Build vs Buy: AI Customer Support

Why the prototype misleads

Connecting a language model to your documentation and getting sensible answers takes a weekend, which makes building look obviously cheaper. That prototype is roughly five percent of a production system, and the remaining ninety-five percent is the part that determines whether it is safe to put in front of customers.

What production actually needs

Retrieval quality tuning, a refusal path, escalation design, permission scoping for any action-taking, audit logging, an evaluation harness, and monitoring. Then a person to maintain all of it as your product and documentation change.

The evaluation problem

You cannot ship customer-facing AI without a way to measure whether it is getting worse. Building that evaluation harness, and keeping the test set current, is genuine ongoing engineering work that most build estimates omit entirely.

The opportunity cost

Every engineer week spent on support infrastructure is a week not spent on the product you sell. That is the real comparison, and it is usually more decisive than the licence cost.

When building makes sense

When support automation is genuinely core to your product, when you have unusual requirements no vendor meets, or when you already run substantial machine learning infrastructure and the marginal cost is genuinely low.

Questions