---
title: "AI Returns Management: Automate Ecommerce Returns and Refunds"
description: Returns eat nearly 16% of retail sales. See how AI approves, routes, and converts returns into exchanges without adding headcount.
image: https://images.unsplash.com/photo-1449247666642-264389f5f5b1?w=1200&q=80
---
[.png)](https://www.getdarwin.ai/en?utm_source=organic)
Industries
[ Automotive](https://www.getdarwin.ai/en/industrias/automotriz?utm_source=organic) [ Education](https://www.getdarwin.ai/en/industrias/educacion?utm_source=organic) [ Real State](https://www.getdarwin.ai/en/industrias/inmobiliario?utm_source=organic) [ Retail](https://www.getdarwin.ai/en/industrias/retail?utm_source=organic) [ Health](https://www.getdarwin.ai/en/industrias/salud?utm_source=organic) [ Insurance](https://www.getdarwin.ai/en/industrias/seguros?utm_source=organic) [ Services](https://www.getdarwin.ai/en/industrias/servicios?utm_source=organic)
IA Workforce
[ Alba, Inbound](https://www.getdarwin.ai/en/worker/alba?utm_source=organic) [ Bruno, Outbound](https://www.getdarwin.ai/en/worker/bruno?utm_source=organic) [ Eva, CX](https://www.getdarwin.ai/en/worker/eva?utm_source=organic) [ Sophia, Post-Sales](https://www.getdarwin.ai/en/worker/sophia?utm_source=organic) [ Lucas, Cobranzas](https://www.getdarwin.ai/en/worker/rio?utm_source=organic)
[About Us](https://www.getdarwin.ai/en/about-us?utm_source=organic) [Partners](https://www.getdarwin.ai/en/partners?utm_source=organic)
Resources
[ Blog](https://blog.getdarwin.ai/en/) [ Press](https://www.getdarwin.ai/en/press?utm_source=organic) [ ROI Calculator](https://www.getdarwin.ai/en/roi-calculator?utm_source=organic) [ Whatsapp Link Generator](https://www.getdarwin.ai/en/whatsapp-link-generator?utm_source=organic) [ Whatsapp QR Generator](https://c2w.getdarwin.ai/en)
EN
[ES](https://blog.getdarwin.ai/es/?hsLang=en) [EN](https://blog.getdarwin.ai/en/) [PT](https://blog.getdarwin.ai/pt-br/?hsLang=en)
[Login](https://app.getdarwin.ai/login) [Try Darwin ](https://app.getdarwin.ai/signup?landing=homepage&language=en&utm_source=organic_homepage&utm_medium=LeadGen)
Industries
[ Industries](https://www.getdarwin.ai/en/industrias/automotriz?utm_source=organic) [ Education](https://www.getdarwin.ai/en/industrias/educacion?utm_source=organic) [ Real State](https://www.getdarwin.ai/en/industrias/inmobiliario?utm_source=organic) [ Retail](https://www.getdarwin.ai/en/industrias/retail?utm_source=organic) [ Health](https://www.getdarwin.ai/en/industrias/salud?utm_source=organic) [ Insurance](https://www.getdarwin.ai/en/industrias/seguros?utm_source=organic) [ Services](https://www.getdarwin.ai/en/industrias/servicios?utm_source=organic)
IA Workforce
[ Alba, Inbound](https://www.getdarwin.ai/en/worker/alba?utm_source=organic) [ Bruno, Outbound](https://www.getdarwin.ai/en/worker/bruno?utm_source=organic) [ Eva, CX](https://www.getdarwin.ai/en/worker/eva?utm_source=organic) [ Sophia, Post-Sales](https://www.getdarwin.ai/en/worker/sophia?utm_source=organic) [ Lucas, Collections](https://www.getdarwin.ai/en/worker/rio?utm_source=organic)
[About Us](https://www.getdarwin.ai/en/about-us?utm_source=organic) [Partners](https://www.getdarwin.ai/en/partners?utm_source=organic)
Resources
[ Blog](https://blog.getdarwin.ai/en/) [ Press](https://www.getdarwin.ai/en/press?utm_source=organic) [ ROI Calculator](https://www.getdarwin.ai/en/roi-calculator?utm_source=organic) [ Whatsapp Link Generator](https://www.getdarwin.ai/en/whatsapp-link-generator?utm_source=organic) [ Whatsapp QR Generator](https://c2w.getdarwin.ai/en)
EN
[ES](https://blog.getdarwin.ai/es/?hsLang=en) [EN](https://blog.getdarwin.ai/en/) [PT](https://blog.getdarwin.ai/pt-br/?hsLang=en)
[Login](https://app.getdarwin.ai/login) [Try Darwin ](https://app.getdarwin.ai/signup?landing=homepage&language=en&utm_source=organic_homepage&utm_medium=LeadGen)
# AI Returns Management: Automate Ecommerce Returns and Refunds
22 de July, 2026
*Last updated: July 21, 2026*
