Novice support agents

Introduction

One of the clearest real-world demonstrations of AI helping novice workers came not from a futuristic robotics lab, but from an ordinary customer-support department. Researchers studying more than 5,000 support agents at a Fortune 500 software company found that a generative AI assistant improved overall productivity by around 14%. But the average hid the most important result: the biggest gains went to newer and lower-skilled agents, whose performance improved by roughly 34%, while top veterans saw little benefit. NBER [OUP Academic]academic.oup.comOUP AcademicGenerative AI at Work* | The Quarterly Journal of Economicsby E Brynjolfsson · 2025 · Cited by 3397 — Finally, we provide evi…

That finding matters because it points toward a broader possibility behind the idea of AI-enabled human flourishing. The most important economic effect of AI may not be replacing experts outright, but making useful expertise easier to access. In this case, the AI system effectively turned years of accumulated workplace experience into live guidance for beginners. Instead of struggling through trial and error, newer staff gained something closer to an on-demand coach.

The study has become central to debates about whether AI assistants can genuinely widen access to capability and know-how — or whether the benefits mainly flow to already elite workers.

What the Fortune 500 support study measured

The best-known evidence comes from the paper “Generative AI at Work” by Erik Brynjolfsson, Danielle Li, and Lindsey Raymond, later published in The Quarterly Journal of Economics. The researchers analysed the staggered deployment of a conversational AI assistant across thousands of customer-support agents. NBER [OUP Academic]academic.oup.comOUP AcademicGenerative AI at Work* | The Quarterly Journal of Economicsby E Brynjolfsson · 2025 · Cited by 3397 — Finally, we provide evi…

The system monitored live customer chats and suggested responses, troubleshooting advice, and conversational phrasing in real time. Unlike a simple chatbot script, it was trained on huge numbers of past customer interactions and learned which approaches tended to solve problems successfully. [brc.org.uk]brc.org.ukgenerative ais impact on customer serviceGenerative AI's impact on customer service5 Oct 2023 — The study examined the staggered deployment of a chat assistant for a Fortune 500… [NBER]nber.orgNBERGenerative AI at Workby E Brynjolfsson · 2023 · Cited by 3217 — Access to the tool increases productivity, as measured by issues reso…

The headline numbers were striking:

Importantly, this was not a laboratory experiment with students performing toy tasks. It was a live workplace deployment inside a large company handling real customer problems under commercial pressure. That made the findings unusually influential in the wider debate over AI and work.

The results also complicated common fears about automation. The AI system did not mainly help the strongest employees become superhuman. Instead, it narrowed part of the gap between newcomers and veterans.

Why beginners improved more than veterans

The central explanation is that beginners lacked access to tacit knowledge — the hard-to-document practical judgement accumulated through experience.

In customer support, much of the job is not contained in formal manuals. Experienced agents gradually learn:

Veteran workers often use these habits automatically without being fully conscious of them. Economists and management researchers call this tacit knowledge because it is difficult to write down explicitly. [Forrester]forrester.comUnlock AI-Led Customer Service with Gen AI & TacitForresterUnlock AI-Led Customer Service with GenAI & Tacit…February 4, 2025 — 4 Feb 2025 — Explore how tacit knowledge and genAI trans…Published: February 4, 2025 [LinkedIn The AI assistant effectively acted as a transmission mechanism for that hidden expertise. Instead of forcing beginners to slowly absorb patte]linkedin.comHow AI is learning to overcome the tacit knowledge problemTacit knowledge includes the skills, intuition, and know-how that people pick u… rns over months or years, the system surfaced useful responses immediately during live interactions.

That helps explain why experienced workers saw much smaller gains. Top performers already possessed many of the behavioural patterns the AI was recommending. In some cases, the AI may even have constrained experts by nudging them toward average approaches rather than their own refined judgement. [arXiv]arxiv.orgarXiv Generative AI at WorkarXiv Generative AI at Work

The researchers described the system as disseminating the practices of high-performing workers across the wider workforce. In effect, the organisation’s best conversational habits became partially reproducible at scale. [ResearchGate]researchgate.net370227920 Generative AI at WorkResearchGate(PDF) Generative AI at Work25 Apr 2023 — Access to the tool increases productivity, as measured by issues resolved per hour…

This matters far beyond customer support. Many jobs depend heavily on informal know-how that is difficult to teach through standard training documents. AI systems may increasingly function as tools for compressing experience curves — allowing people to perform competently much earlier than before.

