The AI Boomerang: Why Cheaper Cognition Brings Work Back to the Business

AI lowers the minimum viable unit of work, bringing neglected customers, cases, reviews, and ideas within economic reach. The management job is turning new capacity into useful activity — not just fewer people.

In February 2024, Klarna announced that its AI assistant had handled 2.3 million customer conversations in one month — two-thirds of chats across 23 markets, in over 35 languages, resolved in under two minutes on average. Klarna described the volume as equivalent to 700 full-time agents and estimated a $40 million profit improvement for the year.

Then the boundary moved. In May 2025, CEO Sebastian Siemiatkowski told Bloomberg the cost-cutting push had gone too far. Klarna planned to recruit human customer-service workers again so customers could always reach a person. Quality, trust, and the option of human contact still carried commercial value.

Automation compresses a familiar body of work. The savings invite broader coverage and higher volume. Human work returns around a changed service model.

Some jobs still disappear. Roles dominated by routine cognitive work can shrink permanently, and the people displaced from those roles have no guaranteed path into the work that grows elsewhere. Executives should state that consequence plainly when they redesign an operation.

The minimum viable unit of work

Every company has useful work that it declines to do. A small customer gets a generic reply because a tailored review costs too much. The work would help, yet its likely value cannot justify a specialist's time. Companies make versions of that decision throughout their operations, usually without recording what they chose to leave undone.

AI lowers the minimum viable unit of work. It reduces the cost of the first useful pass through a document, question, case, or idea. Once that cost falls, work that previously required a specialist's uninterrupted time can begin cheaply and reserve human attention for the difficult parts.

Cashfree Payments, in an AWS-published case study, shortened merchant onboarding from over 24 hours to 10 minutes and resolved support tickets 70% faster using systems built on Amazon Bedrock. EXL's generative AI assistant reduced insurance underwriting from several days to a few hours and cut costs by up to 80%. Orion Health's Oribot made over 500,000 support records searchable in under a minute and reclaimed 50 support-team hours a day.

Productivity expands the operating surface

A controlled field experiment by Dell'Acqua and colleagues, summarized by MIT Sloan, gave consultants realistic assignments and compared participants who used GPT-4 with those who did not. AI improved skilled-worker performance by nearly 40% on tasks inside the model's capability boundary.

The researchers described a 'jagged technological frontier.' On tasks beyond the model's capability, AI users were less likely to reach the correct answer. Confident use of an unsuitable tool can lower performance. Management has to identify suitable tasks, test the boundary, and preserve human accountability where errors carry weight.

The available capacity usually finds a use. Service teams cover additional languages and hours. Compliance teams inspect a larger share of transactions. Sales teams qualify smaller prospects. Product teams run additional tests. Pfizer's collaboration with AWS reports search and extraction tools can save scientists up to 16,000 hours a year. PlanRadar's SiteView reduced visual documentation time by 90%, letting construction teams capture site conditions at a cadence manual photography made uneconomic.

Expertise raises the return on AI

General model capability becomes business performance through domain expertise. An underwriter knows which omission changes the risk. A support lead recognizes when a polite answer avoids the customer's actual problem. A scientist understands whether a retrieved finding applies to the current compound. A site manager can distinguish a harmless visual change from evidence of a costly defect.

Domain knowledge improves AI-assisted work at four points: framing the task, supplying context, evaluating the answer, and steering the process. Role design should keep people who understand customers, failure modes, and operating constraints involved in building the workflow.

Cheap work can create expensive problems

Lower cost removes a useful restraint. A compliance system can flag so many low-risk cases that reviewers lose sight of the important ones. Cheap output creates its own queue when nobody decides which activity deserves attention. Volume needs a value test.

Work intensification is the other risk. Managers may convert every minute saved into a higher target. Employees inherit larger queues, shorter deadlines, and the emotional burden of handling only the hardest cases. A support agent who receives a full day of angry escalations experiences the concentration of strain that automation can cause. Capacity planning should account for that density.

Map the work that becomes viable

Leaders can start with the work the company currently skips, delays, samples, or reserves for its largest customers. Pick one unit of work and measure its labor, delay, error rate, and expected value. Then ask:

• Which step makes this unit too expensive today?
• What evidence and context would AI need for a reliable first pass?
• Where does judgment, empathy, negotiation, or accountability remain human?
• How many additional cases become economically viable if the unit cost falls?
• What bottleneck appears when volume increases?
• Which outcome will prove the added activity is useful?
• Who owns quality, exceptions, workload, and customer recourse as the workflow scales?
• Which roles shrink, and what transition will the company offer the people affected?

Run the first deployment against a bounded set of real cases. Track cost per useful outcome, error severity, review burden, cycle time, employee load, and customer response. Watch for activity that rises without moving an outcome. Keep a human route where trust or material risk requires one.