I automate technical support, end to end

I am a Staff Engineer at DigitalOcean. I have worked in hosting and infrastructure for 17 years, and for the last three I have been building the AI systems that run technical support at Cloudways. Not one chatbot. A system at every stage of a support conversation’s life, from the moment a customer signs up to the review after the ticket closes. All of it is in production, and all of it is measured by the people who use it.

 before the      when the        when an          after the       back into        before the
 ticket          customer asks   engineer picks   conversation    the system       next problem
                                 it up
 ──────────────▶ ──────────────▶ ──────────────▶  ──────────────▶ ──────────────▶  ──────────────▶
 signup          deflection      PlatformOps      QA agent        learning         customer
 verification    escalation      Copilot                          loop             health
 knowledge base                  MCP tools

Before the ticket

Signup verification. Checking a new account used to take a person three to five minutes. Now 70 to 80 percent of signups are checked automatically, low-risk ones in under 30 seconds, and the rest reach a human with the evidence already gathered. Manual review is down 35 percent. Details

Knowledge base. 800+ articles, one per problem. New engineers are productive in three weeks instead of six, and the deflection system answers from the same source. Details

When the customer asks

Deflection. An AI agent handles about 30 percent of incoming conversations on its own, with customer satisfaction above 90 percent, and hands the rest to a person. Verified savings: $270K to $300K a year. Details

When an engineer picks it up

PlatformOps Copilot. A read-only diagnostic agent that investigates live servers for the support team. 68 engineers use it. The first findings report lands in a median of 4 minutes 35 seconds. The engineers grade every answer, and in September 93 percent were graded correct, up from 85 percent in August. Details, and how it is built.

MCP tools. The platform’s API as 43 tools an AI client can call: servers, applications, security, deployments. In testing. Details

After the conversation

QA agent. Every support conversation, 30,000+ a month, is evaluated for clarity, empathy, professionalism and resolution. Coverage went from 10 percent of conversations to all of them. Details

Back into the system

Learning loop. The copilot records what its guidance did not cover, the engineers grade every answer, and a weekly job checks each candidate lesson against the real server before a person decides whether it goes into a runbook. 207 observations so far, 35 refuted by their own check, 23 folded in. Details

Before the next problem

Customer health. Reading support history, usage and incidents together, to see a problem coming before the ticket does. In development. Details

Talk to me

If you are automating any stage of support, I am happy to compare notes. Book a free 30-minute call, ask about mentoring, or write to [email protected]. I am on LinkedIn and GitHub.

Writing

Notes from running these systems. What the measurements say, including the failures.

All posts