Every ecommerce sale carries a hidden second transaction: the possibility of a return. Most teams obsess over winning the first transaction and improvise the second one with a shared inbox, a policy PDF, and a support agent copy-pasting refund instructions at 11 p.m. That improvisation is expensive. According to the [National Retail Federation, retailers expected 15.8% of annual sales, roughly $849.9 billion in merchandise, to come back in 2025](https://nrf.com/media-center/press-releases/consumers-expected-to-return-nearly-850-billion-in-merchandise-in-2025), with online purchases returned at an even higher 19.3% rate. AI returns management turns that chaotic second transaction into a governed, automated workflow: instant eligibility checks, refund-or-exchange decisions made in seconds, fraud screening, and proactive status updates that stop tickets before they start. This guide explains what the technology actually does, how to roll it out in five steps, and which metrics prove it is paying for itself.
## Table of contents
- [Why returns quietly drain ecommerce margins](https://blog.getdarwin.ai/en/ai-returns-management-automate-ecommerce-returns#why-returns-hurt)
- [What AI returns management actually does](https://blog.getdarwin.ai/en/ai-returns-management-automate-ecommerce-returns#what-is-ai-returns-management)
- [A five-step playbook for automating returns](https://blog.getdarwin.ai/en/ai-returns-management-automate-ecommerce-returns#playbook)
- [The metrics that tell you it is working](https://blog.getdarwin.ai/en/ai-returns-management-automate-ecommerce-returns#metrics)
- [Four mistakes to avoid](https://blog.getdarwin.ai/en/ai-returns-management-automate-ecommerce-returns#mistakes)
- [Frequently asked questions](https://blog.getdarwin.ai/en/ai-returns-management-automate-ecommerce-returns#faq)
## Why returns quietly drain ecommerce margins
A return is never just a refund. It is a reverse shipping label, a warehouse inspection, a restocking decision, a payment reversal, and, almost always, a string of customer messages asking where the money is. Each of those steps has a cost attached, and because they sit across logistics, finance, and support, nobody owns the total. The [NRF found that the top reasons retailers now charge for returns are rising processing costs (cited by 40% of retailers) and rising carrier shipping costs (also 40%)](https://nrf.com/media-center/press-releases/consumers-expected-to-return-nearly-850-billion-in-merchandise-in-2025). In other words, the operational burden has grown large enough that merchants are willing to risk customer goodwill just to offset it.
Fraud makes the picture worse. The same [NRF research estimates that 9% of all returns are fraudulent, and reports that 85% of retailers are already deploying AI to detect and prevent return fraud](https://nrf.com/media-center/press-releases/consumers-expected-to-return-nearly-850-billion-in-merchandise-in-2025): empty boxes, counterfeit swaps, overstated quantities. Manual review teams cannot inspect every claim, so they either approve too much and eat the losses or approve too slowly and infuriate honest customers.
Then there is the support load. Every return generates its own micro-lifecycle of anxiety: did you receive it, was it approved, when will I be refunded. These "where is my refund" tickets are the sibling of the [WISMO tickets that already dominate ecommerce support queues](https://blog.getdarwin.ai/en/wismo-automate-where-is-my-order-ai?hsLang=en), and they arrive precisely when the customer relationship is most fragile. A customer whose return goes smoothly often buys again; one who has to chase a refund for two weeks usually does not. As [Shopify's guidance on returns management](https://www.shopify.com/enterprise/blog/ecommerce-returns) argues, the returns experience has become part of the product itself.
## What AI returns management actually does
AI returns management is the use of conversational AI agents and machine-learning decision engines to run the entire post-purchase return journey: intake, eligibility, disposition, communication, and analysis. It replaces the form-plus-inbox model with a system that resolves most cases end to end. Four capabilities matter most.
