Online retail was supposed to make the phone obsolete. It didn't. An AI phone agent for e-commerce customer service is now the most direct way to answer those calls without hiring for peak, provided you scope it around the calls your systems can actually resolve.
This guide is for heads of customer care and e-commerce operations at mid-market and enterprise online retailers. It covers why the phone line still matters, a call matrix for deciding what the AI resolves and what stays human, a cost model you can run on your own numbers, the guardrails we insist on, and a vendor checklist.
Why the phone still matters in e-commerce
Four data points frame the case.
- Online is a growing share of every retail dollar. U.S. e-commerce sales reached $340.2 billion in the second quarter of 2026, seasonally adjusted: 17.1% of total retail sales and 12.2% more than a year earlier (U.S. Census Bureau, quarterly e-commerce report released August 18, 2026). More orders means more deliveries, and more deliveries means more “where is it?”.
- Shoppers still prefer to call. In a survey of 1,000 U.S. consumers published in February 2025, commissioned by voice AI vendor PolyAI and conducted by Dynata, 65% said a phone call was their preferred way to contact customer service for retail and travel brands, including 86% of Gen Z and younger Millennial respondents (PolyAI, February 2025). It is vendor-commissioned research, so read it as a signal, not a census.
- Returns are heavy, and they drive calls. NRF and Happy Returns estimated that 19.3% of online sales would be returned in 2025, against 15.8% for retail overall, and that 71% of consumers are less likely to shop with a retailer again after a poor returns experience (NRF, 2025 Retail Returns Landscape, October 2025).
- Customers are wary of AI that blocks the way to a human. A Gartner survey of 5,728 customers, conducted in December 2023, found that 64% would prefer companies didn't use AI in customer service, and that the top concern is that AI will make it harder to reach a person (Gartner, July 2024).
The last point is not a reason to avoid AI on the phone. It is the design brief. The same PolyAI survey found that 55% of respondents are likely to ask for a representative immediately when they hear a robotic IVR, while 71% would talk to an intelligent voice assistant if it could accurately handle their request. Shoppers do not object to automation. They object to automation that wastes their time.
An e-commerce caller is rarely browsing. They are calling because a parcel, a refund or a charge has already gone wrong.
What an AI phone agent for e-commerce customer service should handle
Do not automate by volume alone. Rank each call type by how often it arrives and how much damage a wrong answer does. This is the starting matrix we would use with an online retailer.
| Call type | AI role | Why |
|---|---|---|
| Order status and delivery tracking | Resolve | The highest-volume call in most stores. A read-only lookup on order and carrier data, after identity verification. |
| Store policy: shipping times, sizing, opening hours, payment options | Resolve | Pure knowledge-base answers, as long as the agent answers only from approved content. |
| Return and exchange requests inside policy | Resolve within rules | Check the return window and item eligibility, start the return, send the label. Exceptions go to a person. |
| Cancel or change an unshipped order | Resolve within rules | Only before fulfilment, and only if your systems confirm it. Once shipped, it becomes a return. |
| Delivery problems: delayed, damaged or “delivered but not received” | Assist | Collect the facts, check carrier status and open a documented case. Reshipping or refunding stays with your rules or your team. |
| Pre-purchase product questions | Assist | Answer from the product catalogue, capture the lead, route high-value buyers to a person live. |
| Refund exceptions and goodwill | Keep human | Judgement calls with real money attached. The agent prepares the case, a person decides. |
| Payment disputes, suspected fraud, account takeover | Keep human (AI collects) | High-risk and adversarial. The AI's job is fast identification and a complete hand-off. |
| Angry or vulnerable callers | Keep human | Transfer at the first clear signal, with the full context, so nobody repeats their story. |
“Keep human” does not mean “AI not involved”. On every call, the agent can answer on the first ring, identify the customer and order, and pass a complete case to the right person. That alone shortens every human-handled call. If you run on Shopify, our Shopify AI phone agent guide walks through what the order lookup and return flow look like on that platform.
A cost model you can run on your own numbers
Here is the calculation we would bring to a budget review. Every input below is illustrative; replace each one with your own telephony and helpdesk data.
Three caveats. First, 60% is an average, not a promise: your first-month figure depends on your call mix, and a store whose calls are mostly complex disputes will land lower. Second, net off the cost of the service and the implementation before calling it a saving. Third, the model ignores the revenue side: abandoned calls during a delivery delay turn into chargebacks and lost repeat orders.
For reference, the averages across Unicall's live deployments are about 60% of calls resolved fully autonomously, a 100% answer rate and cost per ticket down 52% within 90 days. Sample sizes and periods vary by client, so test them against your own baseline during an audit rather than pasting them into a business case.
