Claude agents for an operations team that could not keep up

Challenge

The company sells warehouse and route management software to logistics operators. It was growing fast; its 9-person operations and support team was not. The Head of Operations put it like this: “We are the glue between four tools that do not talk to each other.”

Around 140 tickets a day came into Zendesk. For every relevant incident, someone had to:

• Copy a summary into a Slack channel so product or engineering would see it.
• Log it in a Google Sheet that management reviewed every Monday.
• Look the customer up in HubSpot to find their plan, account manager and whether they were up for renewal.
• Tell sales when a large customer had repeated problems.

We measured the process for two weeks: about 120 hours a month went on moving data from one place to another. And time was not the worst of it. A critical incident took 4 hours on average to reach engineering, and two customers in renewal complained that their account manager knew nothing about their problems.

Solution

We scoped an AI implementation in three two-week deliveries, starting with the workflow with the most hours at stake.

Agent 1: triage and enrichment. When a ticket comes in, Claude reads it, classifies it by type (incident, how-to question, billing, feature request) and severity, and queries HubSpot through its MCP server. It adds the customer’s plan, account manager and renewal status to the ticket. It is the same foundation I describe in AI ticket triage in Zendesk.

Agent 2: coordination. If the incident is serious, it posts a structured summary to the affected team’s Slack channel with a link to the ticket. If the customer is in renewal or has had more than three incidents in 30 days, it sends their account manager a direct message. It also keeps the incident sheet up to date and logs an activity in HubSpot. The technical detail is in Zendesk, Slack and Google Sheets integration.

Agent 3: weekly summary. Every Monday it produces the incident report for management: volume, recurring themes, accounts at risk. It used to take someone half a day.

Rules we set from day one:

• The agents never reply to customers. They only classify, summarise and notify.
• If classification confidence is low, the ticket goes to a review view in Zendesk.
• Everything an agent does is recorded in an audit sheet.

Result

After two months in production:

• 120 fewer hours a month of copying and pasting between tools. The team absorbed 30% more tickets without hiring.
• A critical incident reaches engineering in about 25 minutes, down from 4 hours.
• No key-account incident has gone unreported to sales.
• Classification matches the team’s own in 91% of reviewed cases.

“Now the team works on the tickets instead of moving the tickets around.” Head of Operations (anonymised)

What I would do differently

I would have agreed the format of the Slack summary with engineering earlier. The first version was too long and people ignored it. Cutting it to three lines fixed it: what is happening, who is affected, link.