The AI worked partly as a live coach

A key detail is that the system did not simply automate customer conversations away from humans. The strongest effects appeared when the AI assisted people while they were still making decisions.

That distinction matters because much public discussion of AI and employment focuses narrowly on replacement. But the support-agent study showed a different model: augmentation through continuous coaching.

In practice, newer agents no longer had to rely entirely on:

Instead, the AI provided contextual guidance during the interaction itself.

That changed the tempo of learning. Traditionally, novice workers improve slowly because feedback loops are delayed. They only discover better approaches after repeated exposure to similar situations. The AI system shortened that process by supplying pattern recognition in real time.

The paper also found evidence consistent with accelerated worker learning. Over time, agents appeared to internalise some of the recommended practices rather than merely copying prompts mechanically. [arXiv]arxiv.orgarXiv Generative AI at WorkarXiv Generative AI at Work

This possibility — AI as a tool for distributing expertise rather than merely automating tasks — is one reason the study became important within wider discussions about “intelligence abundance”. If systems can help ordinary workers access high-quality guidance cheaply and quickly, some forms of expertise may become dramatically more scalable.

What customer sentiment adds to the case

The study’s customer-side findings were almost as important as the productivity numbers.

Customers interacting with AI-assisted agents were less likely to ask for supervisors and tended to behave more politely during conversations. [OUP Academic]academic.oup.comOUP AcademicGenerative AI at Work* | The Quarterly Journal of Economicsby E Brynjolfsson · 2025 · Cited by 3397 — Finally, we provide evi… 2arXiv

That matters because productivity improvements in customer service are often achieved by making interactions harsher or more rushed. Faster handling times alone do not necessarily mean better service.

In this case, the evidence suggested something more nuanced. The AI assistance may have improved conversational quality itself:

This reinforced the interpretation that the system was transferring behavioural competence, not merely speeding up typing.

For the broader AI bloom debate, this distinction is important. A future where AI only pushes humans to work faster is not especially attractive. But a future where AI systems help more people communicate effectively, solve problems competently, and gain confidence more quickly points toward a different possibility: broader cognitive empowerment.

Why this study became so influential

Many early AI claims relied on demos, speculative forecasts, or narrow benchmark tests. The support-agent study stood out because it captured AI operating inside an actual organisation with measurable business outcomes.

It also challenged a common assumption from earlier waves of automation. Historically, new technologies often complemented highly skilled workers most strongly, increasing inequality by amplifying existing advantages.

This case showed a different pattern. The largest gains appeared among people with less experience and lower initial performance. NBER [OUP Academic]academic.oup.comOUP AcademicGenerative AI at Work* | The Quarterly Journal of Economicsby E Brynjolfsson · 2025 · Cited by 3397 — Finally, we provide evi…

Economists immediately recognised the significance. If AI can reliably help lower-skilled workers approach competent performance faster, then some forms of expertise could become substantially more democratised.

That does not automatically produce broad prosperity. Companies may still use AI mainly to intensify monitoring, reduce headcount, or weaken bargaining power. Some evidence already suggests firms are experimenting aggressively with AI-driven cost reduction in support operations. [IT Pro]itpro.comDespite wide adoption—nearly two-thirds already use AI tools and 88% plan full deployment within a year—key challenges like data exposure… [Fortune But the support-agent results revealed at least one plausible pathway toward a more optimistic outcome: AI systems that spread practical capa]fortune.comgenerative ai customer service agentsYour next customer service agent could be generative AI16 Feb 2024 — Generative AI is still prone to “hallucinate” and spit out false ans… bility more widely rather than concentrating it solely among existing elites.

The limits of the evidence

The study was influential partly because strong workplace evidence remains relatively rare.

Several important cautions remain.

First, customer support is unusually rich in historical conversational data. That makes it especially suitable for systems trained on large numbers of past interactions. Jobs requiring physical dexterity, deep strategic judgement, or highly novel reasoning may not show the same pattern.

Second, the gains were highly uneven. The strongest improvements appeared in relatively structured environments where success could be measured clearly. [OUP Academic]academic.oup.comOUP AcademicGenerative AI at Work* | The Quarterly Journal of Economicsby E Brynjolfsson · 2025 · Cited by 3397 — Finally, we provide evi…

Third, there is an unresolved tension between assistance and deskilling. If AI systems continuously guide workers, employees may become dependent on them rather than developing deeper independent expertise. Critics worry that organisations could eventually reduce genuine apprenticeship opportunities by relying too heavily on AI guidance.

