Lenny's Newsletter

▶ YouTubecompleted
Back to results

Lenny's Podcast · May 31, 2026

A rational conversation on where AI is actually going | Benedict Evans

AI StrategyPlatform ShiftsEnterprise AdoptionFoundation ModelsValue ChainJob DisplacementDistribution MoatsPricing PowerJevons ParadoxProduct StrategyConsulting/ServicesAI Maturity Curve

Benedict Evans argues AI is as transformative as the internet or mobile — but we're only in a "1997 moment," where most applications haven't been built yet and adoption is still wildly uneven. The bigger structural questions are whether foundation model labs will ever have real pricing power, and whether the value will accrue to applications and distribution rather than the underlying models.

Watch on YouTubePartial match

Topics (13)

The 1997 Internet Analogy for AI

3:19

AI today is at the same maturity stage as the early internet: proven and exciting, but with most of its real-world impact still ahead.

How it works

  • Adoption is spread across a wide distribution: a small group of power users ("bought Mac mini clusters") versus a large majority using it occasionally or not at all.
  • Most of the software, workflows, and business models that will define AI haven't been invented yet.
  • The 1997 analogy warns against both hype ("picking races between Excite and Yahoo") and dismissal — almost all the real winners were unknown at that point.
  • Asking whether AI is "20% bigger than the internet" is unproductive; the useful question is what gets built next.
  • Why you should care

    As a PM, this framing resets your horizon: you're not optimizing for today's tools, you're positioning for a wave that is still building. Decisions made now — on platform bets, build-vs-buy, and team skills — have long lead times before they pay off, exactly like 1997 web strategy calls.

    Think of it like...

    It's like evaluating the internet in 1997 and confidently betting on Excite — you'd miss Google, Amazon, and smartphones entirely.

    Jagged Frontier of AI Capability

    4:13

    AI performance has a jagged, non-obvious edge — excellent at some things, useless at others, and the pattern is hard to predict in advance.

    How it works

  • A software developer may experience AI as a total transformation; a lawyer in the same building may find it interesting but largely irrelevant to their actual work.
  • You often can't tell in advance which tasks are on which side of the frontier.
  • You frequently can't tell *after the fact* either — evaluating whether AI did the job correctly is itself a hard problem.
  • The frontier maps to adoption spread: people who encounter AI in a domain where it works become daily users; everyone else stays at the periphery.
  • Why you should care

    For product decisions, this means you cannot rely on category-level analysis ("AI will disrupt legal") — you need task-level specificity. Discovery work on *which exact tasks* AI handles well in your domain is the prerequisite to any meaningful roadmap bet.

    Think of it like...

    It's like a map where some roads are highways and some are dirt tracks — and the map itself doesn't tell you which is which until you drive them.

    Task vs. Job Distinction

    12:50

    AI replaces tasks, not necessarily jobs — and the difference between the two determines whether a profession shrinks or transforms.

    How it works

  • An elevator attendant's task *was* the job: press the button, go to the floor. Automating the button eliminated the role entirely.
  • A McKinsey engagement's task is the 75-slide deck; the job is diagnosing organizational dysfunction, navigating politics, and surfacing what customers actually think. Claude can make a crappy deck; it cannot do the job.
  • Price elasticity applies: making a task cheaper often means more of the job gets done, not fewer people doing it (the accounting headcount example).
  • The hard question is: in your specific domain, is the task the job, or is it just the visible deliverable?
  • Why you should care

    This is the single most useful analytical frame for assessing AI's impact on your product or team. Before cutting headcount or scope, map which tasks AI can absorb and whether those tasks *are* the value, or just the packaging around the value.

    Think of it like...

    Amazon gets you the SKU if you know what you want — but knowing what SKU to order is a completely different job.

    Jevons Paradox / Price Elasticity of Automation

    13:44

    Making a task cheaper typically increases how much of it gets done, not how few people are needed to do it.

    How it works

  • The Jevons Paradox: cheaper coal led to *more* coal consumption, not less, because more applications became economically viable.
  • Example: Excel made financial modeling cheaper — Goldman Sachs now has *more* analysts doing *more* modeling, not fewer analysts doing the same work.
  • Example: IDEs and libraries handle 90% of iPhone app code — yet there are vastly more software engineers now, not fewer.
  • The accounting headcount chart went up through every wave of automation: adding machines, mainframes, ERP, cloud.
  • Why you should care

    For roadmap and org planning, resist the reflex that "AI does X% of the work → we need X% fewer people." The more likely outcome is that your team does dramatically more, faster — which changes what you should be building toward, not how many seats you cut.

