What happens to the metrics you've spent years optimizing when artificial intelligence starts handling a growing share of your customer interactions? That was the central question at the heart of a thought-provoking think tank session Aspect facilitated at this year's Customer Contact East, a Frost & Sullivan Executive MindXChange event.
The session, titled "KPI Drift: How Agentic AI Is Transforming Contact Center Performance," brought together contact center and workforce leaders from across industries to wrestle with a surprisingly thorny problem: the KPIs that once told you whether your operation was performing are increasingly telling you something different … or nothing at all.
Here's what we learned.
The metrics haven't broken. Their meaning has.
The session opened with a provocative observation: AI isn't just changing contact center operations, it's quietly changing what your existing data means.
Take Average Handle Time (AHT). Several participants shared that after deploying AI, their AHT went up. Others saw it go down. Both outcomes had the same root cause: AI was deflecting easy tier 1 contacts, leaving human agents to handle more complex, higher-effort conversations. If you're using AHT as a performance target, you're likely penalizing your best agents for doing the hardest work.
The same distortion affects nearly every legacy KPI:
- ASA (Average Speed of Answer) looks stable on the surface, but AI deflection is quietly absorbing part of the queue. Staffing models that don't account for fully deflected or contained calls are producing plans built on flawed assumptions.
- Occupancy calculations break down when AI handles a share of volume. The fix: measure occupancy against agent-specific contacts, not total resolved contacts.
- FCR (First Contact Resolution) raises a definitional question: when a bot handles part of the interaction before handing off to a human, what counts as "first"?
- CSAT is now channel-dependent in ways it wasn't before. Participants reported that AI has worsened email CSAT while improving text and chat CSAT, particularly when AI is used for information gathering and handoff quality rather than full resolution.
The group landed on a useful guiding principle: For every KPI, ask “Does it still measure what we intended, or has the meaning changed?”
A better north star: cost per resolved contact
If traditional KPIs are losing their signal, what should replace them?
The session converged on a powerful reframe: a customer interaction is no longer a single unit of work. With AI in the flow, it splits into bot time, customer wait time, and human agent time. Measuring only the agent segment, as most legacy tools do, gives you an incomplete picture at best.
The recommended replacement metrics:
- Cost per resolved contact as the primary financial metric, the only measure that spans the full AI + human journey.
- Human minutes per resolved contact for capacity planning.
- End-to-end time to resolution for CX quality measurement.
One example illustrated the gap vividly: an interaction might run 15 minutes end-to-end, with only 2 minutes involving a live agent. Agent AHT looks excellent. The customer's experience may be anything but.
The group also flagged an emerging cost dynamic worth watching: email is becoming more expensive than phone, not less, due to multi-touch back-and-forth and extended resolution timelines. Some organizations are actively deprioritizing email in favor of real-time channels as a result.
Agent assist: a trust problem as much as a tool problem
Participants who had deployed Agent Assist tools shared candid assessments of what drives adoption, and what kills it.
The strongest insight: don't penalize agents for using agent assist; coach for non-usage. The most effective approach described involved supervisors receiving a daily report of agents who hadn't used agent assist, with a coaching conversation the following day. Reframing adoption as a coaching moment rather than a compliance metric made a measurable difference.
Several other dynamics surfaced:
- The trust bucket problem: Every accurate recommendation builds trust. One wrong or irrelevant answer can empty the bucket entirely, causing agents to abandon the tool. Accuracy, source grounding, and fast correction loops aren't nice-to-haves, they're the foundation.
- The experience gap: Veteran agents who spent years accumulating institutional knowledge can feel threatened when new hires immediately access the same information via AI. Framing agent assist as augmentation and standardization, not replacement, is essential for cultural adoption.
- Knowledge currency is foundational: Agent assist is only as good as the knowledge base behind it. Organizations with frequent policy changes need strong knowledge management discipline, or the tool becomes a source of bad answers at scale.
- Context quality at handoff matters enormously: When AI-to-agent summarization is done well, participants described it as "very good", and agents immediately understand the customer's situation and can engage effectively from the first second. This is one of the highest-leverage levers for both AHT and CSAT.
Organizations that cut onboarding time from 6 weeks to 2 weeks with agent assist weren't doing it through magic. They were doing it through better handoff context, better knowledge access, and better real-time guidance.
Governance is not a post-launch consideration
One of the most consistent themes across the session: AI governance requires dedicated infrastructure, clear ownership, and ongoing operational investment, and most organizations are underestimating this.
Key takeaways from the governance discussion:
- AI agents should be held to the same compliance guardrails as human agents. Regulatory standards don't change because the agent is artificial.
- Real-time routing of compliance exceptions (credit card numbers, SSNs, sensitive disclosures) to human supervisors is table stakes in regulated industries.
- Confidence in compliance decreases as AI handles more volume, because the better AI gets, the more data it requires, and that exposure introduces new risk vectors.
- Adversarial probing, customers and bad actors testing AI system limits, is real and ongoing. It needs to be built into your threat model from day one.
- Auditing AI consumes significant person-hours. This surprised many participants and should be factored into any honest ROI model.
The ROI conversation is harder than vendors suggest
A candid undercurrent ran through the session: the financial case for AI is real, but slower and messier than the market implies.
Several dynamics are creating tension:
- AI platform costs are rising to offset savings from deflection. Year 1 should be planned as a break-even or loss period, with meaningful ROI expected after 12–18 months.
- Vendor pitches lead with cost savings. But for many organizations, the actual driver is CX improvement and brand experience, not cost takeout. Misaligned framing creates disappointment when cost savings don't materialize on the vendor's timeline.
- CFOs still expect a cost narrative, even when the initiative is CX-led. Teams need to be prepared to translate qualitative outcomes into financial terms.
- The market is shifting toward qualitative KPIs such as experience quality, retention, satisfaction over purely quantitative ones. Most organizations haven't resolved the tension between these two frameworks internally.
The session's recommendation: run scenario analysis across deflection rate, AHT impact, AI platform cost, and adoption curves before committing to a business case. Optimistic, realistic, and pessimistic scenarios should all be on the table.
Treat automation as a program, not a project
Perhaps the most actionable takeaway from the entire session: organizations that treat AI deployment as a one-time implementation project are setting themselves up for operational drift. The ones succeeding are treating it as an ongoing program.
What that looks like in practice:
- A dedicated owner with a backlog, a release cadence, and measurable acceptance criteria covering both CX and compliance.
- A recurring AI Ops / Contact Center Ops cadence, weekly or biweekly, to review deflection rates, containment, escalation reasons, and emerging failure modes.
- An automation mix dashboard showing human-handled vs. bot-contained vs. bot-to-human handoff volumes, so that shifts in AHT and CSAT can be interpreted correctly as the mix evolves.
- Rolling workforce planning that continuously adjusts staffing levels, skill mix, and training investment as automation changes the volume and complexity profile. Particular caution was flagged about hollowing out the entry-level pipeline: today's Tier 1 agents are tomorrow's Tier 2 specialists.
What this means for your contact center
The contact center leaders in this session weren't pessimistic about AI. The general consesnus was that AI is delivering real improvements in onboarding speed, agent productivity, customer experience, and operational scalability, but only for organizations willing to do the harder work of redesigning their measurement frameworks, governance models, and operational rhythms alongside it.
The KPIs you trust today may already be telling you less than you think. The question isn't whether to adapt, it's how fast you can.
