I Attended an HBS Lecture on AI and the Future of Work. Here’s My Emotional Journey.

Visit to Boston, June 2026

Every AI presentation follows the same emotional arc. You show up curious, maybe a little skeptical, and within fifteen minutes you’re leaning forward thinking this is incredible, we’re living in the most remarkable moment in human history. By the end, you’re quietly wondering whether to change career advice to your kids.

Last week I attended a lecture by Professor Joe Fuller of Harvard Business School, one of the leading researchers on the future of work. He did not disappoint. What follows is part data summary, part personal reaction, and part genuine optimism. I think the scary slides, as good as they are, are missing something important.

Act One: The Excitement (Slides 1–15, approximately)

Let’s start where Prof Fuller started: with the genuinely astonishing.

AI inference costs have dropped 97% per year at constant capability since late 2023. Not 97% total. Per year. Impressive progress. However, just like switching to Powerpoint made making slides easier, the expectations wound up going up. So we all wound up spending more time on slides than we used to when expectations were low. So, despite the costs of accomplishing a given task going down dramatically, expect overall token usage and costs to continue to climb. Jevons' paradox is alive and well.

Meanwhile, the people actually using these tools are reporting extraordinary productivity gains.

• A recent survey of technical researchers, engineers, and managers found they self-report being 1.5–2.1x more productive today than before AI, with expectations rising to nearly 3x by 2027.

• Companies like Wells Fargo, JP Morgan, P&G, and Cisco have moved from experimentation to coherent enterprise strategies.

• One unnamed large enterprise has invested extensively in an AI-enabled marketing transformation using synthetic customer avatars for training and refining their messaging. They expect this to save them in the range of $100M / year.

• Ohio State University declared itself “AI native” and simply skipped the 18-month faculty committee study that every other institution is still conducting.

• Pharma companies are running old chemical inventories through AI models to find new drug candidates that have been sitting undiscovered in the data for decades. Imagine the benefits from this work.

• The models are getting exponentially better, and they are doing so at lightning speed. Prof Fueller compared the first model he was genuinely excited about to a used Prius, and the current models to Lamborghinis. This change happened in < 2 years.

GenAI's Capabilities are Growing Exponentially. Models are getting exponentially better, and they are being released faster and faster.

At this point in the presentation, the energy in the room was high. People were nodding. A few were quietly calculating how to apply this to their own business. This is the moment, everyone was thinking. What a time to be alive.

And then Prof Fuller clicked to the next section.

Act Two: Hold On, What Happens to Everyone’s Job?

Here’s where the tone shifted.

70% of Americans already view AI negatively, and most of them can’t fully articulate why. They just sense something coming. It turns out their instincts aren’t wrong.

The data on entry-level jobs is sobering. Across both software development and customer service, employment of cognitive workers aged 22–30 has declined sharply since late 2022, exactly when AI tools became widely accessible. So far, older, more senior cohorts have held steady or grown.

Evaporating Entry Level Cognitive Positions

There’s a structural issue that may be masking what is to come. The conventional tech adoption curve is an S-curve: slow, then fast, then plateau. GenAI follows a J-curve: before you get the return, you take a productivity hit while teams invest in learning and process redesign. Many organizations are sitting in the bottom of that J right now, wondering why their ROI hasn’t yet materialized. 

Conventional New Technology Adoption S-Curve vs. the likely AI Adoption J-Curve

Some economists are stating that we are overestimating the labour impacts, but Prof. Fuller disagrees with this take.  In the same way the tide goes out dramatically before a tsunami, the job losses to come may also be masked by the fact that right now so many companies are making the additional investments that will allow them to shed employees in dramatic fashion soon.

Even the tech giants themselves are not immune.  A friend of mine who works at Google, on Gemini, just had her group summarily cut from 7 people to 2 without warning.  One of the criteria for who stayed?  The ones who used AI tools the most to do their work.

Act Three: The Two Slides That Made Me Text My Kids

This is where I want to slow down, because these two slides deserve more than a passing mention.

Slide 1: Every Industry Has Its Own Shape

Before you can understand what AI will do to a workforce, you have to understand what that workforce looks like.

Every industry sector has its own geometry of work. Every sector has different distributions of their workforce between Entry-level roles, Individual Contributors, various management levels and senior leadership. The workforce percentage at each level is shown numerically. Healthcare, for instance, has 23% entry-level jobs, and 57% individual contributor roles.

Prof Fuller’s team mapped a number of industries by their internal job-level distribution: what fraction of workers sit at entry-level, individual contributor, lower management, middle management, senior management, and leadership.  They presented a sample of six in these slides.

Banking and Software both look like diamonds or wide hourglasses . The bulk of their workforce sits at the individual contributor and middle management levels, with relatively thin entry-level cohorts (8–11%). Healthcare and Automotive Manufacturing, by contrast, look like crosses or wide pyramids. They’re bottom-heavy, with entry-level roles representing 21–23% of healthcare workers and a larger share in automotive.

This matters because of what comes next.