### Instant eligibility and policy enforcement
An AI agent reads the order, the purchase date, the item category, and your policy, then tells the customer in seconds whether the return qualifies and what their options are. No queue, no "please allow 2 to 3 business days for a response." Because the agent applies the written policy identically every time, you also eliminate the quiet inconsistency of tired agents approving out-of-policy returns just to close tickets.
### Smart disposition: refund, exchange, or keep it
Not every return should travel back to the warehouse. A decision engine can weigh item value, resale probability, and shipping cost to choose the cheapest good outcome: a standard refund-on-receipt, an instant exchange, store credit with a bonus, or a "keep it" refund for low-value items where reverse logistics costs more than the product. These decisions are exactly the kind of repetitive, rule-plus-judgment calls that AI handles better and faster than a swamped operations team, a point [Forbes has highlighted in its reporting on AI-driven reverse logistics](https://www.forbes.com/sites/dennismitzner/2025/02/11/ai-is-transforming-e-commerce-returns-cutting-costs-and-improving-efficiency/).
### Proactive, conversation-first status updates
The best returns systems talk first. The moment a label is scanned, the package is received, or the refund is issued, the customer gets a message on the channel where they already talk to you, whether that is WhatsApp, email, or web chat. This is the same principle behind [proactive customer support](https://blog.getdarwin.ai/en/proactive-customer-support-ai?hsLang=en): every update you push is a ticket that never gets opened.
### Fraud and abuse screening
Machine-learning models score each return against patterns like serial returning, bracketing, receipt anomalies, and mismatched item weights, flagging only the suspicious minority for human review. Honest customers get instant approvals; the review team concentrates on the cases that deserve scrutiny.
## A five-step playbook for automating returns
### Step 1: Codify the policy as executable rules
Most return policies live as prose full of ambiguity. Rewrite yours as explicit rules: eligibility windows by category, condition requirements, who pays shipping, refund method and timing, exceptions for final-sale and hygiene items. If two experienced agents would decide a case differently, the rule is not specific enough for automation yet.
### Step 2: Deploy a conversational returns agent on the channels customers already use
A returns portal buried in your footer helps the customers who find it. A conversational agent meets everyone else. Shoppers who already [message brands on chat and social channels for support](https://blog.getdarwin.ai/en/ai-chatbots-scale-ecommerce-customer-support-2026?hsLang=en) expect to start a return the same way, in their own words, at any hour. This is where purpose-built AI employees earn their keep: Darwin AI's [Eva, an AI customer experience agent](https://www.getdarwin.ai/en/worker/eva), handles return requests over WhatsApp and other messaging channels, checks the order, applies the policy, issues the label, and keeps the customer informed through to the refund, escalating to a human only when a case falls outside the rules.
### Step 3: Automate disposition decisions
Start conservative: auto-approve the clearly eligible, auto-decline the clearly ineligible, and route the gray zone to people. As the decision engine accumulates outcomes, widen the automated band. Track the override rate, and treat every human override as a rule you have not written down yet.
### Step 4: Save the sale with exchange-first offers
A refund ends a relationship transaction; an exchange continues it. Configure the agent to lead with the exchange or store-credit option when the return reason is size, color, or preference, and to make switching effortless by showing the alternative item in the same conversation. The same conversational infrastructure that [recovers abandoned carts on WhatsApp](https://blog.getdarwin.ai/en/whatsapp-abandoned-cart-recovery-ai-playbook?hsLang=en) can recover a sale that is walking out the door.
### Step 5: Close the loop with product and merchandising
Return reasons are the most honest product feedback you will ever collect. Have the AI tag and cluster them, then feed the clusters to merchandising weekly. A size-runs-small tag on one SKU is an anecdote; the same tag on 4% of a category is a product-page fix that prevents the next thousand returns.
**Key takeaway:** Automating the return is only half the win. The other half is converting refunds into exchanges and converting return reasons into catalog fixes, which is where AI returns management moves from cost reduction to revenue protection.
