One of our longest-running clients is a direct-to-consumer e-commerce company. 25 to 26 different products under their umbrella. Their core competency is digital advertising — they find products, test market demand with landing pages and small ad spends, then build brands around winners and fold them into their portfolio.
Their most successful product had a 21% refund rate. About 40,000 customer support tickets per month, 80% of them refund-related. They had an outsourced team of roughly 33 people handling it, costing $50K-plus a month. And the team had a problem beyond the usual human limitations: they weren’t following the refund pushback logic that leadership had designed.
This is one of our best case studies, and the reason has nothing to do with the AI technology. It has everything to do with the business logic that already existed.
The Logic That Already Worked
The company had built detailed refund pushback procedures. When a customer requests a refund, don’t immediately process it. First, understand the problem. Is it a technical issue with the product? Walk them through troubleshooting. Is the package lost in transit? Offer to ship a replacement instead of a full refund. If the customer still wants out, offer alternatives — an upgraded model at a discount, partial credit toward a future purchase, a 50% refund with a complementary product.
Only after exhausting every option should a full refund be processed, and at that point, escalate to a human for the final decision.
They called it “refund pushback logic.” And they’d proven it worked. Before coming to us, they took a specialized team of three people and had them handle a subset of tickets for six months, rigorously executing the logic. The refund rate dropped by 5%. The “sales save” rate — where customers asking for refunds end up upgrading to a better product — increased by 70%.
The logic was proven. The problem was execution at scale.
Why the Human Team Couldn’t Execute
The offshore team was operating at roughly 60-70% accuracy on following the prescribed logic. They were skipping steps. They were proactively suggesting refunds before customers even asked — the exact opposite of what the playbook said to do. Some customers would express mild dissatisfaction and the agent would immediately offer a full refund without attempting any pushback.
This is the universal problem with scaling through human labor. You have proven logic, clear SOPs, detailed decision trees. But translating that logic downstream — to 33 people, across time zones, with turnover, with varying levels of engagement — and having them execute it appropriately every single time is nearly impossible.
The constraints stack up: humans cost money, they need breaks, they churn, training new people takes time, quality degrades as the team grows, and micro-managing 33 people to ensure compliance with a detailed decision tree is a full-time job in itself.
Now imagine tripling the ticket volume because sales tripled. You’d need to triple the team. Triple the training. Triple the management overhead. Triple the cost. And still have the same execution accuracy problem, just at a larger scale.
What We Built
A full agentic system — not a chatbot — designed from day one to handle the complete customer support workflow, starting with refund pushback.
The system classifies incoming tickets, determines intent, retrieves relevant order and tracking data, and applies the layered pushback logic. First, understand and classify the problem. If it’s a product issue, attempt troubleshooting from the knowledge base. If that doesn’t resolve it, offer an upgraded model. If the customer declines, offer partial credit. Then a 50% refund with a bonus product. Then a 50% refund with credit toward next purchase. Only after exhausting every layer does it escalate to a human for a final decision.
The system integrates directly into their existing helpdesk. It’s customer-facing, runs 24/7, and responds instantly — although we discovered that instant responses aren’t always optimal. There’s a direct correlation between a slight delay in response time and lower refund rates. The psychology of customer support interaction timing is its own rabbit hole.
One important design choice: we don’t respond instantly every time. We calibrate response timing based on what the data shows produces the best outcomes. This is the kind of nuance that only comes from being embedded in the business and optimizing for business metrics, not just technical performance.
The system also has a memory layer. Every interaction makes it smarter. Patterns that would take the organization a year to recognize — the system spots them in weeks because it’s processing 40,000 tickets a month and actively seeking patterns.
The Numbers
The aggregated refund rate dropped from 21% to 16%. Five percentage points. Their goal coming in was 1-2%.
To understand why 5% is transformative for this business, you need to understand the economics of DTC e-commerce.
Everything in this space runs on digital advertising. Their most successful product has maybe one real competitor. Brand loyalty is essentially zero — customers buy the product, not the brand. Like calling a plumber: if the first one doesn’t answer, you call the next one. So ranking at the top of Google ads is everything.
How do you rank higher? Spend more money per click. How can you afford to spend more? Your customer acquisition cost needs to support it. How do you lower your effective CAC? Increase the lifetime value of each customer.
Lifetime value is the average revenue per customer over time. Every refund directly reduces that number. Lower lifetime value means you can afford to spend less per customer acquisition. Spend less, rank lower. Rank lower, get fewer sales. Fewer sales, less revenue. The downward spiral.
Flip it: reduce refunds, increase lifetime value, afford to spend more per click, rank higher, get more sales, more revenue. The flywheel spins in the right direction.
Plus the sales save logic: customers who came in asking for a refund are now walking away with an upgraded product. That’s found revenue — money that would have been returned is now additional revenue. This alone generates roughly $200,000 per month in recovered revenue.
The cost reduction is significant too — from $50K-plus per month for the human team to about $6,500 per month for the AI system. An 87% reduction. But honestly, the cost savings aren’t even the headline. The headline is the structural transformation of customer support from a cost center into a revenue capture mechanism.
The Real Insight
Every ticket is a sales opportunity. Every customer interaction is a chance to increase lifetime value. The customer support function was never just a cost center — it was a retention and revenue mechanism that wasn’t being operated that way.
That reframe changes everything about how you invest in it. You’re not trying to minimize the cost of handling complaints. You’re trying to maximize the revenue captured from every customer interaction. Different objective, completely different system design, completely different ROI equation.
And here’s the thing I want to make clear: the AI technology was necessary but not the source of the value. The value came from the business logic. Their refund pushback system, developed over years, was the actual intellectual property. We just gave it a body that could execute it at scale with consistency that humans couldn’t match.
This is the recurring pattern in almost every project we do. Someone has proven business logic. They know it works at small scale or in theory. The constraint is execution — throughput, accuracy, consistency, cost. AI removes the execution constraint and lets the logic operate the way it was designed to.
If the logic is bad, AI makes it worse faster. If the logic is unproven, you’re running an experiment at scale, which is risky. But if the logic is sound and the execution is the bottleneck — that’s the exact scenario where AI-native transformation produces results that feel almost unfair.
What Comes Next
With refund pushback working, the flywheel kicks in. More cash from reduced refunds means higher lifetime value. Higher LTV means more ad spend budget. More ad spend means more customers. More customers means more tickets. The system scales without proportionally scaling cost. And now we move to the next bottleneck.
Ad creative generation and optimization. Closed-loop testing where the system creates ads, runs them, analyzes performance data, and generates improved versions. If we can reduce the cost to acquire each customer while simultaneously increasing their lifetime value, the compounding is exponential.
Then further into the operation — other products in the portfolio, other departments, other workflows. Each one unlocks the next. Each one compounds the advantage.
This is why we’re building toward a performance-based model. If our systems produce millions in additional revenue, a flat project fee doesn’t capture the value being created. Aligned incentives — where we earn more when the business earns more — keep us motivated to continuously improve, integrate new models, update architectures as the landscape evolves, and hunt for the next bottleneck.
The systems are real. The numbers are real. And the logic — the business logic that the operator spent years developing — is the real asset. AI just lets it run.
