Standing up an AI support team for a D2C brand

The situation

The brand sells home and kitchen products online, with most customer conversations arriving on WhatsApp and email. Ticket volume roughly doubled during every sale. Most tickets were repetitive, such as order status, returns and delivery questions, but they buried the complex cases that needed a person.

The founders had tried an off-the-shelf chatbot. It answered confidently and was often wrong, and customers noticed. They wanted something that would actually help, and they did not have the in-house AI experience to build it.

The brief

Automate the repetitive questions safely, get complex cases to a human faster, and never let the system make a promise the company could not keep.

What we did

We placed a senior AI/ML engineer and a full-stack engineer, both from our vetted network. Their first two weeks were spent not on models but on data: reading several hundred real tickets, building a taxonomy of question types and agreeing with the support lead what a correct answer looked like for each.

• Evaluation before launch. Four hundred real, anonymised tickets became a golden set with approved answers. Every change to prompts, retrieval or models was scored against it before release.
• Retrieval over the truth. Answers were grounded in the company's policies and live order data, retrieved at question time, rather than in whatever the model happened to remember.
• Read-only tools first. The agent could look up orders and delivery status. It could not issue refunds or change orders; those went to a person, with the case already summarised.
• Guardrails. Checks for prompt injection, personal data in logs and answers outside the policy scope, with an easy "talk to a person" path at every step.
• Gradual rollout. The system started on email, then a share of WhatsApp traffic, then all of it once the evaluation scores and customer ratings held up.

How it went

By the end of the engagement, a large share of tickets was being resolved without a human agent, and the human team was spending its time on the cases that genuinely needed judgement. First response times fell from hours to minutes. Customer satisfaction held steady, which, given how the earlier chatbot had gone, was the metric the founders cared about most.

What we'd tell another team

• Build the evaluation set first. Without it, every change is a guess.
• Never give an AI system the power to spend money or make commitments without a human check.
• Measure deflection honestly. A ticket the customer gave up on is not a ticket you resolved.

The engineers who do well on AI projects are the ones who spend the first fortnight reading tickets, not tuning prompts.
Saurav Kumar Jha, Founder & CEO