There is also the question of surveillance. Real-time coaching systems can easily become real-time monitoring systems. Employers may use AI not only to assist workers but to measure, rank, pressure, and standardise them more aggressively.

Recent backlash against poorly implemented customer-service AI shows these risks are real. Many companies deploying AI support systems have encountered problems involving hallucinations, governance failures, and deteriorating customer trust. [IT Pro]itpro.comDespite wide adoption—nearly two-thirds already use AI tools and 88% plan full deployment within a year—key challenges like data exposure… [TechRadar So the optimistic interpretation is conditional. AI coaching appears capable of helping novice workers substantially in some environments. Bu]techradar.comNotably, 74% have rolled back or shut down at least one AI agent due to governance issues. This trend suggests that companies with more m… t whether that translates into broader human flourishing depends heavily on how organisations deploy the technology and who captures the gains.

Why this matters beyond call centres

The support-agent findings matter because they may represent an early glimpse of a much larger shift.

Modern economies contain enormous quantities of fragmented practical knowledge:

Much of this knowledge remains locked inside individual careers because it is difficult to formalise.

Generative AI systems may increasingly function as interfaces to that accumulated experience. The deeper significance of the support-agent study is not really about call centres. It is about the possibility that useful competence itself could become more accessible.

If AI systems can help people learn faster, avoid common mistakes, and access expert-level guidance cheaply, the long-term implications extend into education, healthcare, scientific work, and many other domains tied to human capability.

That possibility sits near the centre of the broader AI bloom argument. The optimistic case is not merely that machines become more intelligent. It is that intelligence — practical problem-solving ability, accumulated know-how, and access to expertise — becomes more abundant for ordinary people.

Endnotes

1. Source: nber.org
   Link: https://www.nber.org/papers/w31161
   Source snippet

   NBERGenerative AI at Workby E Brynjolfsson · 2023 · Cited by 3217 — Access to the tool increases productivity, as measured by issues reso...

2. Source: academic.oup.com
   Link: https://academic.oup.com/qje/article/140/2/889/7990658
   Source snippet

   OUP AcademicGenerative AI at Work* | The Quarterly Journal of Economicsby E Brynjolfsson · 2025 · Cited by 3397 — Finally, we provide evi...

3. Source: arxiv.org
   Title: arXiv Generative AI at Work
   Link: https://arxiv.org/abs/2304.11771

4. Source: brc.org.uk
   Title: generative ais impact on customer service
   Link: https://brc.org.uk/news-and-events/news/technology/2024/generative-ais-impact-on-customer-service/
   Source snippet

   Generative AI's impact on customer service5 Oct 2023 — The study examined the staggered deployment of a chat assistant for a Fortune 500...

5. Source: arxiv.org
   Link: https://arxiv.org/pdf/2304.11771
   Source snippet

   Generative AI at Workby E Brynjolfsson · 2023 · Cited by 3229 — We find that access to AI assistance increases the productivity of agents...

6. Source: nber.org
   Link: https://www.nber.org/system/files/working_papers/w31161/w31161.pdf
   Source snippet

   NBERNBER WORKING PAPER SERIES GENERATIVE AI AT...by E Brynjolfsson · 2023 · Cited by 3397 — In addition, we find that AI assistance impr...

7. Source: forrester.com
   Title: Unlock AI-Led Customer Service with Gen AI & Tacit
   Link: https://www.forrester.com/blogs/youre-overlooking-the-key-to-customer-service-automation-tacit-knowledge/
   Source snippet

   ForresterUnlock AI-Led Customer Service with GenAI & Tacit...February 4, 2025 — 4 Feb 2025 — Explore how tacit knowledge and genAI trans...

   Published: February 4, 2025

8. Source: linkedin.com
   Link: https://www.linkedin.com/posts/mitch-gunnels_the-tacit-knowledge-conundrum-and-how-ai-activity-7348350489236807681-2qty
   Source snippet

   How AI is learning to overcome the tacit knowledge problemTacit knowledge includes the skills, intuition, and know-how that people pick u...

9. Source: researchgate.net
   Title: 370227920 Generative AI at Work
   Link: https://www.researchgate.net/publication/370227920_Generative_AI_at_Work
   Source snippet

   ResearchGate(PDF) Generative AI at Work25 Apr 2023 — Access to the tool increases productivity, as measured by issues resolved per hour...