## The metrics that tell you it is working
Do not measure the program on refund volume alone; measure the full economics of the second transaction.
| Metric | What it tells you | Healthy direction |
| --- | --- | --- |
| Cost per return processed | Total reverse-logistics and labor cost divided by returns handled | Falling |
| Refund cycle time | Hours from request to money back | Falling |
| Exchange rate vs. refund rate | Share of returns saved as exchanges or credit | Rising |
| Return-related ticket share | Portion of support volume caused by return status questions | Falling |
| Repeat purchase after return | Whether a smooth return preserves the customer | Rising |
| Fraud catch precision | Flagged cases that reviewers confirm as abuse | Rising |
The exchange-versus-refund split deserves special attention. It is the clearest signal that the system is not just processing returns more cheaply but actively keeping revenue in the business.
## Four mistakes to avoid
**Automating a policy you have not fixed.** If the underlying policy is confusing or hostile, AI will only enforce the confusion faster. Fix the policy first, then automate it.
**Hiding the human exit.** Customers tolerate automation when escape is easy. Bury the escalation path and every edge case becomes a public complaint.
**Treating fraud screening as a blanket slowdown.** Punishing all customers with delays because a minority abuses the policy destroys the loyalty the program is meant to protect. Score cases individually and keep the honest path instant.
**Ignoring the data exhaust.** If return reasons are not flowing to product, merchandising, and marketing teams every week, you are running an expensive logistics program and skipping the free intelligence program that comes with it.
## Frequently asked questions
### What is AI returns management?
It is the use of conversational AI agents and machine-learning decision engines to automate the ecommerce returns process end to end: intake, eligibility checks, refund or exchange decisions, fraud screening, customer communication, and return-reason analysis.
### Does automating returns increase fraud risk?
Done properly, it lowers it. The [NRF estimates 9% of returns are fraudulent and reports that 85% of retailers already use AI to detect return fraud](https://nrf.com/media-center/press-releases/consumers-expected-to-return-nearly-850-billion-in-merchandise-in-2025), because models can screen every single case for abuse patterns, something no manual review team can do.
### Will customers accept an AI agent handling their return?
Customers care about speed, clarity, and getting their money back, not about who processes the request. An agent that resolves a return in one conversation at midnight beats a human reply that arrives two days later. The key is a clearly available human escalation path for complex cases.
### Where should a small team start?
Start with the highest-volume, lowest-ambiguity slice: standard-window returns of full-priced items in resalable condition. Automate that band completely, measure override rates for a few weeks, then expand coverage category by category.
**Turn returns from a cost center into a retention engine.** Eva, Darwin AI's customer experience agent, handles returns, refunds, and status updates on WhatsApp around the clock.
[Meet Eva](https://www.getdarwin.ai/en/worker/eva)
Compartir [](https://twitter.com/intent/tweet?url=&text=) [mailto:?subject=&body=](mailto:?subject=&body=)
[](https://www.getdarwin.ai/)
## Blog posts
[View All ](https://blog.getdarwin.ai/?hsLang=en)
[](https://blog.getdarwin.ai/en/support-capacity-planning-demand-spikes?hsLang=en)
[Support Capacity Planning: Handle Demand Spikes Without Hiring](https://blog.getdarwin.ai/en/support-capacity-planning-demand-spikes?hsLang=en)
Peak season doubles your ticket volume, not your headcount. The forecasting and automation math that keeps CSAT steady when demand spikes.
Lautaro Schiaffino • 13 Aug 2026
[](https://blog.getdarwin.ai/en/speed-to-lead-response-time-b2b?hsLang=en)
[Speed to Lead: Why Response Time Decides Who Wins the Deal](https://blog.getdarwin.ai/en/speed-to-lead-response-time-b2b?hsLang=en)
Most B2B teams reply to inbound leads in hours. Here is what that costs, how to measure it honestly, and how to close the gap without hiring…
Lautaro Schiaffino • 10 Aug 2026
[](https://blog.getdarwin.ai/en/average-handle-time-reduce-aht-without-hurting-csat?hsLang=en)
[Average Handle Time: How to Cut AHT Without Hurting CSAT](https://blog.getdarwin.ai/en/average-handle-time-reduce-aht-without-hurting-csat?hsLang=en)
Most AHT programs shave seconds and lose customers. Here is how to find the minutes hiding in wrap-up, hold and context - without rushing a …
Lautaro Schiaffino • 07 Aug 2026
[](https://blog.getdarwin.ai/en/customer-effort-score-ces-measure-reduce-ai?hsLang=en)
[Customer Effort Score: How to Measure CES and Reduce It with AI](https://blog.getdarwin.ai/en/customer-effort-score-ces-measure-reduce-ai?hsLang=en)
The metric that catches what CSAT misses: how to measure customer effort, read it against containment, and cut steps out of every resolution…
Lautaro Schiaffino • 06 Aug 2026
[](https://blog.getdarwin.ai/en/ai-credit-management-approve-b2b-orders?hsLang=en)
[AI Credit Management: Approve More B2B Orders, Write Off Less](https://blog.getdarwin.ai/en/ai-credit-management-approve-b2b-orders?hsLang=en)
Late payments hit a huge share of B2B receivables. See how AI credit scoring decides who gets terms, and how much, before the invoice exists…
Lautaro Schiaffino • 05 Aug 2026
[](https://blog.getdarwin.ai/en/how-to-reduce-dso-b2b-playbook?hsLang=en)
[How to Reduce DSO: The B2B Playbook for Faster Cash](https://blog.getdarwin.ai/en/how-to-reduce-dso-b2b-playbook?hsLang=en)
Global B2B DSO runs ~45 days on Net 30 terms. Here are the levers that actually close that gap, ranked by how many days each one removes.
