Case study · Google

How we prepared responses for more than 70% of support inquiries

I led the development of an AI support agent at Google. It investigated technical issues and wrote replies for support specialists to review and send.

By

Talk about a similar project →How it worked ↓
70%+
of inquiries received review-ready replies
1 day
customer response time
Cleared
the growing inquiry backlog

01 · The problem

Support work was piling up

Technical inquiries arrived as email conversations. For each one, a support specialist read the thread, checked the relevant product systems and backing data, consulted technical documentation, and wrote a reply.

As the number of inquiries grew, so did the backlog. Much of the team’s time went into gathering the same kinds of evidence again and again.

02 · The system

How the agent worked

The agent gathered the evidence and wrote a response. A support specialist reviewed it before anything went to the customer.

  1. 01forum

    Read the conversation

    The agent read the full email thread to understand the customer’s issue and its context.

  2. 02account_tree

    Plan the investigation

    Google ADK routed the inquiry through the workflow and selected the evidence it needed.

  3. 03database

    Check the source systems

    Tools looked up the current system status and backing data. The agent did not have to rely on model recall.

  4. 04menu_book

    Consult the documentation

    The agent used relevant internal and external documentation to write the draft.

  5. 05person_check

    Hand it to support

    The system saved the response for a specialist to review, edit if needed, and send.

03 · The safeguards

Checks before support saw the response

Each draft went through a set of validation checks before it reached the support team.

policy

Information boundaries

One check looked for internal information that should not appear in a customer reply.

fact_check

A second model review

An LLM judge checked the draft again for quality and consistency.

verified_user

Support approval

A support specialist made the final decision about what to send.

04 · The result

The backlog cleared, and replies went out within a day

For more than 70% of technical inquiries, the workflow handled the investigation and produced a reply ready for review. The team cleared its growing backlog and brought customer response time down to one day.

The pilot led to a decision to integrate the agent into Merchant Center Assistant, the AI chat assistant available to Merchant Center users.

A similar problem?

Does your team repeat the same investigation for every request?

I can help you map the workflow, work out where an agent is useful, and keep the right decisions with your team.

Book an architecture call →