10. Source: fortune.com
    Title: generative ai customer service agents
    Link: https://fortune.com/2024/02/16/generative-ai-customer-service-agents/
    Source snippet

    Your next customer service agent could be generative AI16 Feb 2024 — Generative AI is still prone to “hallucinate” and spit out false ans...

11. Source: arxiv.org
    Title: arXiv AI and jobs. A review of theory, estimates, and evidence
    Link: https://arxiv.org/abs/2509.15265

12. Source: techradar.com
    Link: https://www.techradar.com/pro/the-most-advanced-organizations-arent-failing-less-theyre-seeing-failures-sooner-many-firms-are-already-having-to-roll-back-ai-customer-service-tools
    Source snippet

    Notably, 74% have rolled back or shut down at least one AI agent due to governance issues. This trend suggests that companies with more m...

13. Source: brynjolfsson.com
    Link: https://www.brynjolfsson.com/
    Source snippet

    Erik BrynjolfssonErik Brynjolfsson is the Jerry Yang and Akiko Yamazaki Professor and Senior Fellow at the Stanford Institute for Human-C...

14. Source: linkedin.com
    Link: https://www.linkedin.com/posts/robynnstorey_ineedahuman-isitreallycustomerservice-activity-7306358484101431296-FFCM
    Source snippet

    AI in customer service: a win for companies, not customersWhen companies say they are using AI to improve customer service…. It means the...

15. Source: linkedin.com
    Link: https://www.linkedin.com/posts/alignai_ai-agents-can-automate-a-lot-but-what-they-activity-7432779346798424064-gYZx
    Source snippet

    Capturing Tacit Knowledge for AI Success | AlignAI posted...AI agents can automate a lot. But what they can't do on their own is the thi...

16. Source: linkedin.com
    Title: Erik Brynjolfsson
    Link: https://www.linkedin.com/in/erikbrynjolfsson
    Source snippet

    Professor, Writer, Speaker, InventorErik Brynjolfsson is the Jerry Yang and Akiko Yamazaki Professor and Senior Fellow at the Stanford In...

17. Source: nber.org
    Title: Erik Brynjolfsson The Economics of Transformative AI. Book
    Link: https://www.nber.org/people/erik_brynjolfsson
    Source snippet

    Erik BrynjolfssonThe Economics of Transformative AI. Book - Conference Volume. editors - Ajay K. Agrawal, Erik Brynjolfsson & Anton Korin...

18. Source: arxiv.org
    Link: https://arxiv.org/pdf/2508.19942
    Source snippet

    Socially Interactive Agents for Preserving and Transferring...by M Benderoth · 2025 · Cited by 4 — In this paper, we proposed the applic...

19. Source: researchgate.net
    Link: https://www.researchgate.net/publication/395015958_Socially_Interactive_Agents_for_Preserving_and_Transferring_Tacit_Knowledge_in_Organizations
    Source snippet

    Socially Interactive Agents for Preserving and Transferring...28 Aug 2025 — This paper introduces a novel approach to tackle the challen...

20. Source: siepr.stanford.edu
    Link: https://siepr.stanford.edu/publications/working-paper/generative-ai-work
    Source snippet

    ed per hour, by 14% on average, including a 34% improvement for novice and low-skilled...Read more...

21. Source: itpro.com
    Link: https://www.itpro.com/technology/artificial-intelligence/ai-agents-arent-cutting-it-in-customer-service
    Source snippet

    Despite wide adoption—nearly two-thirds already use AI tools and 88% plan full deployment within a year—key challenges like data exposure...

22. Source: Wikipedia
    Title: Erik Brynjolfsson
    Link: https://en.wikipedia.org/wiki/Erik_Brynjolfsson
    Source snippet

    Erik BrynjolfssonErik Brynjolfsson is an American academic, author and inventor. He is the Jerry Yang and Akiko Yamazaki Professor and...

23. Source: economics.stanford.edu
    Title: erik brynjolfsson
    Link: https://economics.stanford.edu/people/erik-brynjolfsson
    Source snippet

    is the Jerry Yang and Akiko Yamazaki Professor and Senior Fellow at the Stanford Institute for Human-Centered AI (HAI), and Director of...

24. Source: gsb.stanford.edu
    Title: erik brynjolfsson
    Link: https://www.gsb.stanford.edu/faculty-research/faculty/erik-brynjolfsson
    Source snippet

    He also holds appointments at SIEPR, the Stanford University Department of Economics.Read more...

