Contact Centre Agent Assist: One Large Study, Gains Concentrated Among Novices

The use case

Between autumn 2020 and winter 2021, a Fortune 500 software company switched on a tool that read each technical-support chat as it happened and proposed replies and links to internal documentation. The agent could accept, edit or ignore every suggestion¹. This is agent assist: AI that sits beside a human agent, not in front of the customer.

The scope here is narrow. AI means a machine-learning language model. The study tool was a large language model fine-tuned on past conversations labelled with outcomes such as resolution and handling time, and trained to favour empathy and documentation¹. Older technology does similar work and is not credited to AI: keyword search of a knowledge base, scripted prompts, rules-based next-best-action and manual quality scoring. Customer-facing chatbots and fully autonomous agents are out of scope.

Where it is running

No source opened gives the share of contact centres that run machine-learning agent assist. The named deployments below are the ones the evidence supports.

Gartner surveys show pressure rather than deployment. In a survey of 321 service leaders in October 2025, 91% reported executive pressure to implement AI (self-reported, all AI)⁷. In a survey of 265 leaders in April and May 2025, 75% reported larger AI budgets (self-reported, all AI)⁶. Neither gives an adoption rate for agent assist.

What the evidence shows

Average effect. Access to the tool raised issues resolved per hour by 15% on average, or 0.30 chats against a pre-treatment mean of about two (measured without a randomised control group, staggered rollout, 2020-21)¹. The gain came from shorter chats: handle time fell 3.7 minutes, 8.5% from a 43-minute baseline¹. Resolution rates rose 1.3 percentage points, which the authors call modest and insignificant, and the customer score showed no meaningful change¹.

Who gains. The gain is concentrated in the least experienced agents, and the most skilled agents saw no productivity gain. The lowest-skill quintile resolved 36% more chats per hour; agents with under a month of tenure gained 0.7 resolutions per hour; agents with more than a year gained nothing measurable (measured without a randomised control group, 2020-21)¹. The top-skill group showed small but statistically significant falls in resolution rates and customer satisfaction¹.

Why novices gain. The system was trained on text from agents the firm rated as top performers, so it hands newer agents the phrasing and documentation those agents use¹. The authors found adherence to suggestions averaged about 35 to 38% (the paper gives both figures), and that returns were higher where agents followed suggestions more¹.

Independent replication. A randomised field experiment at Alibaba, where agents were randomly assigned access to a drafting assistant, reports faster service and better customer ratings, no significant effect on customer retrials, gains concentrated in lower performers, and declines in quality for top performers (abstract only, randomised trial, preprint, July 2026)². No sample size or effect size could be read, so the direction is usable and the magnitude is not.

Randomisation caveat for the first study. Its own randomised pilot covered about 50 workers, and the authors have no information on the control group¹. The 15% figure therefore rests on comparing staggered start dates.

What it takes: cost, integration, governance

Cost. No public price per agent or per contact was found in the sources opened. This is a gap, not a finding.

Integration and people. Rollout in the main study was limited by licence caps and run by managers through a three-hour online onboarding session¹. In Gartner's April and May 2025 survey the typical leader planned to add five full-time roles to manage AI investments (self-reported plan, all AI)⁶. In October 2025, 58% of leaders planned to upskill agents as knowledge managers who review and curate AI content (self-reported, all AI)⁷.

Rules. The EU AI Act recitals say AI systems that detect the emotional state of people in workplace situations should be prohibited, defined by inference from biometric data, and treat AI used in employment and monitoring of workers as high-risk⁹. Only the recitals could be read here; application dates and the exact wording of the operative articles were not retrieved, so a vendor feature that scores agent or caller emotion needs a legal check before use in the EU.

In the United States, the live risk is call-recording law. A federal court in San Francisco declined on 10 February 2025 to dismiss a class action under California's wiretapping statute against Google over its contact-centre AI service. The order describes the service as able to give a human agent a real-time transcript, article suggestions and smart replies, and records the allegation that callers are not told⁴. The judge applied a test based on a provider's capability to use call data for its own purposes⁴.

Where it falls short

Narrow base. The strongest result is one firm, one tool and one period, now almost six years old. The authors state that the paper is not designed to show aggregate employment or wage effects¹. No newer controlled estimate with readable numbers was found.

Skilled staff. Both the 2020-21 study and the 2026 Alibaba preprint report that top performers gain nothing or lose quality¹². The first study's authors warn that high adherence now may reduce the quality or diversity of solutions used to train the tool later¹. The tool's value is borrowed from people it does not replace.

Customer outcomes. Customer scores did not move in the main study¹. The effect is faster handling and an easier job for novices, not proven better service.

Neighbouring result. A randomised preprint on supervised autonomous AI at Alibaba, a different design from agent assist, found chat duration fell 3.2% overall and 16.8% on AI-eligible chats, but customer ratings fell by 0.412 points on those chats (randomised field experiment, August 2024)³.

Legal. The Ambriz ruling decided only that the claims were adequately pleaded, not that any call data was misused; status after February 2025 was not checked⁴. No programme withdrawals were found.

Headcount. In a December 2025 Q&A Gartner reported that over 80% of leaders expect to reduce agent headcount within 18 months (self-reported expectation, all AI)⁸. The February 2026 release of the same survey frames the figure as moving agents into new roles⁷. The two framings should not be merged.

Phase: In early use — why

The hypothesis was In wide use. The evidence confirms machine-learning agent assist at production scale in one large firm and in a randomised trial at Alibaba, but no source states how many organisations run it. Because the share of genuine AI use cannot be established, the lower phase applies.

What to watch next

• Full text of the Alibaba agent-assist preprint (version of 31 July 2026), which would add sample size and effect size².
• Any merits ruling or class-certification decision in Ambriz v. Google, 23-cv-05437⁴.
• Commission guidelines and the operative articles of the AI Act on emotion inference at work, which were not read in this research⁹.
• A controlled study that reports customer-outcome and cost data after 2024.

Questions to ask before adopting

• Which of our agents are novices, and how would gains for them be separated from effects on our best performers?
• What was the tool trained on, and whose conversations count as the gold standard?
• Do callers and agents know a third party transcribes and analyses calls, and does local recording law allow it?
• Does any feature infer emotion, and is that lawful for staff and for customers in each country?
• What is the full cost per agent, including onboarding, knowledge upkeep and review roles, against measured handle-time savings?