We put Claude into a client's support flow. Here's what actually moved the needle.

What we tried first (and why it flopped)

Our first instinct was a chatbot. Of course it was. Drop Claude on the homepage, train it on the docs, done. Right?

The customers hated it.

Not because the answers were wrong — Claude was surprisingly good. They hated it because they'd already tried three AI chatbots that month that couldn't help them, and they assumed ours was another one.

Deflection, it turns out, is a trust problem before it is a technology problem.

What actually worked

We moved the AI behind the scenes instead of in front of it.

When a ticket comes in, Claude:

• Drafts a response (never sends it on its own)
• Tags the ticket by category
• Flags anything urgent for human eyes
• Suggests the knowledge base article if one exists

A support agent sees all of this when they open the ticket. They can send the draft in one click, tweak it, or ignore it.

The result: 4× faster resolution time, and customer satisfaction went *up*, not down — because every response still felt human.

The lesson

Everyone wants to put AI in the fanciest, most visible place. Sometimes the biggest win is using it somewhere no customer ever sees.

We now lead every AI project with the same question: 'What's the least visible place this could add value?' It's a better starting point than 'where do we put the chatbot.'