    Think of it like...

    Building a faster highway doesn't reduce traffic — it creates more of it, because more trips now make sense.

    Foundation Model Pricing Power (or Lack Thereof)

    38:32

    Foundation model companies probably won't capture most of AI's value because the models look increasingly like undifferentiated commodity infrastructure.

    How it works

  • Models don't appear to have strong network effects: one model winning doesn't lock users into that model, so competition persists indefinitely.
  • From a user perspective, models are converging — a normal person cannot reliably distinguish between Gemini and Claude for most tasks.
  • Telecom analogy: mobile data consumption is on a perfect exponential curve, mobile stocks have gone nowhere in 25 years. Volume ≠ margin.
  • Sam Altman's "sell intelligence like electricity" framing ignores that utility infrastructure historically captures almost none of the value created by what runs on top of it.
  • Why you should care

    If you're choosing a model vendor, long-term lock-in risk is lower than it feels. The application and distribution layers are where durable value is more likely to accumulate — which reinforces the case for investing in your own product differentiation rather than betting on one model partner.

    Think of it like...

    Noble-prize-winning flat panel screen technology is still a low-margin commodity — scientific complexity doesn't create pricing power.

    Distribution as the Dominant Moat in a Commodity Model Market

    43:53

    In a world of commoditized models, whoever has the most distribution wins — not whoever has the best model.

    How it works

  • Browser analogy: the rendering engine can vary in quality, but the product is a thin wrapper — Microsoft won browsers through distribution (bundling with Windows), not through technical superiority.
  • Google is using Android and Search surfaces to drive Gemini adoption regardless of model quality delta.
  • Meta distributed an "adequate" model across every surface it owns and ended up with usage comparable to ChatGPT — despite being written off in tech circles.
  • The strategic implication: an adequately performing model with massive distribution beats a superior model with weak distribution.
  • Why you should care

    For enterprise product strategy, your AI feature doesn't need to be the best model — it needs to be the default in the workflow your users are already in. Embedding AI inside existing high-frequency surfaces beats launching a standalone AI product almost every time.

    Think of it like...

    Microsoft won the browser war not by building a better browser, but by making IE the default on a billion Windows machines.

    Enterprise Deployment Lag

    21:33

    Even if AI is ready, large organizations aren't — their adoption clock runs on 3-10 year cycles, not the 2-week Twitter doomer timeline.

    How it works

  • Enterprise software sales cycles run 18+ months; ripping out SAP and replacing it with AI-native tooling is a multi-year program, not a quarterly decision.
  • Workflow redesign is itself a project requiring 5-10 people spending 1-2 months just to scope what could be automated — before any actual implementation.
  • This is why AI labs are investing in forward-deployed engineers and consultancies: enterprises don't have internal capacity to run these transformation projects.
  • Most SaaS companies founded before ChatGPT could have been built years earlier — the delay was people figuring out the application, not the technology being unavailable.
  • Why you should care

    If you're in enterprise product, your competitive threat from AI-native startups is real but slower than it looks. Your window to embed AI into existing workflows before a replacement product captures your users is measured in years — but that window is finite.

    Think of it like...

    No one tears out SAP on a Tuesday — enterprise transformation happens on the timescale of construction projects, not app updates.

    Lump of Labor Fallacy in AI Job Displacement

    0:07

    History shows that automation destroys specific jobs and creates entirely different new ones — the net is more jobs and more prosperity, though the transition is painful.

    How it works

  • Since 1800, every major technology wave (steam, electricity, computers) automated jobs and created different jobs. We went from 90% agricultural labor to today's economy.
  • The new jobs always look implausible in advance: "railway engineer" made no sense before railways.
  • AI adoption speed is faster because it stands on existing infrastructure (internet, smartphones) — but the structural dynamic is the same.
  • The real risk is frictional pain during transition: hollowed-out towns, displaced careers — real harm even if aggregate outcomes are positive.
  • Why you should care

    For org planning and hiring strategy, the question is not "how many roles does AI eliminate" but "what new capability and role shapes does AI make possible that we should be building toward." Headcount reduction is the short-term frame; capability expansion is the durable one.

