AI deflection is doing what the business case promised. Containment is climbing, self-service is doing its job, and fewer calls are reaching human agents. The agent-side data is the part that’s getting more interesting: handle times running longer than projected; CSAT varying more across queues; and attrition moving in unfamiliar directions.
What’s emerging now is the second half of the AI deflection operating model, and it’s where the next wave of strategic CX investment is starting to concentrate.
The work reaching agents has changed shape
The business case for contact center AI has been strong for years: reduce case volume and cost by automating routine work and deflecting simple queries. This then frees human agents for the interactions that need their judgment. That logic is delivering. The implementation reality is turning out to have a second layer worth designing for as carefully as the first. The work now reaching human agents has changed shape, and the operating model around those agents has the opportunity to change with it.
Today, some interactions are arriving pre-frustrated, having already tried and failed to resolve through self-service. Complexity is up, as is the emotional charge. As a result, the judgment required of human agents is harder to script.
McKinsey’s The Contact Center Crossroads, published in March 2025, framed the underlying dynamic well: human agents are still where contact centers route the work that turns on empathy and situational judgment, and that’s the work AI hasn’t been able to absorb. In concentrated form, this kind of workload calls for a different kind of operational support than the contact center industry has historically built around it.
Recalibrating the metrics
Average handle time was conceived as an efficiency metric for a workload that included easy calls. Strip the easy calls out, and AHT measures something different. First contact resolution (FCR) rates assume that an agent has a reasonable shot at resolving the issue inside the call. When the issues themselves are more complex, FCR shifts mechanically without anyone on the floor getting worse at their job. Occupancy stays high, but the cognitive intensity of that occupancy is meaningfully different from what the workforce planning model was calibrated for.
What the dashboard reads as a performance regression is, in many cases, a measurement question. The composition of the work has changed. The metrics have yet to catch up.
The other half of the agent experience question
This shift isn’t happening in isolation. CX Today’s The Algorithm Never Blinks, published in October 2025, surfaces a pattern happening alongside workload concentration. The same wave of investment that deflected easy interactions also expanded what the contact center observes about the humans handling the rest. AI now analyzes 100% of interactions, where traditional QA programs sampled 1-3%. Real-time prompts feed the agent during the call. Sentiment scoring and coaching nudges run between calls.
That’s a real and durable advance in operational visibility, and it appears that wellbeing data is catching up. According to Omdia’s 2025 Digital CX Survey (cited in the CX Today piece), 75% of North American contact center leaders are paying close attention to AI’s impact on agent wellbeing. According to ICMI research cited in the same piece, 45% of organizations and 55% of contact centers don’t yet measure employee satisfaction or stress.
Both figures point to the same opportunity: pairing AI-driven monitoring with parallel investment in agent experience telemetry. Leaders who do so will get the operational signal they need to keep the system performing well as containment rates climb further.
Four layers of the agent-side redesign
The natural starting point is the AI side of the system: better routing, smarter escalation, and more nuanced sentiment analysis. The larger opportunity is on the human side of the operating model, and it tends to involve different sponsors than the original AI deployment did.
Onboarding is the first place where redesign pays compounding returns. Agents spending a workday on concentrated complex interactions need different preparation than agents handling a mixed load. The traditional ramp model, where new hires graduate through easier calls before tackling harder ones, was built for a workload distribution that AI deflection has now changed. Curricula and ongoing development can be redesigned around the work that actually shows up. Organizations that move on this early will see the gains compound as containment climbs.
Real-time support is the next layer. McKinsey’s research found that the organizations achieving meaningful year-over-year reductions in human-handled interaction volume got there by first solving the data, integration, and process problems AI couldn’t fix on its own. The same logic applies to agent assist.
CRM integration that pops the right customer record and an AI copilot that surfaces the relevant customer history at the moment an agent needs them can materially reduce cognitive load and meaningfully improve the interaction. The integration work involved in getting that right is the same discipline that made the deflection deployment succeed in the first place.
Measurement is where the operational gains become visible. The existing KPI stack can be extended to distinguish a composition shift from a performance regression. Adjusting AHT benchmarks for the new mix of interactions is the conceptually easy part. Tracking emotional load across a shift, or measuring whether agents have the room and authority to apply judgment when their experience says to, is harder and more consequential.
Why this matters in the next budget cycle
The agentic layer is moving from pilot to production faster than the previous wave of AI did. Deloitte’s 2026 State of AI in the Enterprise, based on a survey of 3,235 enterprise leaders across 24 countries, found that a quarter of organizations have already moved 40% or more of their AI experiments into production, with another 54% expecting to cross that threshold within three to six months. The same research identifies customer support as the function where leaders expect agentic AI to have its highest enterprise impact.
Product roadmaps line up with that timing. Salesforce has Agentforce Contact Center moving toward general availability, with GA in the US already rolled out and European platform readiness scheduled for September and October of 2026. Cisco has agentic capabilities shipping across Webex Contact Center. Containment rates are about to step up again, and the composition shift that’s manageable at 30% deflection becomes a defining operational variable at 50% or 60%.
The organizations getting the next phase of AI investment right are treating it as an integration challenge across systems and people, not an automation challenge inside a single technology stack. The deflection logic was the foundation. Designing a contact center that surrounds the agents who handle what AI couldn’t is the next strategic project, and it’s where the most durable CX advantage is now being built.
For practitioners who’ve spent years working at the intersection of telephony, CRM, and AI orchestration, the brief hasn’t fundamentally changed. The composition of the work has, and the redesign that follows is the kind of strategic project that compounds in value the longer it runs. Making the system perform as one across what AI handles and what it leaves to humans is the same work it’s always been, only with deeper questions to answer.
