Most retailers are carrying more customer service load than they were a few years ago, and the cost of carrying it keeps climbing. Shoppers moved to digital channels and stayed there. They expect an answer at midnight as readily as at noon, and they move between the website, the app, the phone, and the store without thinking twice about it. Every one of those channels generates contacts, and most of those contacts are routine: where is my order, how do I return this, why was I charged twice, do you have it in stock. None of it is complicated. All of it is expensive when a person has to handle every one. The question for most service teams is no longer whether to automate, but how to do it without making the experience worse.
Business problem
Customer service is a high-volume function where most of the volume repeats. Order and delivery status, returns and exchanges, billing and payment questions, product availability, account changes: these are the bulk of inbound contacts, and they pull associates and call center staff away from the work that moves revenue. Queues build. Customers wait. Some give up and don't come back.
The bulk of inbound contacts - high volume, low complexity, and expensive one conversation at a time.
The cost isn't only operational. Slow or clumsy service drives cart abandonment, erodes loyalty, and feeds churn, and the load keeps growing as customers expect support on more channels and around the clock. A retailer that staffs for peak pays for idle capacity in the quiet hours. A retailer that staffs for the average drowns every time there's a promotion or a shipping delay.
Challenges
A few things make this hard to solve cleanly.
Repetitive, low complexity
Status checks, returns, and billing questions are a large share of inbound contacts, and they're expensive to handle one conversation at a time.
Fragmented channels
Website, app, phone, email, and store, usually with no single connected view of the person. The customer repeats themselves and the agent starts cold.
Sensitive and regulated
Anything touching payments, personal data, or financing sits under privacy and consumer-protection rules, so automation needs real guardrails and a clean path to a human.
Inconsistent quality
Holding one standard across many locations, shifts, and agents of varying experience is genuinely hard.
Staff time is the constraint
Associates pulled into routine service can't focus on selling and high-touch relationships, which is the work that creates value.
Nobody wants a bad bot
People hate talking to bots. They don't want to be deceived, they don't want slow or repetitive, they want their problem solved and a fast route to a human when the bot can't do it.
Approach
The model that works is an AI agent as the first layer of service across every channel, with people handling the exceptions.
A conversational agent resolves routine inquiries end to end: order and delivery status, returns and exchange initiation, billing and payment questions, product and availability lookups, account updates. It works across chat, the app, email, and voice, drawing on a single customer view so the shopper doesn't have to repeat themselves when they move between channels. On anything payment-sensitive or data-sensitive, it stays inside preset guardrails, and it hands off to a human associate, with the full conversation attached, the moment a request falls outside its authority or the customer simply asks for a person.
One agent across every channel. It resolves the routine, and routes the sensitive or complex to a person with full context.
Roll it out in stages rather than switching it on all at once. Start with read-only and informational tasks: status, lookups, scheduling. Once the controls and the confidence are there, move into transactional actions like processing returns, taking payments, and updating accounts. Staging isn't caution for its own sake. Each phase produces the evidence you need to justify the next one.
Read-only & informational
Status, lookups, scheduling. Low risk, fast to trust, and it builds the controls and the evidence.
Transactional actions
Processing returns, taking payments, updating accounts - once the guardrails and the confidence are there.
Why an agent beats the usual alternatives
The other options each solve part of the problem and then hit a wall.
Adding headcount scales cost in a straight line with volume. Every spike, whether seasonal, promotional, or a stockout that goes viral, means hiring, training, and eventually letting people go, and it does nothing for after-hours demand. An AI agent absorbs the swings without a matching jump in cost, and it covers nights and weekends by default.
Outsourcing to a BPO lowers the per-contact price but trades away control of quality and brand voice, adds handoff friction, and still scales with volume. It tends to widen the consistency gap between channels rather than close it.
Rules-based chatbots and IVR trees are cheap and brittle. They follow fixed decision trees, break the moment a customer phrases something in an unexpected way, and often dump the customer into a queue anyway. That's why the older chatbots deflected frustration more than they resolved problems: they couldn't take action across systems. An AI agent understands intent in plain language, reasons through a multi-step request, and finishes the task, processing the return or taking the payment instead of pointing at a help article.
Static self-service like FAQ pages and help centers puts the work on the customer to find the answer and interpret it, and the content goes stale as policies change. An agent grounded in that same knowledge gives the answer in conversation, shaped to the specific account and the specific question.
Action plus judgment. Traditional automation can inform. An agent can resolve - working across order, payment, and account systems inside guardrails, while routing the genuinely complex or sensitive cases to a person with full context. That combination is what lets a retailer raise service quality and lower cost at the same time, instead of trading one for the other.
Klarna went all-in on AI customer service and reported its assistant doing the work of roughly 850 agents, then publicly pulled back toward a hybrid model after finding that complex and emotionally charged cases needed human judgment the agent couldn't reliably provide. The lesson isn't that the automation failed. It's that the winning design is AI handling volume so people can handle nuance, which is exactly why this approach keeps humans on the sensitive contacts.
“The winning design is AI handling volume so people can handle nuance.
Outcomes
The rollout cut cost and lifted revenue at the same time, which was the whole point.
Why Strongly
Agentic systems like this are non-trivial: the number of integration points, the scale of operations, and the bar customers set are all high. Plenty of companies specialize in customer support software, and several focus on support for e-commerce and retail specifically. Strongly doesn't specialize in any one vertical. What we provide is the building blocks that accelerate solutions like this one. Instead of the usual build-versus-buy decision, think of us as a third option: buy 80% of the plumbing, build the 20% that's specific to your business and the experience you want to deliver. Here, that 20% was mostly integrations, agent training for personality, skills, and guardrails, and quality assurance testing.
A third option between build and buy: buy the 80% of plumbing, build the 20% that's specific to your business.
The Strongly capabilities this build relied on:
Streaming workflows
Answering calls, multimodal models that handle speech, and tool calling for database lookups, so every call runs with high accuracy. Elastic, scaling up and down with volume so you hold capacity for peaks without paying for it in the quiet hours.
Data connectors
Strongly connects to more than 500 external tools and services, which makes hooking into retail CRM, ERP, and POS systems secure and scalable.
Governance and guardrails
Occasionally a caller would try to trick the agent - an adversarial attack. It was critical that the agent never leak PII or run a transaction it wasn't approved to run.
Observability and monitoring
Transcripts and recordings for every call were a hard requirement, both for auditing and for future improvement.
Put an agent in front of your service load
This is the first in a series on agentic AI in production. If you're weighing how to automate customer service without making it worse, we can help you scope the 20% that's specific to your business.
Schedule a demo