Escaping the Trap of Hollow Gains: How We Can Build Long-Term Value through AI

By Nishanth Bhimireddy, Hamed Mehrabi, and Riccardo Vecchiato

AI strengthens or weakens firms depending on integration, culture, and whether leaders track its impact on capabilities and performance.

Consider two companies that used AI to enhance customer service. At a Fortune 500 software firm, an AI assistant suggested responses to support employees. The number of issues resolved per hour rose by 15% on average, most of all among newer staff, and the evidence suggests that employees who had used the assistant for some months remained faster even during system outages, when the assistant was unavailable.¹ However, Klarna took a different route. It announced that its AI assistant was doing the work of hundreds of full-time agents.² Little more than a year later it was hiring people again, and its chief executive conceded that cost had weighed too heavily in the delivery of customer services and that quality had suffered.³

The health sector provides similar evidence. In Sweden, AI-supported breast screening found more cancers without more false alarms and substantially reduced radiologists’ reading workload.^4,5 In Poland, experienced endoscopists who had become used to AI assistance detected noticeably fewer precancerous growths when they worked without it.⁶

AI can clearly improve today’s results. Whether it also strengthens or weakens the capabilities that produce those results is a separate question, and the answer depends largely on how AI is applied.

What AI can quietly draw down

An organisation’s capabilities are the things it can reliably do well. They might include identifying and solving customer problems, maintaining consistent product quality, developing new products, building strong customer relationships, or making sound operational decisions under pressure. Capabilities are built up over time through repeated use, accumulated experience and shared knowledge, and they decline when they are no longer exercised.⁷ They also cannot be rebuilt quickly: money cannot compress years of experience into months.

AI changes the process through which capabilities are built and renewed. A capability rests on a combination of people, routines, knowledge and culture, and the use of AI changes that combination. When AI is introduced on the assumption that it can simply replace some of these elements, essential ingredients of the capability-building process can go missing. Output (i.e., today’s results) can hold steady or even improve while the capability behind those results weakens. A study of an accounting firm that relied on automated software for several years found that it had lost the ability to perform a key accounting process without the software, something neither staff nor managers noticed until the software was replaced.⁸

The pipeline is also at risk. Much of an organisation’s expertise is renewed as new employees do routine work alongside experienced colleagues. When AI absorbs that work, junior staff lose the practice ground where judgement is formed, and firms may find years later that too few people can oversee the systems that replaced them.^9-11

The result is what we call “hollow gains”: output that looks healthy today while a capability deficit builds beneath it. Klarna’s reversal shows what can happen when that deficit is recognised (too) late.

Others have warned that AI can erode the expertise firms compete on, and half of senior executives in one recent survey already see de-skilling in their organisations.¹² Still, there is another side of the ledger and less attention has gone to it.

The other side of the ledger

AI can also strengthen capabilities. The customer-support assistant in the Fortune 500 software firm appeared to spread the practices of the most capable employees to the less experienced ones, and the gains held when the tool was unavailable.¹ Likewise, AI supports capability development when it creates an enabling context (e.g., providing advice and feedback).¹³

AI also changes which capabilities matter. Firms that adopt it more intensively advertise more managerial roles, and increasingly seek stakeholder management, creativity and growth-oriented skills rather than routine administrative ones.¹⁴

Whether AI builds or erodes capabilities depends on how it is integrated into the processes, routines, and knowledge of a company.

A tension to manage, not a choice to make

Today’s results draw on capabilities, while the capability-using process renews capabilities themselves and enhances long-term results. With AI, however, they can pull against each other. The quickest way to raise today’s output is often to remove the very processes and routines through which capabilities are built. More important, while AI’s effect on output is visible almost immediately, its effect on capabilities emerges only over time. That is why the impact of AI on capabilities is so easily neglected.

Paradox research suggests that such tensions are best managed by attending to both demands over time, not by choosing one.¹⁵ Studies of AI make a similar point: firms that lean too far towards automation, often through bypassing established routines, can erode capabilities they later need.^16,17  

Design for both

We propose that leaders treat every significant AI addition as having two entries: one for output and one for the capabilities behind it. Managing both starts with design.