Lautaro Schiaffino • 04 Aug 2026
[](https://blog.getdarwin.ai/en/cost-per-resolution-support-metric?hsLang=en)
[Cost Per Resolution: The Support Metric That Exposes Real Cost](https://blog.getdarwin.ai/en/cost-per-resolution-support-metric?hsLang=en)
Your cost-per-ticket number is hiding a 2.3x multiplier. Here is how to find the real cost of every problem your support team solves.
Lautaro Schiaffino • 03 Aug 2026
[](https://blog.getdarwin.ai/en/containment-rate-ai-support-metric?hsLang=en)
[Containment Rate: The Metric That Shows If Your AI Support Works](https://blog.getdarwin.ai/en/containment-rate-ai-support-metric?hsLang=en)
Most teams calculate containment rate wrong and celebrate the result. Here are real benchmarks, the three signals that expose fake containme…
Lautaro Schiaffino • 31 Jul 2026
[](https://blog.getdarwin.ai/en/sales-to-customer-success-handoff-b2b-playbook?hsLang=en)
[Sales to Customer Success Handoff: The B2B Playbook That Sticks](https://blog.getdarwin.ai/en/sales-to-customer-success-handoff-b2b-playbook?hsLang=en)
The handoff is where deal context goes to die. The brief, the five steps and the four metrics that keep renewals from quietly slipping away.
Lautaro Schiaffino • 30 Jul 2026
[](https://blog.getdarwin.ai/en/ai-order-taking-whatsapp-email-orders-erp?hsLang=en)
[AI Order Taking: Turn WhatsApp and Email Orders into ERP Data](https://blog.getdarwin.ai/en/ai-order-taking-whatsapp-email-orders-erp?hsLang=en)
Your buyers order by chat and email. Here is what it really costs to retype those orders, and how AI captures them clean the first time.
Lautaro Schiaffino • 29 Jul 2026
[](https://blog.getdarwin.ai/en/ai-chargeback-management-prevent-win-disputes?hsLang=en)
[AI Chargeback Management: How to Prevent and Win More Disputes](https://blog.getdarwin.ai/en/ai-chargeback-management-prevent-win-disputes?hsLang=en)
Chargebacks cost far more than the refund. See how AI prevents disputes, automates evidence, and recovers revenue most merchants write off.
Lautaro Schiaffino • 27 Jul 2026
[](https://blog.getdarwin.ai/en/automated-payment-reminders-invoices-paid?hsLang=en)
[Automated Payment Reminders: Get Invoices Paid Without the Chase](https://blog.getdarwin.ai/en/automated-payment-reminders-invoices-paid?hsLang=en)
59% of businesses carry overdue invoices. See how automated, conversational reminders collect the money before you ever have to chase.
