contacts a buyer when an order remains unpaid after creation, using the timing and message content you configure. It is designed to help buyers who encountered checkout or payment trouble or still have a product question, then move any reply into the normal Customer Service workflow.

This is not a broadcast tool. The system continues to check order state and buyer engagement. It stops later reminders after payment, cancellation, or buyer engagement. A no-reachout control and configuration A/B tests let you separate natural payment from incremental impact created by outreach.

#Open the configuration

Open Shops, select the Shop, then choose AI Customer Service → Unpaid order reachout. The configuration has two parts:

1. Follow-up schedule: whether outreach runs, when it sends, and what each message says;
2. Evaluation & optimization: how much traffic remains in a no-reachout control and whether complete strategies are compared.

AI Customer Service must be enabled and the Shop must remain authorized. Scheduled outreach runs in the backend, so the settings page does not need to stay open. Keep the Shop’s responsible Desktop online to handle buyer replies when they enter Customer Service.

#Configure the follow-up schedule

After enabling Reach out to unpaid orders, configure up to three stages. Each stage includes:

• Enabled state: temporarily disable a stage without deleting it;
• Minutes after order creation: a whole number from 1–2879; enabled stages cannot use the same delay;
• Reminder template: an empty template uses the localized default for the Shop region.

Stages are ordered by delay. Enabling an empty configuration creates one default stage. The current default is 24 hours after order creation for Mexico and three minutes for other supported regions. Review that timing for the Shop’s market, buyer decision cycle, and messaging requirements; a default is not automatically the best strategy for every Shop.

Templates support three placeholders:

Placeholder	Replaced at send time with
{{order_id}}	Order ID
{{product_count}}	Number of items in the order
{{shop_name}}	Shop name

The first message should help with payment or product questions without creating false urgency. Do not promise an unconfigured discount, stock level, or delivery date. A second or third stage should add help or context instead of repeating the same sentence.

#How one order moves through outreach

When the platform reports a new unpaid order, the system records the outreach plan active at that time and calculates each stage from the order’s creation time. Later configuration changes do not rewrite the plan snapshot already assigned to that order.

Before every send, the system checks that:

• AI Customer Service and unpaid-order outreach are still enabled;
• the order is still unpaid;
• the order has not been cancelled;
• the buyer has not engaged after an earlier outreach;
• the buyer can be identified and a Customer Service conversation can be opened;
• the order is not assigned to the no-reachout control.

Payment or cancellation stops all pending stages. If the buyer replies after the first outreach, later automated reminders stop and the normal Customer Service conversation continues. If several stages have already become due, the system skips an older superseded stage instead of sending several reminders in quick succession.

#Why “associated paid” is not incremental payment

Open Customer Service → Performance → Unpaid reachout to review eligible orders, reached orders, associated paid orders, associated units, and associated GMV by date and stage.

“Associated paid” means that payment occurred within the attribution window after outreach. Some buyers would have returned and paid without a message, so post-reachout payment is useful for monitoring delivery and the funnel but does not prove that outreach created additional payment.

To answer whether outreach truly adds value, retain a randomized no-reachout control.

#Experiment layer one: the no-reachout control

When Keep a control group is enabled, every new eligible order receives a stable order-level random assignment:

• Control: remains part of the eligible population but receives no outreach message;
• Treatment: follows the current production outreach plan.

The holdout can be 1%–20% and defaults to 5%. One order stays in one group; it does not move between groups during the experiment. Because control and treatment run concurrently under the same market and traffic conditions, their payment difference within the fixed outcome window can support an incremental-impact comparison.

The primary incrementality metric is payment rate within the outcome window. GMV and units per assigned order are secondary metrics. In the current version, outcome maturity is measured from order creation: 48 hours for Mexico and 60 minutes for other supported regions. An order that has not completed that window remains immature and should not enter a final comparison.

Changing the holdout percentage, enabling or disabling evaluation, or changing the production stages while evaluation is active starts a new experiment version so different configurations are not mixed.