Peak season and returns: the real stress test
Average weeks do not break an e-commerce support team. Peak does. Black Friday, the holiday run-up and the January returns wave compress months of contacts into a few weeks, and NRF's 2025 report found that retailers expected 17% of holiday sales to be returned.
The usual answers to that peak each have a catch:
- Seasonal hires need training before the volume arrives, then leave in January.
- An outsourced overflow partner works, but costs rise with volume and quality is hard to see. We break down the second-order costs of an overloaded queue in the hidden cost of call center overflow.
- Shorter phone hours or a hidden number move the problem to email, chat and chargebacks.
An AI phone agent scales with call volume rather than headcount, which is why peak is where the business case is clearest. The catch is timing: an agent that goes live in mid-November learns your edge cases during your busiest week. Plan to be in production, on overflow at least, by early autumn.
Six guardrails for voice AI in online retail
- Verify identity before revealing anything. Order details, addresses and delivery instructions are personal data. Match the caller to the order (for example phone number plus order number or email) before the agent reads anything out.
- Put refund and cancellation rules in code, not in the prompt. Return windows, value thresholds and fulfilment status should be checked by your systems, not interpreted by a language model. We explain why in Confidently Wrong.
- Keep card numbers out of the conversation. If a call needs a payment, send a secure payment link or transfer to a compliant payment flow. Ask any vendor how card data is kept out of recordings and transcripts.
- Answer only from approved content. Policies, product facts and delivery promises come from your knowledge base and catalogue. When the agent is not sure, it says so and routes the call.
- Make the human one sentence away. “Can I speak to someone?” should trigger a warm transfer with context, not a loop. This is the direct answer to Gartner's top customer concern.
- Outbound needs consent. In February 2024, the FCC ruled that AI-generated voices count as “artificial” voices under the Telephone Consumer Protection Act, so such calls generally require the called party's prior express consent (FCC Declaratory Ruling 24-17). Delivery and abandoned-cart call campaigns deserve their own compliance review.
A 4–6 week rollout that leaves your phone system alone
Unicall deployments typically go live in 4 to 6 weeks without replacing the existing phone system or helpdesk. For an online retailer, the sequence looks like this:
- Week 1 — call-flow audit. Classify a month of calls by reason, outcome and time of day. Order status and returns usually dominate; after-hours volume usually surprises people.
- Weeks 2–3 — integrations and design. Connect the store platform, helpdesk and CRM; load policies, return rules and transfer rules.
- Weeks 4–5 — testing. Replay real scenarios, including the awkward ones: a caller with two orders, a gift recipient who does not know the order number, a return one day outside the window.
- Week 6 — go live on overflow and after-hours first. Then widen to first-ring answering once the transcripts look right.
Buyer checklist: choosing an AI phone agent for your online store
Compare pricing models too. Per-seat pricing makes little sense for an agent that replaces hold queues rather than desks; Unicall's managed service is billed on ticket volume and starts from $1,000 a month (see pricing). If you are weighing a voice agent against the AI built into your helpdesk, our Gorgias comparison sets out the differences on public information.
FAQ
Which e-commerce calls should an AI phone agent handle first?
Start with order status, delivery questions, return and exchange requests that fit your written policy, cancellations of unshipped orders, and store-policy questions such as shipping times and sizing. They are frequent, rules-based and can be checked against your order and shipping data, so they are the safest place to prove the agent works.
Do online shoppers really still call customer service?
Yes. In a survey of 1,000 U.S. consumers published in February 2025 commissioned by PolyAI and conducted by Dynata, 65% said a phone call was their preferred way to contact customer service for retail and travel brands, including 86% of Gen Z and younger Millennial respondents. Shoppers tend to call when something has already gone wrong, which is why those calls matter more than their volume suggests.
Can an AI phone agent issue refunds or cancel orders?
It can, within rules you set and enforce in your systems: for example, cancel an order only before fulfilment, or start a return only inside the return window and below a value threshold. Refund exceptions, goodwill gestures and anything outside policy should go to a person, with the case already documented.
Will customers be annoyed to reach an AI on the phone?
Some will be if the AI blocks them from a person. A Gartner survey published in July 2024 found the top customer concern about AI in service is that it will make it harder to reach a human. Answer on the first ring, resolve what you can quickly, and transfer with context the moment the caller asks for a person or the agent is unsure.
How long does it take to put an AI phone agent on an online store's support line?
With a done-for-you provider, a first scope covering order status, returns and store-policy questions typically goes live in 4 to 6 weeks without replacing your phone system or helpdesk. Access to your store platform, helpdesk and shipping data drives most of that timeline, so start at least two months before your peak season.
Unicall.ai builds and operates done-for-you AI voice agents that answer every call 24/7, take real actions in your systems, and hand over to your people, with full context, the moment they should.