Lautaro Schiaffino • 24 Jul 2026
.png)
[Try Darwin Now](https://app.getdarwin.ai/signup?landing=homepage&language=en&utm_source=organic_homepage&utm_medium=LeadGen)
#### SiteMap
- [Home](https://blog.getdarwin.ai/?hsLang=en)
- [Partners](https://www.getdarwin.ai/en/partners?utm_source=organic)
-
- [Industries](https://www.getdarwin.ai/en/industrias?utm_source=organic)
- [API](https://api.getdarwin.ai/)
- [Jobs](https://apply.workable.com/darwin-ai/)
#### Industries
- [Automotive](https://www.getdarwin.ai/en/industrias/automotriz?utm_source=organic)
- [Education](https://www.getdarwin.ai/en/industrias/educacion?utm_source=organic)
- [Real Estate](https://www.getdarwin.ai/en/industrias/inmobiliario?utm_source=organic)
- [Retail](https://www.getdarwin.ai/en/industrias/retail?utm_source=organic)
- [Health](https://www.getdarwin.ai/en/industrias/salud?utm_source=organic)
- [Insurance](https://www.getdarwin.ai/en/industrias/seguros?utm_source=organic)
- [Services](https://www.getdarwin.ai/en/industrias/servicios?utm_source=organic)
#### Recursos
- [Blog](https://blog.getdarwin.ai/en)
- [ROI Calculator](https://www.getdarwin.ai/en/roi-calculator?utm_source=organic)
- [WhatsApp Link Generator](https://www.getdarwin.ai/en/whatsapp-link-generator?utm_source=organic)
- [WhatsApp QR Generator](https://c2w.getdarwin.ai/en)
- [Help Center](https://help.getdarwin.ai/en/)
#### Legal
- [Privacy policies](https://www.getdarwin.ai/en/politicas-de-privacidad?utm_source=organic)
- [Terms of Use](https://www.getdarwin.ai/en/tyc?utm_source=organic)
© 2025, Papagaio US Corp
[](https://www.linkedin.com/company/getdarwinai) [](https://www.instagram.com/darwinai.en/) [](https://x.com/getdarwinai) [](https://www.tiktok.com/@getdarwin)
```json
{
"@context" : "https://schema.org",
"@type" : "BlogPosting",
"author" : {
"@type" : "Person",
"name" : "Lautaro Schiaffino",
"url" : "https://blog.getdarwin.ai/en/author/lautaro-schiaffino"
},
"dateModified" : "2026-07-22T12:00:00.471Z",
"datePublished" : "2026-07-22T12:00:00.000Z",
"headline" : "AI Returns Management: Automate Ecommerce Returns and Refunds",
"image" : [ "https://images.unsplash.com/photo-1449247666642-264389f5f5b1?w=1200&q=80" ],
"mainEntityOfPage" : {
"@id" : "https://blog.getdarwin.ai/en/ai-returns-management-automate-ecommerce-returns",
"@type" : "WebPage"
},
"publisher" : {
"@type" : "Organization",
"logo" : {
"@type" : "ImageObject",
"url" : "https://blog.getdarwin.ai/hubfs/Darwin%20AI%20Logo.jpg"
},
"name" : "Darwin AI"
}
}
```
```json
{
"@context" : "https://schema.org",
"@type" : "FAQPage",
"mainEntity" : [ {
"@type" : "Question",
"acceptedAnswer" : {
"@type" : "Answer",
"text" : "It is the use of conversational AI agents and machine-learning decision engines to automate the ecommerce returns process end to end: intake, eligibility checks, refund or exchange decisions, fraud screening, customer communication, and return-reason analysis."
},
"name" : "What is AI returns management?"
}, {
"@type" : "Question",
"acceptedAnswer" : {
"@type" : "Answer",
"text" : "Done properly, it lowers it. The NRF estimates 9% of returns are fraudulent and reports that 85% of retailers already use AI to detect return fraud, because models can screen every case for abuse patterns, something no manual review team can do."
},
"name" : "Does automating returns increase fraud risk?"
}, {
"@type" : "Question",
"acceptedAnswer" : {
"@type" : "Answer",
"text" : "Customers care about speed, clarity, and getting their money back, not about who processes the request. An agent that resolves a return in one conversation at midnight beats a human reply that arrives two days later. The key is a clearly available human escalation path for complex cases."
},
"name" : "Will customers accept an AI agent handling their return?"
}, {
"@type" : "Question",
"acceptedAnswer" : {
"@type" : "Answer",
"text" : "Start with the highest-volume, lowest-ambiguity slice: standard-window returns of full-priced items in resalable condition. Automate that band completely, measure override rates for a few weeks, then expand coverage category by category."
},
"name" : "Where should a small team start?"
} ]
}
```