First, map how the capability is built. Before AI enters a process, identify what the capability relies on: which people, routines, knowledge and relationships sustain it, and where new employees develop judgement. Much of this appears in no process map, which is why AI integration can remove it without anyone consciously deciding to do so.¹⁰

Second, design the integration to serve both entries. Some organisational choices give output and capability building their own time or place. People attempt a task before turning to AI. Teams regularly complete selected work without AI so that core capabilities remain in use. Difficult cases, together with a proportion of routine ones, remain with people so expertise continues to develop.^10,12,17 Other organisational choices integrate the two, so that improving output also builds the capability. AI can suggest how to respond, explain why a recommendation has been made, or direct people to relevant knowledge so that they absorb good practice while working, as in the customer-support study of the Fortune 500 software company.^1,18

Monitor both, and redesign your solution

No solution is perfect, and AI’s effects change as tools, people and work evolve. The third step is therefore to monitor both entries. Output is easy to track. Capabilities need their own test, and the examples at the beginning of this article point to a simple one: how effectively can the work be performed when AI is removed from the process? Periodic AI-off assessments, similar to the way pilots continue to practise manual flying, are not intended to replace AI. Instead, they help organisations determine whether the underlying capabilities are being maintained, strengthened, or gradually eroded.

Reading the two entries together gives four possible positions, each calling for a different response (Figure 1). Compounding is the goal: AI raises output and strengthens the capability behind it. Borrowing is the position to watch most closely because hollow gains can mask a growing capability deficit. Investing, where capability grows before output does, is typical early on and requires patience. Losing calls for a rethink. The readings then feed back into design, so that AI integration becomes a continuing cycle rather than a one-off decision. Each significant integration also needs a named owner accountable for the capability, not only the system.

Figure 1. Reading the ledger: what AI is doing to output and capability

What leaders should do

• Treat AI integration as capability design, not tool selection. Map how a capability is built before AI changes it and design the integration to support output and capability building together.
• Monitor both entries and redesign. Give each significant integration a named owner and a regular AI-off check, read its position on the matrix, and feed what is learned back into design.
• Sequence adoption. Start where mistakes are cheap and learning is rich and extend AI to core capabilities once both entries can be seen clearly. Klarna appears to have done the opposite, letting AI take over much of a core customer relationship while freezing hiring. For a time, the output metrics looked strong; meanwhile, the underlying capability was being overlooked. In Europe, the AI Act’s human-oversight obligations for most standalone high-risk systems apply from December 2027, adding a regulatory reason to keep the capability to oversee AI in-house. ^19,20

The most dangerous AI implementation is not the one that fails visibly. It is the one that delivers hollow gains: succeeding today while quietly eroding the capabilities needed tomorrow.

Whether AI strengthens or weakens organisational capabilities is a design choice, and firms can manage it only if they monitor both sides of the ledger.

About the Authors

Nishanth Bhimireddy is Lecturer in International Business and Strategy at Kingston University, London. His research both studies and applies new digital technologies, using AI, machine learning and natural language processing to analyse online discourse. He previously worked on large-scale enterprise systems at Toyota North America and Western Union. 

Hamed Mehrabi, PhD, is Senior Lecturer in Marketing at Kingston University London. His research focuses on how organisations build capabilities and leverage artificial intelligence, analytics, and innovation to enhance performance and competitiveness. He has published in leading journals and engages in industry-focused research and consultancy in innovation and technology management. 

Riccardo Vecchiato is Professor of strategic management at Kingston University London, where he leads the research hub on Foresight, Creativity, and Future of the Work. His research addresses decision making under uncertainty and technology management. On these themes, he publishes in leading journals and acts as consultant in major corporations. 

References

1. Brynjolfsson, E., Li, D., & Raymond, L. (2025). “Generative AI at work”. Quarterly Journal of Economics, 140(2), 889–942.

2. Klarna. (2024, February 27). Klarna AI assistant handles two-thirds of customer service chats in its first month [Press release]. https://www.klarna.com/international/press/klarna-ai-assistant-handles-two-thirds-of-customer-service-chats-in-its-first-month/.

