What Microsoft learned scaling AI across 45,000 contact center agents

There's a lot of AI transformation fatigue out there right now.

And yet, according to CSC’s Customer Support Reimagined with Generative AI eBook, only 38% of customer success processes are currently automated – suggesting that for most teams, the transformation hasn't really started.

At Microsoft, we feel it too – and we've found that mapping the journey into three distinct stages has made a real difference, both for our teams and for how we communicate progress internally.

With roughly 45,000 agents across 25 centers worldwide, processing hundreds of millions of cases every year for customers in 192 countries, we've had to get this right.

Here's what we've learned.

The three stages

Stage 1: Assistive AI

This was the early Copilot era. The focus was on literacy, belief, and getting AI into people's hands in a way that actually helped them day-to-day. Think:

• Drafting emails with Copilot
• RAG-based solutions giving support engineers answers to highly technical questions
• Self-service deployments in customer portals
• Personal action logs and deliverable creation

The gains were real. People came in with genuinely exciting stories: "I built this thing, it runs every Friday evening and summarizes every action I need to take." But here's the honest truth: it didn't move the needle on organizational KPIs.

Saving five minutes drafting an email sounds great until you do the math across hundreds of millions of cases. You can't cash that check. The gains are too small, spread across too many people doing too many things at once.

The biggest risk at this stage? False confidence and early declarations of victory. A lot of companies announced they'd "landed AI," cut delivery partners, then quietly rehired everyone six months later.

Be cautious here. The wins are fast, but the impact is shallow.

This post is for subscribers only

Become a member to get access to all content

Subscribe now