Slide 2: The AI and the Future of Work Disruption Is Not Uniform

This is the slide I keep coming back to. It overlays each of those workforce shapes with a prediction of how GenAI will affect each job level, broken into four categories:

•       Automation: The task is fully replaced with no human required

•       Augmentation: Human still needed, but AI does a substantial share of the work (fewer people for the same or more output)

•       Lower Potential: The task changes marginally

•       Non-Language: The task is structurally outside the reach of LLMs.  This represents physical, hands-on, sensory work that a language model simply cannot do

The source methodology comes from Accenture’s task-level analysis of over 19,000 individual work tasks, cross-referenced against Lightcast labor market data. One important caveat: the slide doesn’t specify a target year, it represents current and near-term exposure at today’s AI frontier. It is not a projection to a specific year like 2030 or 2040.

Here’s what jumps out:

• Banking and Software are almost entirely dark. Look at the individual contributor bars for both industries: they are dominated by automation and augmentation colors, with almost no light grey (non-language). Roughly a third of roles at every level appear to be fully automatable, with another third facing augmentation.

• Healthcare looks different. The individual contributor level in healthcare , which represents 57% of all healthcare workers, has a substantial non-language component. A large portion of what healthcare workers do is physical: administering medications, performing procedures, providing hands-on care. Language models can’t do any of that.

This is why I texted my kids, who are still evaluating career options.  Healthcare is looking pretty good in the AI era.

Close up of the Healthcare and Automotive Manufacturing sectors that shows predicted AI impact on different job levels. The bottom level is Entry-level roles, followed by Individual Contributors, various management levels and Senior Leadership. Dark blue shows the proportion of today's jobs expected to be fully automated by AI. Medium blue shows the proportion expected to be augmented by AI. Light blue shows the proportion of jobs expected to have little impact from AI. The gray areas are expected to have no impact from AI, as they are non-language activities.

A few more observations:

•       The data makes me want to see this analysis for industries Prof. Fuller didn’t present: legal (where AI is already displacing junior associates), media and animation, and the skilled trades. The trades are both physically non-language and facing a severe labour shortage. They may well emerge as a great career option in the AI era.

•       The “Non-Language” category is specifically about LLMs and GenAI. Physical robotics is a separate conversation. For instance, automotive manufacturing may face a different threat on a longer timeline.

•       These slides don’t tell you the timeframe. The exposure exists. The pace of realization will vary enormously by industry, regulation, capital availability, and organizational will.  Just like my friend’s group at Google, whether your job is fully automated away or merely augmented may also depend on individual willingness to adapt and embrace new ways of working.

What These Artificial Intelligence Slides Don’t Show

I want to end with what I think is the most important point, and the one thing that was conspicuously absent from this short presentation.

These charts are beautifully rigorous about measuring what currently exists. They can tell you how many entry-level banking jobs are structurally exposed to automation. They cannot tell you about the jobs that will exist in five years that don’t exist today.

Consider: in 2000, there were no social media managers, no app developers, no data scientists, no prompt engineers, no AI trainers. The internet destroyed entire categories of work and created entirely new ones that are now massive employers.

The historical pattern of general-purpose technology adoption is consistent: disruption is real and painful in transition, but total employment does not collapse. The industrial revolution did not end human work. The computing revolution did not end human work. The internet did not end human work.

I’m not dismissing the pain of the transition. It’s real, it’s massive, it’s unequal, and it lands hardest on people with the fewest options to adapt. I hope that both public and private organizations will take on the important mission of retraining displaced folks. But I part ways with the implicit narrative in some of these slides, which is that the exposure we can measure today is the full picture.

People are more adaptable than these charts can model. That’s not wishful thinking — it’s the entire history of human labour. What are your thoughts? Share them in the comments. If you liked this post, please send it to a friend.

Landon Steele is a startup consultant, angel investor, and advisor to early-stage founders. She is based in Vancouver, BC. She works with founders and the ecosystems that fund them across Canada and the US. Learn more at steeleconsultinggroup.com

_____________________

Sources:

Lecture Source:

Professor Joe Fuller, Harvard Business School, Managing the Future of Work Project — lecture attended June 5, 2026. Credit to him for the slides in this post.

Data & Research Sources:

1. Accenture — "A New Era of Generative AI for Everyone" — The foundational methodology for the workforce exposure analysis (automation / augmentation / lower potential / non-language categories), based on analysis of 19,000+ tasks across 867 occupations using O*NET and BLS data.

2. HBS Managing the Future of Work Project + Accenture Research based on Lightcast — Lightcast provided the labor market job posting data layered on top of the Accenture methodology.

3. Korinek, A. & McKelvey, P. (2026) — "Measuring the AI Economy," Peterson Institute for International Economics, Working Paper 26-9 — The 97% per year inference cost collapse chart.

4. Measuring the Task Completion Time Horizon of Frontier AI Models.

5. metr.org — "Measuring the Self-Reported Impact of Early-2026 AI on Technical Worker Productivity" (May 2026) — The 1.5–2.9x productivity multiplier survey of 349 technical workers.

6. Brynjolfsson, Chandar, and Chen (2025) — "Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of AI" — The evaporating entry-level cognitive positions data

7. Fuller, J., Sigelman, M., and Fenlon, M. — "The Expertise Upheaval," Harvard Business Review, March 2025 — Referenced on the asymmetric career pathways slide.