AI in contact centres: productivity, service quality and worker transition

What measured support research and company reports tell us—and what they cannot establish about jobs.

A contact centre is a useful place to examine AI because its work produces records that can be checked: a customer arrives with a problem, someone attempts to resolve it, and the customer may return if the answer fails. Yet the same industry also attracts sweeping claims about disappearing jobs. A faster conversation, a task that software can perform and an employee who has lost a job are different observations. Managers, workers and readers need to keep them separate.

What the support study measured

Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied the staggered introduction of a conversational assistant using records from 5,172 customer-support agents. The revised paper reports an average 15% improvement in issues resolved per hour. Results differed across workers: less experienced and lower-skilled agents improved speed and quality; highly skilled agents saw smaller speed gains and small quality declines. The public author manuscript was revised in November 2024 and the journal issue is dated May 2025.

These findings concern a particular assistant, workforce and operating setting. They establish neither a universal uplift for every BPO nor an automatic staffing reduction. Before using the result in a business case, check whether your own workers handle similar questions, have comparable knowledge support and can refuse an unsuitable suggestion. A voice workflow involving identity checks or a disputed payment is also different from a text-support workflow.

What a company's capacity claim means

OpenAI's Klarna case study describes the company's early assistant rollout in February 2024. It reports work equivalent to 700 full-time support agents and an estimated $40 million improvement in profit for 2024. Those are attributed company/vendor claims. The estimate was a forecast in that account; it should not be silently relabelled an independently audited saving. Equivalence to a quantity of work is also not documentation that the same number of people were dismissed because of AI.

For a reader investigating workforce changes, the missing evidence matters. Look for dated headcount disclosures, affected locations, redeployments, hiring restrictions and the employer's stated reasons. A company's total headcount can change through acquisitions, seasonal demand, attrition or reorganisations. It takes more evidence to isolate the effect of automation than to report a smaller number of employees.

Klarna’s later FY2025 annual filing provides a separate company-reported outcome: it estimates $39 million in 2024 cost savings and says its assistant handled 80% of customer-service chats during 2025. These figures are management estimates and internal service-log reporting, not an independent causal audit. Cost savings and the earlier profit forecast are different measures and should not be treated as a like-for-like comparison.

Exposure is a starting point for research

The ILO's May 2025 update estimates that one in four workers worldwide holds an occupation with some exposure to generative AI. Its assessment is based on tasks and distinguishes degrees of exposure. The brief emphasises continued human input and the possibility of work changing rather than simply vanishing. It is not a count of redundancies that have already happened, and it does not tell a particular contact centre how many posts to remove.

Use exposure research to identify tasks worth examining: finding a policy, summarising a conversation, translating a draft or suggesting the next question. Then inspect the duties around those tasks. Responsibility for an exception, checking a customer's circumstances and explaining an unpopular decision may remain with a person even when the first draft is automated.

An illustrative workflow to test

Consider a support team dealing with delivery enquiries. Begin with read-only assistance: retrieve the approved delivery policy and draft an explanation. Let the agent review it before sending. Keep address changes, refunds, suspected fraud and exceptional promises outside the pilot. This is an illustrative operating design, not a description of the companies in the cited research.

Collect a baseline for the same types of cases and compare more than speed. Count first-contact resolutions, repeat contacts, escalation, incorrect promises and the minutes people spend checking suggestions. Include difficult conversations and different language needs. A quick answer that makes a customer call again can shift work rather than remove it.

Record how newer and experienced agents use the tool. If a suggestion appears convincing but conflicts with the policy, can someone explain and correct it? Can a worker decline it without a performance penalty? Keep a clear route to human help. The deployment should leave a usable record of what the assistant proposed and what the responsible person chose, with customer information handled under the organisation's applicable policies.

Decisions for workers and organisations

The OECD's workplace research associates training and worker consultation with better outcomes. That association is useful guidance, not proof that a particular course will prevent every displacement. Give workers time to learn the changed workflow, invite them to identify failure patterns and make the responsibilities of supervision explicit.

Organisations should publish their own evidence before making sweeping claims: which task improved, over what period, for which cases, at what full operating cost and with what quality trade-offs. Workers can ask for the same details when a role changes. The questions are practical: what work is being removed, what work is being added, who carries the risk, and what support is available for the transition?

This article is educational analysis. It does not endorse a vendor, establish a worldwide AI-attributed job-loss total or guarantee a financial return. Source dates describe the evidence; the publication date describes this analysis.