How to Actually Implement AI in Customer Support (Most People Skip These Steps)

Yesterday I joined Emily Lampert's webinar about how Anthropic is using Intercom for AI-powered support. Here’s the real AI implementation playbook straight from Anthropic!

Tell AI its role.

Don’t assume AI knows how to behave -- you have to train it like a new hire.

• "You are a support expert."

• "Here's our style guide on how to talk to users. Headings {here}; Bullets {there}."

Feed it the right content.

AI is only as good as what you give it. If your help docs are weak, AI won’t magically fix them.

• Upload existing product guides and marketing materials.

• Review your support site articles and restructure to be more easily consumed by AI.

• Have AI review and rewrite your top-used support macros as snippets.

Get humans involved -- don’t build AI in isolation.

Your support team knows what’s missing. Make them part of the process.

• Anthropic ran a 3-4 day hackathon where support specialists doubled the number of snippets to be used by AI.

• AI works best when trained by the people who actually do the job.

Test quality before launch.

Want to know if AI is ready? Be a user.

• Run real queries and see if AI’s responses are clear, correct, and actually helpful.

• If it’s confusing for you, it’ll be worse for customers.

Keep getting better.

Continuously audit what AI is saying -- don’t “set it and forget it.”

• Make sure your Support Ops team is reviewing AI performance (Resolution Rate) looking for content gaps at least weekly.

• Get your Specialist humans seeing each escalated case as an opportunity to improve AI performance -- real-time content optimization.

Most companies implement AI like it’s a magic switch. It’s not.

If you don’t invest in content, structure, and human oversight, your AI is just an expensive FAQ bot.

AI isn’t replacing humans -- it’s amplifying their expertise.

The question is: Are you training AI to be an expert on your team? What else have you seen work well?