Every outsourcing provider is publishing a version of ‘AI will transform retail customer experience’. None of them is wrong and none of them is useful, because the interesting question is not whether to automate. It is which contacts, in what order, and where to stop.

This post gives two lists and the reasoning behind each. Centro sells both the technology and the people who handle what the technology should not, so read the second list knowing there is an interest behind it. The reasoning stands on its own and you can check it against your own queue.

Where retail customer support automation actually is

Consumer willingness is further ahead than most retail roadmaps assume. In a Verint survey of 5,000 US consumers conducted in January and February 2026, 69% said they would accept fully automated service if it resolved their issue, rising to 93% among younger customers. In the same survey 78% said they prioritise the fastest resolution over their preferred channel.

The condition attached to that willingness is the whole story. It is conditional on resolution. The same survey found 79% would switch to a competitor after a single negative experience. Customers are not resistant to automation. They are resistant to automation that does not work, and they leave over it.

Retail also has a specific trust gap. Zendesk’s 2026 CX Trends research found retail CX leaders trail other industries by 8 to 12 points on willingness to invest in AI decision transparency. That matters more in retail than in most sectors, because two of the highest-volume automated decisions in retail (return eligibility and fraud triage) are decisions made about the customer rather than for them.

Automate these

Each of these is high volume, deterministic, and has an answer that exists in a system rather than in someone’s judgement. Each also carries a condition, and the condition is not optional.

The one worth doing first

Proactive exception notification is last in the table and first in return on effort. Every other item on the list handles a contact more cheaply. This one stops the contact happening. Retailers consistently sequence it last because it sits between support and logistics and neither team owns it.

Do not automate these

The common thread is not complexity. It is consequence. Each of these is a contact where being wrong costs materially more than handling it costs.

Delivery failure with a deadline attached

A birthday, a wedding, a flight, a gift for a child. The customer is not asking a question, they are asking for a decision: reship, refund, expedite, or compensate. Automation can gather the facts. It should not deliver the outcome, because the outcome is a commercial judgement about a customer relationship.

High-value order disputes

Set a threshold appropriate to your average order value and route everything above it to a person, regardless of the question. The economics are straightforward: the cost of an agent handling the contact is trivial against the value of the order and the customer behind it.

Fraud triage and its outcomes

Never deliver a fraud decision through an automated channel. NRF’s 2025 returns research found 9% of returns were fraudulent, and Appriss Retail put the cost of fraudulent returns and claims at $103 billion in 2024, so the problem is real and the pressure to automate the response is real with it. But the same research found 45% of consumers think it is acceptable to bend the rules on returns, which means a meaningful share of the people your model flags are ordinary customers rather than organised fraudsters. An automated refusal delivered to a legitimate customer is the worst single output a retail support system can produce.

Loyalty and membership escalations

Your loyalty programme exists to identify customers worth keeping. Routing them to automation when they escalate inverts the purpose of the programme.

Anything the customer has already tried once

This is the rule that matters most and is implemented least. A repeat contact means the first attempt failed. Sending the customer back through the same automated flow that already did not work is how a minor issue becomes a review, a social post, or a churned customer. Detect the repeat and route it to a tenured agent, every time, regardless of the contact type.

Score your own contact types

Rather than adopting the two lists above wholesale, score your own contact types on three dimensions. Volume tells you whether automation is worth building. Variability tells you whether it can work. Consequence of error tells you whether it should.

Add the three scores. 12 and above, automate. 8 to 11, assist the agent rather than replacing them. Below 8, keep it human and staff it properly. Consequence of error carries a veto: anything scoring 1 on consequence stays human regardless of its total.

The handoff is where programmes fail

Most automation programmes are evaluated on containment rate, which measures how many contacts never reached an agent. It is the wrong primary metric, because it is maximised by making escalation difficult.

• Carry the context across. The agent must see what the customer already said and what the system already tried. A customer repeating themselves after an automated attempt experiences the automation as an obstacle they had to defeat.
• Make the exit obvious. A visible route to a person increases trust in the automation and, counter-intuitively, is not used by most customers when the automation works.
• Escalate on sentiment and repetition, not only on intent. The second time a customer rephrases the same question, the flow has failed.
• Measure resolution rate, not containment. With 85% of retail leaders saying customers drop brands that miss first contact resolution, containment without resolution is a metric that improves while the business gets worse.

What this does to your agent profile

Automate the six contact types in the first list and the work left over is harder, not easier. Fewer agents, handling a queue with no simple contacts in it, where every interaction is an exception, a judgement or a customer who has already failed once.

Three consequences worth planning for rather than discovering:

1. Hiring profile changes. You are recruiting for judgement and product depth rather than throughput.
2. Pay expectations change with it. A queue of exceptions is a more demanding job and the market prices it accordingly.
3. Average handle time rises, and that is correct. If AHT goes up after automating the simple volume, the programme is working. Teams that keep AHT as a headline metric through an automation programme end up punishing the outcome they paid for.

A 12-month sequence

Note what is not in the sequence: a platform purchase. The measurement in Q1 tells you what you need, and buying before that step is how retailers end up with capability they have no volume for.

Retail customer support automation: the short version

Automate order status, return initiation, store and stock queries, resets, booking, and proactive notification. Keep delivery failures with deadlines, high-value disputes, fraud outcomes, loyalty escalations and every repeat contact in front of a person. Score your own contact types on volume, variability and consequence rather than adopting anyone’s list, measure resolution rather than containment, and expect handle time to rise, because that is what success looks like.

Sources

• Verint, survey of 5,000 US consumers, fieldwork January to February 2026, published 12 May 2026. Vendor-run survey measuring stated intent. https://www.businesswire.com/news/home/20260512146246/en/
• Zendesk, CX Trends 2026, Retail. AI transparency gap and first contact resolution figures. https://cxtrends.zendesk.com/reports/retail/
• National Retail Federation and Happy Returns, 2025 Retail Returns Landscape, 15 October 2025. Return fraud share. https://nrf.com/research/2025-retail-returns-landscape
• Appriss Retail, fraudulent returns and claims cost, $103 billion in 2024. https://www.businesswire.com/news/home/20241230601195/en/