The AI Productivity Paradox in the Contact Centre

AI is improving productivity. So why can agents look less productive?

AI is supposed to make the contact centre more productive, and in many cases it does. But as AI takes on more routine work, I think we may be measuring some of that success the wrong way.

As automation handles simpler enquiries and Agent Assist removes repetitive tasks, the work that still reaches an agent increasingly consists of the difficult stuff: exceptions, complaints, complex billing issues, retention conversations and problems that cross multiple systems. The contact centre can therefore become more productive overall while the agents handling what remains appear less productive on the dashboard. That is what I think of as the AI productivity paradox.

AI really can improve agent productivity

Good evidence shows that generative AI can improve frontline productivity. Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied more than 5,000 customer-support agents using a generative AI assistant and found productivity rose by about 14% on average, measured by issues resolved per hour. For novice and lower-skilled workers, the gain was around 34%.

What interests me most is why. The researchers found that AI helped spread the practices of stronger agents, surfacing knowledge and behaviours in real time that once took months of experience to develop. Agent Assist can therefore do more than save a few seconds; it can help newer agents move closer to experienced-agent performance much faster.

But AI is changing the work humans receive

The first contacts organizations tend to automate are predictable and high volume: password resets, order-status enquiries, appointment changes, basic account questions and straightforward transactions. Those interactions also used to give agents a natural mix of easy and difficult work.

Remove the simpler contacts and the workload changes. Agents increasingly receive exceptions, failed self-service journeys, emotionally difficult conversations and problems requiring judgement or several systems at once. Industry reporting has started to identify this shift directly: as routine issues are automated or deflected, people are left with a greater share of the complex and emotionally charged work. Contact volume may fall, but complexity moves toward the people who remain in the loop.

The measurement trap

Consider a simple example. Before automation, a contact centre handles 1,000 interactions, all handled by agents at an average handle time of six minutes.

Now AI resolves the 400 simplest interactions, leaving 600 for agents. Because those 600 were always the harder ones, average handle time rises to eight minutes. The dashboard reports that human AHT has jumped 33% and interactions per hour have fallen, so if you look only at the traditional agent dashboard, productivity appears to have deteriorated.

But the organization has eliminated 400 human-handled interactions and concentrated its people on the work that actually needs them. The agents did not suddenly become less capable; the mix of work changed. We are comparing today’s more complex human workload with yesterday’s blended workload and risk calling the difference a performance problem.

AHT still matters, but it needs context

I am not suggesting Average Handle Time should disappear. AHT remains useful for forecasting, capacity planning, queue modelling, costing and spotting operational anomalies. The danger is treating it as a simple measure of individual productivity.

Picture two retention calls. One agent finishes in eight minutes and the customer leaves; another spends fourteen minutes understanding the issue, finds the right solution and keeps the customer. Which interaction created more value? As AI takes over predictable work, time becomes an increasingly incomplete proxy for productivity.

WFM has the same challenge

Historical workforce-management data increasingly reflects a blend of work that still needs a person and work that can now be automated. If AI removes a meaningful share of simpler demand, last year’s blended AHT and workload patterns may no longer be a good predictor of next year’s human workload.

A 20% reduction in contacts does not automatically translate into a 20% reduction in human capacity. The interactions that remain may take longer, require different skills and vary more, so forecasting has to consider not only how much demand exists but what kind of demand reaches people after automation has taken its share.

AI may also change what an expert agent is

If Agent Assist can distribute the knowledge and practices of experienced employees, the value of expertise starts to shift. Knowing where to find an answer or memorizing procedures may matter less, while judgement, empathy, negotiation, exception handling, complex problem-solving and knowing when to override an AI recommendation may matter more.

That could reshape recruitment, training and coaching, while also changing the role of senior agents. AI can capture and distribute their knowledge, freeing them to focus on situations where experience genuinely matters.

There is an employee experience question too

There is another issue worth watching. If AI keeps removing easier interactions, an agent’s day could become a steady stream of exceptions, escalations and emotionally demanding conversations. AI may reduce the quantity of human work while increasing its intensity.

We do not yet have enough evidence to say this is happening uniformly across the industry, so I would treat it as an operational risk rather than an established outcome. But if customer-service shifts increasingly consist of the hardest conversations, organizations may need to rethink scheduling, breaks, coaching and how workload is measured.

A better productivity dashboard

I would keep AHT, after-call work, occupancy and service level because operations still need them. But I would add four additional lenses:

Complexity: What intent was involved? How many systems or actions did the interaction require?

AI contribution: Was the interaction resolved autonomously, AI-assisted or handled entirely by a person? Were AI recommendations accepted, edited or ignored?

Outcome: Was the customer's objective resolved? Did they contact us again? What happened to CSAT, retention or another relevant business measure?

Economics: What did the successful outcome cost across AI, workflow and human resources?

A complex interaction can cost more and still be the better outcome, while a cheap automated interaction that fails and generates repeat contacts may be poor economics. The goal should be to use the most appropriate resource, human or AI, to produce the right customer and business outcome.

The real risk

The pattern is becoming clearer. AI automates simpler interactions and Agent Assist accelerates routine parts of human ones, leaving employees with a greater share of complex work. Traditional human productivity metrics can then move in the wrong direction even while overall operating productivity improves.

That does not make the old metrics useless; it means their context has changed. As automation absorbs predictable work, the human role becomes more concentrated around judgement, empathy, exceptions and difficult problem-solving, which are exactly the parts of customer service that are hardest to measure with a stopwatch.

Perhaps the better question is not simply how many interactions an agent handled per hour, but whether the combination of AI and people produced better customer outcomes, at the right cost and with the right experience.

The real risk is not that AI will make our agents less productive. It is that we will use yesterday’s metrics to measure tomorrow’s work.

I’d be interested to hear what others are seeing. Are your contact centre metrics evolving as AI changes the work reaching agents?

Sources and research notes

Brynjolfsson, Li & Raymond, “Generative AI at Work.” NBER Working Paper 31161; published in The Quarterly Journal of Economics (2025). The working-paper results report 5,179 customer-support agents, about a 14% average productivity improvement, and about 34% for novice and lower-skilled workers.

NBER Digest, “Measuring the Productivity Impact of Generative AI.” Summary of the customer-support study and the larger gains for less-experienced workers.

NiCE, “How AI is redefining workforce management.” 2026 analysis of capacity planning for a hybrid human and AI workforce; reports CMSWire research that 76% of contact centre agents handle multiple channels most or all of the time.

CX Dive, “Agents are overloaded. AI often makes it worse, experts say.” The shift toward more complex and emotionally charged human-handled work as routine contacts are automated.

Aspect, “How Do You Forecast a Contact Center Where AI and Humans Share the Work?” How AI deflection changes the composition, duration and variability of the work reaching humans.

Amazon Science, “Contact Complexity in Customer Service.” Research on defining and predicting contact complexity and matching complexity to agent capability.