Skip to content
Fast Company Executive Board

The Fast Company Executive Board is a private, fee-based network of influential leaders, experts, executives, and entrepreneurs who share their insights with our audience.

Customer context must survive the handoff from AI to human

Design and measure your system to capture and transfer the information properly.

[Photo: Getty Images]

Profile photo of Amit Mathradas

Chief Executive Officer · Nintex

AI now handles more of the customer conversation than it used to: the routine questions, the simple lookups, and the first few minutes of almost every interaction. According to McKinsey, AI can help, tackling upwards of 60% of addressable care volume. This would allow humans to concentrate on the customer interactions that bring loyalty and value.

That’s the vision I call Humantic CX: human and AI agents working as one continuous system, each making the other better. Done right, it works well. But the more AI carries the front of a conversation, the more often it changes hands to a human. Every handoff is a place where something can be lost, which happens more often than you’d think.

In our recent research, 96% of organizations believe they preserve context when a conversation moves from AI to a human agent. However, 83% of customers said they had to repeat themselves on some occasions. That gap exists because most companies measure whether a handoff happened, not whether the context survived or was serviced well.

THIS ISN’T A TRANSCRIPTION PROBLEM

On hearing that customers repeat themselves, many organizations assume the transcript wasn’t good enough; a sharper summary and a more complete record will fix it.

It won’t. A transcript, however accurate, records what was said, not what’s already been established or resolved. A customer who verifies their identity with an AI agent and is then asked to verify again by a human has hit an architecture that never told the second system what the first already knew. The same happens with everything else a customer establishes early on: what they’ve already tried, confirmed, or been told. It’s in the record, but missing from the moment the human needs it.

That’s the real failure, the design of what travels with the customer and what doesn’t.

WHY CONTEXT KEEPS DISAPPEARING AT THE HANDOFF

When we asked business leaders what breaks during these transfers, the answers split into a few categories. Visibility is one. We found that over a third of leaders said agents lack visibility into the prior AI interaction because nothing was built to surface the data when needed.

Fidelity is another. Inaccurate transcripts, or redaction and compliance rules, strip out the exact detail an agent needs to pick up the thread.

But the deepest problem is that even a perfect, compliant transcript is still just a written account. Handing an agent a wall of text and asking them to reconstruct what’s already resolved isn’t meaningfully different, from the customer’s seat, than making them explain it again. There is still an uncomfortable wait.

WHAT IT COSTS AGENTS AND CUSTOMERS

Agents are absorbing this cost more than most dashboards show. Across leaders we surveyed, 61% said human agents are now fielding fewer routine inquiries; AI took the easy calls, as promised. But leaders are split on whether that made the job easier or harder. The honest read is that AI didn’t relieve the workload; it concentrated it.

What’s left for human agents are the judgment calls and escalations. They handle conversations where missing context does the most damage, because there’s no script to fall back on when you don’t know what already happened.

For customers, the cost is less about time than about how the delay gets interpreted. A slow start with a human agent feels like ordinary, forgettable friction. But providing your information to an AI agent and then asking them to repeat everything to the human agent feels different. It’s noticeable proof that the AI agent wasted your time, which is harder for customers to let go of, and it’s costlier for brands.

THE METRIC HIDING THE PROBLEM

Ninety-nine percent of organizations we surveyed reported that their customer experience has improved in the last 10 years. Only two-thirds of consumers agree. That gap is how a localized failure survives underneath a positive-looking trend. “Has service improved?” is broad enough to let it hide.

Containment rate is the clearest example of this gap. The containment rate rewards resolving a conversation in that call and not escalating it. However, it says nothing about whether context survived into the calls that do escalate. A high containment number next to a spike in complaints is a problem, if you know where to look.

If I were auditing customer satisfaction today, I’d want three numbers instead of one blended score:

  • How often customers repeat information at the point of transfer
  • How frustrated they are in that moment, and whether it affects their willingness to use an AI channel again
  • Whether human agents themselves, not leadership, say they had what they needed before they said hello to the caller

Right now, all three get folded into a single “handoff effectiveness” score that lets organizations grade their own homework. Pull them apart, and the gap disappears.

DESIGN FOR CONTEXT TO SURVIVE

None of this gets fixed by handing the agent a longer transcript. The fix is to ensure the right context is passed along at the right time, not simply capturing a large volume of information. That takes three deliberate choices:

First, what’s already been tried and confirmed by AI should arrive to the human agent as structured information they can act on immediately. It should not be facts buried in a paragraph they must read and interpret first.

Second, design instead of build the handoff. Give the handoff its own success criteria, audited by the metrics above.

Third, close the integration seams. Every bridge between a separate CRM, contact center, knowledge base, and AI model is a place where context can fall through, even when each system works fine on its own. A unified platform won’t perfect the status summary or judgment by itself, but it removes that seam.

The companies that solve this will stop treating context as a transcript sent by default, and start designing it to give the human agent the information needed up front.

Amit Mathradas is CEO of Five9.