Innovation & Tech

AI Resolved 15% More Issues Per Hour — Until It Had to Act Alone

AI-assisted agents closed 15% more issues per hour, but Alibaba's own data shows lower customer ratings when AI runs solo. Human oversight remains the deciding variable.

By Olivia Hart

4 min read

Best Supporting Agent: Why Agentic AI in the Contact Centre Still Needs a Human Lead
Best Supporting Agent: Why Agentic AI in the Contact Centre Still Needs a Human LeadMV-Fotos / pixabay

What's News

  • A study of over 5,000 support agents found AI assistance raised resolutions by about 15% per hour, with the largest gains for inexperienced agents.
  • In 2026, Alibaba's customer service operations recorded shorter AI conversations but lower customer ratings in AI-eligible interactions.
  • Recommended guardrails include approval limits for high-value refunds and automatic escalation when data conflicts.
  • The EU AI Act generally bans workplace emotion detection by AI, with call centres cited as an example by the European Commission.

AI-assisted agents resolved about 15% more customer issues per hour than unassisted colleagues, according to a study of more than 5,000 support workers. That headline number, however, came from a setup where people still made every decision — a detail that matters as contact centres hand agentic AI the authority to act on customers' behalf.

Agentic AI marks a shift from answering questions to taking action. Given the right permissions, these systems can retrieve account records, check eligibility and update a booking across multiple systems without waiting for an agent. For contact centres pressured to handle more conversations with fewer resources, the appeal is obvious. The risk is equally obvious: someone must define what happens when a customer needs an exception, or when an automated decision is technically correct but plainly unfair.

What does the evidence actually show?

The 15% productivity gain is real, but its distribution matters. The biggest improvements went to less experienced agents, who benefited from helpful suggestions appearing while they worked. A well-designed assistant also changes the agent's role before the conversation even starts — surfacing and summarizing prior interactions instead of forcing the agent to spend minutes hunting for context.

Yet the study's fine print deserves attention. As the analysis itself puts it: "The AI was helping agents do their jobs, with people still making the decisions." Whether the gains survive full automation is a question the data does not answer.

Speed is not the same as service. An AI agent might resolve a billing inquiry in seconds, but if it misreads the customer's situation or applies the wrong policy, that quick resolution turns into another complaint.

Where does human judgment still matter?

Consider a billing dispute. The AI checks the account and confirms the charge matches the contract. But the customer was promised a credit after a service outage, and that promise was never entered into the billing system. The AI answers from available records. An experienced agent recognizes something is missing and investigates.

The cautionary evidence is already in. In 2026, Alibaba's customer service operations showed that AI shortened conversations but received lower customer ratings in AI-eligible interactions. The research also found that the effectiveness of human intervention depended on when agents stepped in and what had gone wrong. Knowing when to involve a person, in short, matters as much as knowing what AI can handle.

Handoffs deserve equal scrutiny. Customers should not repeat their concerns at every transfer, and agents need visibility into what the AI has already done — especially when the caller is already frustrated.

How should centres set boundaries?

The recommended framework is built on five practical measures:

  • Define approved tasks: let AI independently handle items such as order-status checks or appointment updates.
  • Set approval limits: require agent authorization for high-value refunds, account changes or policy exceptions.
  • Establish escalation rules: involve an agent automatically when information conflicts, a request falls outside approved procedures, or the system cannot answer reliably.
  • Control data access: give AI only the customer data needed for the approved task.
  • Review performance regularly: monitor incorrect responses and complaints to locate missing safeguards.

Preparing a refund recommendation carries less risk than issuing a large credit without approval. Those limits should be reviewed and adjusted as the technology performs in real interactions, so automation can expand without eroding accountability.

What should contact centres actually measure?

A centre can celebrate lower average handling time while customers quietly lose confidence. If someone calls back the next morning because an AI agent promised a change that never happened, the first conversation was not a success. Reporting should make that visible: repeat contacts, reopened cases, correction rates, complaints and how often a human must reverse an automated action.

Frontline agents are an underused source of operational intelligence. They hear the awkward pauses. Give them a quick way to flag errors — and show them what changed as a result. If every correction disappears into a generic feedback form, the team will stop bothering.

There is also a legal ceiling on monitoring. The EU AI Act generally prohibits using AI to identify agents' emotions in the workplace, with limited exceptions, and the European Commission has cited call centres as an example. Businesses should check regional law before adopting tools that claim to detect emotion through a worker's voice.

The fully automated contact centre makes a neat presentation slide. Actual customers change their minds, misunderstand instructions and bring circumstances no system may judge fairly. AI's success, on this evidence, still depends on how well it supports agents — not on replacing them.

Source: unite.ai

Share this article:

More from Olivia Hart

Olivia Hart

Show full bio

Staff writer covering people, launches and day-to-day practice in entrepreneurship, startups and venture funding at Business Bearings.

2,535 articles

Related articles

« Previous article