#Experiment layer two: compare A/B/C outreach plans

The control answers “should we send?” A configuration A/B test answers “which outreach plan should we use?” After outreach and evaluation are enabled and saved, select Configure A/B test:

1. Plan A is the current production configuration. Its share of Treatment traffic can change, but its stages cannot be edited inside the experiment;
2. copy A or add B, C, and other plans with different delays, stage counts, or messages; one experiment can contain 2–20 plans;
3. every plan needs at least one enabled stage, and complete configurations must be distinct;
4. percentages are calculated within Treatment traffic, must total 100%, and each plan must receive at least 1%;
5. save a server draft before launching when more review is needed.

For example, a 5% holdout leaves 95% for the configuration experiment. If A and B each receive 50% of Treatment traffic, total traffic is approximately 5% no outreach, 47.5% A, and 47.5% B.

While a configuration experiment is running, the Shop’s base stages are locked so the production baseline cannot change mid-test. Create a new version or stop the current experiment when the plans need to change.

#Why the experiment layers stay separate

When both layers run, incrementality compares only:

• no-reachout Control;
• current production configuration A.

Exploratory B and C plans are excluded from the primary incrementality comparison and analyzed only in the configuration test. This prevents several different messages and delays from being blended into one Treatment arm, which would make it impossible to answer whether the production plan itself adds value.

Configuration experiments compare:

• payment rate within the outcome window;
• GMV per assigned order;
• units per assigned order;
• payment latency, where lower is better;
• message send failure rate as a guardrail. The current non-inferiority margin is one percentage point.

For an interpretable result, change one primary factor at a time—for example, the first message or the first-send delay. If copy, timing, and stage count all change together, a winning plan will not reveal which change produced the result.

#Review results in Experiment analysis

Open Customer Service → Experiments. The page separates Realtime and History:

• Realtime includes running experiments and stopped experiments whose order outcomes are still maturing;
• History contains experiments whose outcome window has finished and can be read as final.

Check experiment quality first:

• Assigned: orders entered into the experiment;
• Matured: orders that completed the full outcome window;
• Allocation quality: whether observed assignment materially differs from the configured split;
• Started and data-as-of times: the business period represented by the result.

Stopping an experiment prevents new orders from entering. It does not instantly finalize orders already assigned. Existing cohorts must finish the outcome window, so a stopped experiment first appears as ended and maturing before it becomes final.

#Read the payment-progress chart

The payment-progress chart begins at minute one after order creation. It shows how payment changes over time for each arm and marks each plan’s outreach times.

Before the first message is sent, the curves mainly represent natural payment behavior and cannot evaluate outreach impact. Differences become meaningful only after orders reach an outreach time and continue through observation.

The page offers a modeled signal and raw steps. Segments with fewer than 100 assigned orders, a 95% interval wider than ten percentage points, or observation coverage below 80% are visually de-emphasized. Treat these as directional signals rather than final evidence.

#Read statistical comparisons

For each metric, the page reports observed values, relative difference, a 95% confidence interval, and a p-value.

Use this order:

1. Check maturity: incomplete outcome windows should not enter a final payment-rate comparison;
2. check allocation quality: investigate a material mismatch before interpreting outcomes;
3. check effect size: decide whether the difference is large enough to matter operationally;
4. read the uncertainty range: a wide interval means the true effect remains uncertain;
5. check the guardrail: a higher payment rate does not justify a materially worse send-failure rate;
6. make the business decision: include message quality, complaint risk, and Shop policy.

Do not stop because a live curve is temporarily ahead, and do not repeatedly edit variants during a run. Peeking and stopping on the current leader increases the chance of mistaking random variation for a real effect.

#Stop an experiment and adopt a plan

Stop after the planned observation period or when a business-risk guardrail requires immediate action. Wait until the status becomes Final, then review the completed result in History.

For a stopped configuration experiment, select a plan and use Apply plan to write its stages and messages back to the Shop’s base configuration. Adoption is a business decision, not a payment-rate-only decision; review delivery failures, message quality, buyer feedback, and operating risk as well.

#Recommended first launch

1. Start with one stage and verify placeholders, tone, and timing.
2. Keep a 5% control and first establish whether outreach adds value over natural payment.
3. After execution and data look healthy, create an A/B test that changes one factor.
4. Monitor allocation, maturity, and send failures in Realtime without declaring a winner early.
5. When the experiment is final, record the hypothesis, configurations, dates, sample, and decision before applying the winning plan.

The goal is not to send more messages. It is to help buyers who genuinely encountered a payment or product problem without over-contacting everyone—and to verify the incremental result with a control.