    Think of it like...

    You can always see the job that's about to disappear, but the job that replaces it doesn't exist yet — so it always looks like there won't be one.

    Value Stack Migration (Infrastructure vs. Application Layer)

    35:24

    The money in tech platform shifts almost never sits at the infrastructure layer — it accumulates in the applications and experiences built on top.

    How it works

  • Telecom built the global mobile network, captures almost none of the app economy value. Apple and developers captured it.
  • AWS is commodity cloud infrastructure; the value is in the software companies running on it — customers don't standardize on AWS, they standardize on the apps.
  • Windows had actual leverage because apps were tied to it — foundation models don't appear to have that same lock-in structure.
  • The model layer looks more like AWS (commodity, interchangeable) than Windows (lock-in, platform leverage).
  • Why you should care

    This tells you where to build: application differentiation, workflow integration, and proprietary data are more defensible than model selection. If you're debating whether to deep-integrate with one model provider, the AWS analogy suggests the risk of lock-in is lower than it feels — but the opportunity from building great apps is real.

    Think of it like...

    The telecom built the pipes, Apple built the App Store, and developers built the apps — the pipes are a utility, the apps are the business.

    Doing the Old Thing More vs. Creating the New Thing

    1:01:41

    The first wave of any new platform is faster/cheaper versions of old things; the real value comes from new things that only that platform makes possible.

    How it works

  • Early e-commerce was just catalog retail online; Amazon's real innovation was a searchable global inventory at scale — a new category.
  • Early mobile apps were desktop software shrunk to a screen; then cameras, notifications, and location enabled entirely new businesses.
  • Spotify is not an online music store — it redefined the question from "buy a track" to "access all music for a flat fee."
  • AI is still mostly in phase one: doing existing tasks faster. The phase-two products haven't been imagined yet.
  • Why you should care

    Most AI product roadmaps right now are phase-one work — automating existing workflows. That's valuable but defensible only temporarily. The durable competitive question is: what does AI make *possible* that couldn't exist before? Discovery work should be hunting for those phase-two opportunities.

    Think of it like...

    Printing emails out was phase one of the internet; Gmail, Twitter, and Slack were phase two.

    O*NET-Style Job Exposure Analysis as a Flawed Framework

    1:03:33

    Breaking a profession into sub-tasks and scoring each for AI-replaceability sounds rigorous but produces misleading numbers.

    How it works

  • The expert systems problem: trying to recognize a cat by building logical sub-detectors (edge, ear, eye) leads to 700 steps and still doesn't work. Jobs decompose the same way — the model breaks.
  • The inverse error: taxi drivers looked impossible to automate via the internet in 1997, yet Uber restructured the entire industry.
  • Conversely, "personal trainer" looks human and physical — but an AI watching you via your iPhone camera and building your program is already plausible.
  • Quantitative exposure scores create false confidence in predictions that are structurally impossible to make.
  • Why you should care

    Don't use percentage-exposure frameworks to justify product or headcount decisions. The real signal is empirical: run small experiments where AI touches actual workflows, observe what happens, and iterate. Analytical decomposition is not a substitute for hands-on evaluation.

    Think of it like...

    You can't predict which buildings will survive an earthquake by analyzing their blueprints — you need real-world stress tests.

    The Forward-Deployed Engineer as Modern Consultant

    11:22

    AI transformation in enterprises requires human experts on-site to figure out and implement new workflows — exactly what consultancies have always sold.

    How it works

  • Enterprises don't have spare teams sitting around to rethink internal operations; transformation is a discrete project requiring external capacity.
  • Scoping what can be automated with AI is itself a 5-10 person, 1-2 month engagement — before any implementation begins.
  • Ironically, the companies most invested in AI (OpenAI, Anthropic) are the ones most aggressively building or acquiring consulting-style delivery capacity.
  • The forward-deployed engineer is structurally identical to an Accenture engagement — just run by people in San Francisco instead of Bangalore.
  • Why you should care

    For enterprise PMs, this signals that AI adoption is a change management and services problem as much as a technology problem. Products that bundle deployment support, workflow consulting, or "opinionated" implementation guides will have adoption advantages over raw API offerings.

