← All articles
Agent design • Field note

Your chatbot explains the fix. It should run it instead.

A support bot that answers with instructions leaves the customer to do the work. The way out is an agent that decides what needs doing and hands the doing to automation your team already trusts.

David Soden • 5 min read • 2 November 2026
A man in glasses on the phone at a laptop, frowning as he works through a problem at home
Step four of five, and the customer is on the phone anyway.

What happened

Two Red Hat Asia Pacific leaders told iTnews Asia that companies should treat agentic AI as an infrastructure problem rather than another application. Steve Shirkey, who runs Red Hat's AI platform in the region, said many enterprises are stuck in what he called a "chatbot loop": AI projects that never get past conversational assistants to systems that deliver a measurable business result.

His colleague Chip McClelland, who leads the Ansible automation platform there, made the less fashionable point in the same interview. Most enterprise automation is still plain task-based scripting, and it still beats agents where consistency matters. Shirkey put it bluntly: "If you need to patch 1,000 servers identically, you want a static script, not a reasoning agent." Their way out of the loop is agents that orchestrate the work while established automation carries out the governed steps.

Why this matters to your business

In customer service the chatbot loop has a familiar look. A customer asks to change a delivery address. The bot replies with five accurate steps and a link to the account page. The customer gets stuck on step four and calls. Your dashboard counts the chat as contained and the call as a new contact. You paid for both, and the customer did most of the work.

That gap is easy to miss because the bot's answer was correct. Accuracy is the wrong test for this kind of request. The customer did not want to know how to change the address. They wanted it changed, and a correct explanation of how leaves them exactly where they started.

Close view of server racks with green status lights in a dim data center
The scripts and APIs that already run the same way every time are the safest thing an agent can call.

Why this is a CX-Builder use case

Red Hat's argument is most useful read as a split of labour. The agent reads the request, asks for anything missing, confirms the details, and picks which procedure applies. Your team already runs that procedure the same way every time: the address change API, the appointment rebooking call, the invoice resend. The agent calls it and doesn't improvise it.

That puts the model where judgment is needed, in reading a messy request and choosing among your procedures, and keeps the action itself predictable. CX-Builder builds that shape directly: an agent with a short list of tools, each one wrapping a single fixed action, with a human approval step in front of anything that moves money or can't be undone.

Shirkey also warned that failure modes change completely once an AI can act, prompt injection included. So keep the tools narrow rather than staying in the loop. An agent that can change one field on the signed-in customer's own order has very little room to do damage, however cleverly someone words a message to it.

What this looks like if you build it

Start with one request your bot currently answers with instructions, ideally a frequent one. Wrap the existing API call or workflow as a custom tool with fixed inputs, so the agent supplies an order number and a new address and nothing else. In an agentflow, the agent gathers and confirms those inputs, then calls the tool and reads the result back to the customer. A human-in-the-loop node holds anything above a value threshold for one-click review, and every call is logged.

The tool decides how the change is made. The agent only decides whether to make it and with which details. Your existing automation stays the system that does the work, which is the part McClelland was defending.

A support agent wearing a headset looks at her computer screen in an office
People review the actions that need them, and stop walking customers through the ones that don't.

The takeaway

Pull last month's bot transcripts and find the most common request where the answer was a list of steps the customer had to follow. If that action already runs through an API or script your team trusts, give the agent that one action, gate it where the stakes call for it, and watch whether the matching tickets and calls go down.

All articles Install CX-Builder View on GitHub