Kogan rebuilds its customer service bot as an AI agent on Agentforce

Many companies shut down an automated support project when the first bot gives poor answers. Kogan did the opposite.

The Australian e-tailer's first bot closed about 6% of customer enquiries. The company rebuilt it as an AI agent on Salesforce's Agentforce. The agent now handles 10 types of request, as CTO Goran Stefkovski explained at the Dreamforce conference, according to iTnews on 14 October 2026^[1].

From web forms to a bot that rarely resolved anything

Kogan's customer support had always run through web forms and email.

Stefkovski saw AI as a way to ask less effort of customers. The second goal was to respond faster.

The first step was a deterministic bot, which follows hand-written rules. The CTO describes it as "the technology we had available".

That bot resolved about 6% of incoming enquiries. Kogan decided to keep the project and replace the engine.

Six weeks from hack day to production

The rebuild was done with the Salesforce team that works inside customer companies. Salesforce calls these engineers forward-deployed engineers.

The figures the company has made public, as reported by iTnews:

• Starting point: about 6% of enquiries resolved by the deterministic bot.
• Timeline: the first use cases were tested at an internal "hack day", with production after about six weeks.
• Scope: the agent handles 10 use cases, from order status to starting a return.
• Weight of the main use case: WISMO ("where's my order?") enquiries make up about 60% of incoming requests.
• Tools: four Agentforce services, namely NextGen Authoring, Agent Script, Observability and Testing Center.

Marketing-Interactive^[2] also describes the project as a six-week sprint.

All of these figures come from Kogan. The outlets report them, and so far no one has checked them independently. The report I was able to read also doesn't say what share of enquiries the new agent closes on its own.

Saying 6% out loud

The most useful part of the story is the low number.

Stefkovski said you have to accept a starting point. In his view, expecting 60, 90 or 100% from day one is unrealistic. He says this is where many companies fail, because they're afraid to show customers something that still needs work.

The second choice was the target. Kogan focused on a single question that accounts for more than half of the volume on its own.

A third use case is still being developed. The agent uses image recognition to assess shipping damage in returns. The same check helps catch duplicate or fake photos sent by customers.

Four steps for your customer service

1. Measure how many enquiries your bot closes today and write the number down.
2. Count enquiries by type and start with the one that carries the most weight.
3. Try the first use cases in an internal day, then give yourself a fixed number of weeks to reach production.
4. Build testing and response monitoring into the project from day one.

Which question fills your inbox?

Every customer service team has its own question that makes up 60% of the volume.

Do you already know yours? And what starting percentage would you be willing to admit to?