McKinsey's State of AI: voice agents are the highest-value use case — if you can get past the pilot

Every year the headline number goes up, and every year the harder number barely moves. McKinsey's State of AI report captures it cleanly: 88% of organizations now use AI in at least one business function — up from 78% the year before — but only about a third have scaled it past pilots.

That gap is the whole story. And for voice AI in customer service, it's also the opportunity.

The value gap: adoption is easy, EBIT is not

The report's most telling pair of numbers isn't about adoption at all — it's about realized value. Per McKinsey's State of AI, 64% of organizations point to AI-driven innovation, but only 39% report an actual enterprise-level EBIT impact.

In other words: nearly everyone is doing AI; far fewer are banking it. The difference between the two is almost never the model — it's whether the thing made it into daily operations at scale, or stalled as an impressive pilot.

Where the value concentrates: customer service and conversation

McKinsey's State of AI points to conversational interfaces and call-center / customer-service automation as among the highest-value applications of AI. That's not a coincidence — it's where the work is repetitive, high-volume, and directly tied to revenue and retention.

Map that onto voice AI and the conclusion is direct: an AI voice agent that answers and resolves customer calls sits exactly in the area the report calls most valuable. Every missed call, every hold queue, every after-hours voicemail is measurable lost value — and it's precisely the surface a voice agent addresses.

Agentic AI: everyone's experimenting, few are scaling

The report's read on agentic AI mirrors the broader pattern. Per McKinsey's State of AI, about 62% of organizations are experimenting with AI agents, but only 23% are scaling them — and high performers scale agents roughly three times as often as everyone else.

So the leaders aren't the ones running the most pilots. They're the ones who got agents into production. That's the divide that actually separates outcomes.

Why the pilots stall — and what "in production" takes

Here's the part the survey numbers imply but don't spell out: the gap between experimenting and scaling is a production-engineering gap, not a model gap.

A voice-AI pilot demos beautifully in a controlled setting. Then it meets a real phone line, and hits everything a demo dodged:

• Telephony and SIP variability — real networks add jitter and delay you don't control.
• Voicemail — a large share of outbound calls hit a machine; the agent has to know.
• Barge-in — callers interrupt, and an agent that talks over them feels broken.
• Latency — past the sub-500ms human band, callers talk over the agent or hang up.
• Human handoff — when the agent should pass to a person, it has to do it with context.

None of these are AI problems. They're the unglamorous plumbing of a call that survives contact with a real customer — and they're exactly where pilots die. (We went deeper on this in the boardroom mandate: why pilots die in production.)

Call2Me: the production example, not the pilot

Call2Me is built around that production layer — sub-500ms latency, barge-in, voicemail detection, real SIP/telephony, a built-in knowledge base, 9 languages, and clean human handoff. It's a working, in-production version of exactly the customer-service voice AI McKinsey's State of AI flags as highest-value.

If the report's lesson is that value comes from scaling, not experimenting, then the practical move is to start with something already built to survive a real call — and put it on your line this week, not next quarter. For the buyer's checklist, see how to choose a voice AI provider; for the cost math, how much an AI receptionist costs; and for the full landscape of options, the best AI answering services & AI receptionists compared.

Source: McKinsey & Company, The State of AI. Figures cited are from McKinsey's State of AI survey.