Building, buying, and partnering for e-commerce AI customer service: costs and control

A home goods company recently needed to evaluate their customer service system. Three proposals arrived: hiring a team to build a custom LLM, integrating a SaaS platform, or bringing in consultants. The owner's first question was: "Which saves the most money?" This frames the decision incorrectly. Cost savings isn't the deciding factor. What matters is whether the system can handle the 200 exception rules in your return policy.

Choosing an AI customer service system means deciding how much cost to trade for control. Three options exist: build your own LLM, integrate a SaaS platform, or work with consultants. The differences lie not in feature lists but in cost structure and who owns your customer and order data.

Three paths: the math

Building your own LLM means maintaining the model, vector database, and retrieval pipeline yourself. Upfront investment is significant. A production-ready system costs six figures monthly (in Taiwan dollars) across engineering, compute, and data annotation, and takes three to six months to reach stable output. Once running, marginal costs approach zero. Your data remains in your data center. You can integrate ERP, payment systems, and customer segmentation tools without restrictions.

SaaS integration is the fastest path. Sign a contract, and you could have a bot handling inquiries like "where's my package?" within days. Upfront costs are low, but pricing scales with conversation volume or seat count. As usage increases, costs increase. Data control is more critical. Customer complaints, order details, and member behavior feed into a third-party cloud. Many e-commerce companies manage the monthly cost. They struggle when their return and exchange rules don't fit the platform's standard workflow.

Consultant-led deployment is the third option. Engineers work on-site to build customer service around your data and workflows, deploy it, and stay until staff actually use it. This approach leverages open-source models and existing systems, customizing each integration for your needs rather than rebuilding components.

Cost and control: side by side

Beautiful demos and cheap pricing mislead

The SaaS proposal appeared cheapest and polished in the demo. The company tested their return policy: "returnable within seven days if unopened, no damage from use." The system simply repeated the terms verbatim. A realistic customer query like "I opened it but it's the wrong size" produced no useful response. Extra FAQ entries won't solve this. The system needs to understand order status, shipping progress, and customer history to respond correctly.

The solution wasn't a simple choice between three options. Instead, a layered approach was used: a lightweight bot handles routine questions like shipping status and invoice requests, while return-and-refund decisions and price disputes go to an embedded RAG system using the company's internal policies as knowledge sources. After three months, average first-response time fell from six minutes to 40 seconds, freeing staff to address the emotionally complex complaints that systems can't handle well.

Selection has no standard answer, only one rule: don't let data control become a hidden cost you didn't account for. To decide, ask yourself two questions: How non-standard are your workflows? Can your customer data stay in-house? Once you've answered clearly, the right path becomes obvious. At Tenten, we help clients work through these questions, then deploy engineers on-site to build a customer service system that works.