AI takes the easy support tickets. The human queue gets harder.

Decagon says Assist works inside Salesforce, Zendesk and Front rather than sending representatives into another application. When a Decagon customer-service agent escalates a conversation, the human can receive a summary of what happened, a drafted response and an action such as processing a return or updating an account. Voice calls can include live transcription and guided next steps; chat can include real-time translation.

The company also added an analytics hub that breaks out adoption and outcomes by team, representative and type of work. Managers can see which suggested actions are used, where handle time changes and whether customer satisfaction moves.

Those are vendor claims from a launch announcement, not an independent result for every support team. Decagon quotes Aura’s vice president of customer experience saying an earlier version improved consistency, speed and accuracy across three channels. The public post does not publish Aura’s baseline, sample size or before-and-after numbers. A buyer still needs a pilot on its own queue.

Genesys surveyed 5,811 consumers and 1,560 customer-experience and business leaders for its 2026 State of Customer Experience report. Ninety-one percent of leaders said human representatives would remain critical three years from now, and 90% expected human interactions to become more complex or emotionally charged.

That shift breaks a lazy comparison. If AI resolves more routine questions, the average human case will take longer even when the representative is doing better work. A team can improve the customer journey while its raw average handle time rises. It can also lower handle time by rushing the cases that most need patience. The number alone cannot tell the difference.

Gartner found the same pressure from the customer side. In a survey of 3,566 B2B and B2C customers, 87% said access to a human was essential when a company uses generative AI for service. Half said GenAI had made service easier. Customers are not rejecting automation outright; they want it to work and let go when it does not.

For the first month, keep the old queue and the AI-assisted queue comparable by case type. Put order-status questions beside order-status questions and billing disputes beside billing disputes. If the assisted team receives a heavier mix, say so before anyone calls a longer handle time a regression.

Then follow the case past the first reply. Did the refund, cancellation or account change actually land? Did the customer return within a day or two? How often did the representative correct the AI summary, ignore the suggested action or reopen a source document because the suggestion was incomplete? Those edits are product evidence, not disobedience.

The representative should also be able to mark why a suggestion was skipped: wrong policy, missing account detail, bad tone, stale customer history or simply not useful for this case. Ten repeated skips in one category may reveal a broken knowledge article or action. A leaderboard of who clicked the assistant most often will reveal mostly who understood the incentive.

Priya Rao would split results by case type and compare the full outcome: resolution, repeat contact, customer satisfaction, human correction and time spent after the transfer. If the human queue contains harder work, the denominator has to show it. Otherwise the rollout can punish the people carrying the exceptions.

Cass Bell is wary of the rep-level adoption view. A dashboard that shows who uses or ignores AI can become a loyalty test long before anyone proves the suggestions are good. He would keep individual usage out of performance files during the pilot and treat a skipped suggestion as a question about the system first.

Priya is asking for better measurement. Cass is asking who that measurement can hurt. A sensible rollout needs both: enough detail to find the broken handoff, and a boundary that stops experimental software from quietly grading the employee.

An AI assistant for customer service should reduce searching, repeated context and after-call cleanup. Decagon Assist is aimed directly at that boring middle. The buying decision should still include what happens to staffing and expectations after routine volume drops.

Do not promise the same number of cases per hour from a queue that now contains more judgment, anger and exceptions. Adjust targets by case mix. Give new representatives practice on complete low-stakes cases instead of training them only to review difficult AI handoffs. Keep breaks and escalation support intact when the emotional load rises.

A better summary should leave the representative with more room for the hard call, spare the customer from starting over and keep the same unresolved case from coming back tomorrow.