3. Ivanova, I. (2025, May 9). As Klarna flips from AI-first to hiring people again, a new landmark survey reveals most AI projects fail to deliver. Fortune. https://fortune.com/2025/05/09/klarna-ai-humans-return-on-investment/.

4. Gommers, J., Hernström, V., Josefsson, V., Sartor, H., Schmidt, D., Hjelmgren, A., … & Lång, K. (2026). Interval cancer, sensitivity, and specificity comparing AI-supported mammography screening with standard double reading without AI in the MASAI study: a randomised, controlled, non-inferiority, single-blinded, population-based, screening-accuracy trial. The Lancet, 407(10527), 505-514.

5. Lång, K., Josefsson, V., Larsson, A. M., Larsson, S., Högberg, C., Sartor, H., … & Rosso, A. (2023). Artificial intelligence-supported screen reading versus standard double reading in the Mammography Screening with Artificial Intelligence trial (MASAI): a clinical safety analysis of a randomised, controlled, non-inferiority, single-blinded, screening accuracy study. The Lancet Oncology, 24(8), 936-944.

6. Budzyń, K., Romańczyk, M., Kitala, D., Kołodziej, P., Bugajski, M., Adami, H. O., … & Mori, Y. (2025). Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study. The Lancet Gastroenterology & Hepatology, 10(10), 896-903.

7. Helfat, C. E., & Peteraf, M. A. (2003). “The dynamic resource-based view: Capability lifecycles”. Strategic Management Journal, 24(10), 997–1010.

8. Rinta-Kahila, T., Penttinen, E., Salovaara, A., Soliman, W., & Ruissalo, J. (2023). “The vicious circles of skill erosion: A case study of cognitive automation”. Journal of the Association for Information Systems, 24(5), 1378–1412.

9. Beane, M. (2019). “Shadow learning: Building robotic surgical skill when approved means fail”. Administrative Science Quarterly, 64(1), 87–123.

10. Sudhir, K. (2026, June 19). “How to design agentic systems around the implicit rules that govern your company”. Harvard Business Review.

11. Kenny, G., & Pogrebna, G. (2026, April 1). “Don’t let AI destroy the skills that make your company competitive”. Harvard Business Review.

12. Goel, S., Martin, D., & Kaffe, C. (2026, June 17). “When everyone uses AI, companies risk losing critical skills”. Boston Consulting Group.

13. Hillebrand, L., Raisch, S., & Schad, J. (2025). “Managing with artificial intelligence: An integrative framework”. Academy of Management Annals, 19(1), 343–375.

14. Alekseeva, L., Azar, J., Giné, M., & Samila, S. (2026). “Artificial intelligence adoption and the demand for managerial expertise”. Strategic Management Journal. https://doi.org/10.1002/smj.70099

15. Smith, W. K., & Lewis, M. W. (2011). “Toward a theory of paradox: A dynamic equilibrium model of organizing”. Academy of Management Review, 36(2), 381–403.

16. Raisch, S., & Krakowski, S. (2021). “Artificial intelligence and management: The automation–augmentation paradox”. Academy of Management Review, 46(1), 192–210.

17. Yang, T. (2026). “The paradox of efficiency: Institutional interfaces, residuals, and the erosion of innovation drivers in AI-augmented organizations”. Frontiers in Artificial Intelligence, 9, 1864238.

18. Armstrong, B., & Shah, J. (2026). Humans in the loop: The evolution of work in early experiments with generative AI [Industry report]. MIT Industrial Performance Center. https://ipc.mit.edu/wp-content/uploads/2026/04/Humans_in_the_Loop_full_r01M.pdf.

19. Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act), Official Journal of the European Union, L series, 12 July 2024. http://data.europa.eu/eli/reg/2024/1689/oj

20. Hickman, T., Lorenz, S., Hainsdorf, C., & Teetzmann, C. (2026, August 4). EU AI Omnibus enters into force, amending the AI Act. White & Case. https://www.whitecase.com/insight-alert/eu-ai-omnibus-enters-force-amending-ai-act.