25. Source: x.com
    Title: Erik Brynjolfsson (@erikbryn) / Posts / XErik Brynjolfsson (@erikbryn)
    Link: https://x.com/erikbryn
    Source snippet

    Posts - Director @DigEconLab Co-founder, @Workhelix @StanfordHAI @SIEPR @Stanford amazon.com/Sec... | X (formerly Twitter)...

Additional References

1. Source: bitter-sweet.ai
   Link: https://www.bitter-sweet.ai/articles/generative-ai-at-work-brynjolfsson-2023/
   Source snippet

   14% Productivity Boost for Customer Service AgentsThe first large-scale field experiment showing AI assistants increased worker productiv...

2. Source: sternstrategy.com
   Link: https://sternstrategy.com/speakers/erik-brynjolfsson/
   Source snippet

   Erik Brynjolfsson AI and Economist SpeakerErik Brynjolfsson, an optimistic economist, explores how technological advances, like AI, impac...

3. Source: analysisgroup.com
   Link: https://www.analysisgroup.com/people/affiliated-experts/erik-brynjolfsson/

4. Source: serviceji.co
   Link: https://serviceji.co/blog/ai-customer-support-backlash/
   Source snippet

   ServiceJiThe AI Customer Support Backlash: Why Companies Are...5 days ago — AI can make support faster, but forced automation is creatin...

5. Source: tomshardware.com
   Link: [https://www.tomshardware.com/tech-industry/intel-shifts-customer-support-to-microsoft
   Source snippet

   This move includes the reduction of phone and direct social media support, redirecting users to online case systems and [community]({{ 'ai-bloom-abun/ai-bloom-abun-98d3a6-long-future-b-29cc56-abundant-know-ca00d9-community-hea...

6. Source: 3c-online.co.uk
   Link: https://3c-online.co.uk/accessible-customer-service-expertise/
   Source snippet

   Transforming Tacit Knowledge into Accessible ExpertiseDiscover how transforming tacit knowledge into accessible, structured expertise can...

7. Source: cxnetwork.com
   Title: ai wont replace human customer service agents gartner predicts
   Link: https://www.cxnetwork.com/artificial-intelligence/news/ai-wont-replace-human-customer-service-agents-gartner-predicts
   Source snippet

   Replace human customer service agents15 Sept 2025 — Predictions that artificial intelligence (AI) agents will replace customer service ag...

8. Source: searchunify.com
   Title: capture tacit knowledge with intelligent knowledge agents
   Link: https://www.searchunify.com/resource-center/blog/capture-tacit-knowledge-with-intelligent-knowledge-agents/
   Source snippet

   Use Intelligent Knowledge Agents To Capture Tacit...4 Oct 2024 — Discover how knowledge agents can help your organization capture and le...

9. Source: cmr.berkeley.edu
   Title: tacit knowledge is your next competitive moat
   Link: https://cmr.berkeley.edu/2026/03/tacit-knowledge-is-your-next-competitive-moat/
   Source snippet

   California Management ReviewTacit Knowledge Is Your Next Competitive Moat16 Mar 2026 — Today, as agentic and generative AI become more in...

10. Source: customerexperiencedive.com
    Title: consumers prefer human led customer service
    Link: https://www.customerexperiencedive.com/news/consumers-prefer-human-led-customer-service/814226/
    Source snippet

    Consumers prefer talking to people in customer service9 Mar 2026 — Two-thirds of consumers are not confident with how companies use gener...

Amazon book picks

Further Reading

Books and field guides related to Novice support agents. Use these as the next step if you want deeper reading beyond the article.

Book

The Shift: The Future of Work is Already Here

By Lynda Gratton

An unhysterical look at the future of employment.

See on Amazon

Book

Artificial Intelligence in Economics and Finance Theories

By Tankiso Moloi, Tshilidzi Marwala

As Artificial Intelligence (AI) seizes all aspects of human life, there is a fundamental shift in the way in which humans are thinking of...

See on Amazon

Book

Artificial Intelligence in Education

By Ulrich Hoppe, Maria Feliza Verdejo et al.

This work reports on research into intelligent systems, models, and architectures for educational computing applications. It covers a wid...

See on Amazon

Book

Intelligent Tutoring Systems

By Gilles Gauthier, Claude Frasson et al.

ITS 2000 is the fifth international conference on Intelligent Tutoring Systems. The preceding conferences were organized in Montreal in 1...

See on Amazon

Browse more on Amazon

As an Amazon Associate I earn from qualifying purchases.