The most consequential number in customer service is no longer how many tickets a rep can close in a day. It is the deflection rate, the share of incoming queries resolved without a human ever touching them. Industry estimates put that figure at roughly 65% in 2025, up from about 52% in 2023. Roughly 80% of companies now use or plan to deploy AI chatbots, against around 5% in 2020. The shift since the pandemic is one of the fastest tooling rollouts any frontline profession has seen.
That is the headline. The more interesting story is what it does to the work that remains.
The bifurcation of the queue
When a deflection layer pulls 65% of incoming queries off the floor, the queries that reach a human rep are not a representative sample of the old queue. They are the hard ones. Edge cases, angry customers, multi-product issues, things the bot tried to resolve and failed at. Reps using AI agents are reportedly handling about 13.8% more inquiries per hour, which sounds like an efficiency win, but the inquiries themselves are heavier on average than they used to be.
This produces a paradox most contact-centre managers are now living. Average handle time goes up even though efficiency tools are getting better, because the easy calls are gone. CSAT scores get noisier because the remaining contacts are by definition the harder ones to satisfy. Coaching has to be rebuilt around tier-2 work rather than tier-1 throughput.
The reps who thrive in this environment are the ones who can do something a bot cannot: take a tangled situation, work out what is actually going on, and reach a resolution the customer accepts. That skill was always present in the job. It is now the entire job.
What AI does sitting next to the rep
Two AI surfaces matter for a working rep. The first is the customer-facing layer (the chatbot, the voicebot, the help-centre search) which the rep does not control but inherits the consequences of when handover happens. The second is the agent-assist layer, which sits inside the rep’s console.
Agent-assist tools from Salesforce, Zendesk, Intercom, and Gorgias now do a few things consistently well: surface the right knowledge-base article based on the live conversation, draft a response the rep can edit and send, summarise the customer’s history before the rep opens the ticket, and flag sentiment when a call is heading sideways. The reported time savings cluster around 20% less time on routine cases and roughly 45% off call time on average for teams that have integrated these tools properly. Those numbers depend heavily on the integration, the knowledge base, and the rep’s willingness to actually use the suggestions rather than write from scratch.
The failure mode worth naming: reps who copy AI suggestions verbatim without checking them. This is how customers receive confidently wrong information about return policies, warranty terms, or account details. Every contact centre that has rolled out agent-assist at scale is now grappling with the same QA question, which is when a hallucinated response counts as a coaching issue, a tooling issue, or a policy gap.
The job market underneath the trend
Gartner has projected up to USD 80 billion in contact-centre labour cost savings by the end of 2026 from AI deployment. Whether that figure lands or not, the directional pressure is real. Big retailers, telcos, and BPOs have been quietly reducing tier-1 headcount through attrition rather than layoffs, holding hiring flat while volumes grow.
The reps still being hired are not the same reps as five years ago. The job description has shifted toward problem-solving aptitude, written communication, and comfort with multiple tools running at once. Entry-level roles that consisted of reading a script have largely gone to the bot. The reps who remain are doing work that looks more like account management or technical support than traditional call-centre service.
The pay is following the work. Senior CX specialists, retention specialists, and complex-case agents are paid meaningfully better than the tier-1 rep job they evolved out of. The bottom of the pyramid has thinned and the middle has thickened.
What the next twelve months change
Three shifts are visible in the data and the rollout pipeline.
First, voice deflection finally catches up to chat deflection. The voicebots of 2026 are using LLMs rather than the old IVR decision trees, which means they handle more nuanced spoken queries. The contact centres that historically routed voice to humans by default will start to deflect voice at rates closer to chat.
Second, AI-driven QA replaces sample-based QA. Instead of supervisors listening to 2% of calls, every interaction gets scored automatically against rubrics. This will reshape how reps are coached and, in some organisations, how they are paid.
Third, the workforce model splits cleanly. There will be a small high-skill rep population working complex cases, and a much smaller bot-supervision layer monitoring AI conversations and stepping in when needed. The middle, the generic frontline rep working a queue of routine queries, is the part of the job AI is genuinely automating.
The reps building career durability are doing two things in parallel: getting fluent with the tools so they can move faster than peers, and building the diagnostic and communication craft that makes them the human a customer is happy to be transferred to. Both matter. Neither is optional anymore.