    Think of it like...

    A forward-deployed engineer is just an Accenture consultant with a GitHub account and a San Francisco ZIP code.

    AGI Redefinition Problem

    27:28

    AGI is a moving target that gets redefined every time AI achieves something — which makes it nearly useless as a planning concept.

    How it works

  • Larry Tesler's observation: "AI is whatever machines can't do yet" — once machines can do it, people say "that's just software."
  • Image recognition, sentiment analysis — both were called AI, then reclassified as mere software once they worked reliably.
  • Current AGI definitions have shifted from "has a soul" to "can perform X% of economically valuable work" — which an IBM mainframe in 1975 arguably met.
  • We have no theory of human intelligence, no theory of why LLMs work, and no theory of how much better they'll get — all forecasting is "vibes-based."
  • Why you should care

    Don't build product strategy around AGI timelines or definitions — they're definitionally unstable. Focus instead on empirical capability assessments of specific tasks today and near-term. Scenario planning against AGI milestones is mostly a distraction from shipping.

    Think of it like...

    AGI is like the horizon — you can walk toward it forever and it always stays the same distance away.

    Why this matters to you

    Your AI roadmap is probably too near-term.

    We're in a 1997 moment — most of the applications that will define the AI era haven't been invented yet. Optimizing your current product for today's model capabilities is necessary but insufficient; the bigger strategic question is what you'll build when the platform matures.

    Foundation model lock-in risk is lower than you think — but so is their long-term leverage.

    The structural case for foundation models commoditizing over time (no network effects, converging quality, utility-style margin pressure) means you should invest in application-layer differentiation, not model-layer dependency. The value will be in your workflows, data, and distribution.

    Enterprise AI adoption is a services and change management problem, not just a technology problem.

    The reason AI labs are buying consultancies is the same reason your customers will need help: they don't have spare capacity to redesign their own workflows. Products that come with embedded deployment support, opinionated workflows, or strong professional services partnerships will win on adoption, not just features.

    Relevant transcript passages (6)
    14:52
    what Amazon does is get you the skew. If you know what the skew is, if you know what skew you want, you want that microphone stand. You know this part number, you can go to Amazon and get it. If you don't know what microphone to get, probably shouldn't start on Amazon.

    Crisp illustration of the task-vs-job distinction applied to e-commerce — directly transferable to AI product scoping decisions.

    14:02
    you know this is a joke I made on Twitter back when it was Twitter was like young people won't believe this but before invest before Excel junior investment bankers worked really long hours and now thanks to Excel Goldman's associates all work at lunchtime on Fridays.

    Sharpest delivery of the Jevons Paradox point — automation made the task cheaper and produced more demand, not fewer workers.

    12:15
    you would think AI is going like consultants were going to be gone. No, we don't need all these people anymore. AI is going to do their work. Instead, like the most cutting edge AI labs are the ones most investing in these folks.

    The forward-deployed engineer paradox — useful as a signal for how enterprise AI adoption actually works vs. how it's narrativized.

    44:58
    clearly what's happening now is that Google is using distribution to drive um to drive Gemini and like what's the difference between Gemini and for and and and like if you're you know if you're using this stuff all day then you know but like normal person there's no difference

    Direct evidence of the distribution-as-moat thesis playing out in market behavior right now.

    47:16
    the model is just like the dumb thing underneath the funny way of putting it the dumb thing underneath that powers the feature the model is the commodity that powers different decisions about what the feature should be and what different distribution

    Most direct articulation of the value stack migration thesis — the model is infrastructure, the feature is the product.

    1:03:33
    you can't kind of look at a senior partner at a law firm and say, well, 17% of their work could be automated like this is horshit. You can't do that.

    Cuts through the O*NET-style job exposure analysis that is widely used in enterprise AI business cases — important caveat for any PM using such frameworks to justify roadmap bets.

    Key Insights (7)
    • AI is as big as the internet or mobile — and only as big. The comparison is a ceiling as well as a floor, and it implies a long, uneven rollout ahead.
    • Most AI usage data is either missing (model labs publish no meaningful DAU figures) or methodologically unreliable (survey-based). Strategic decisions made on this data should carry wide uncertainty margins.
    • The 'task vs. job' frame is the most operationally useful lens for assessing AI's impact on any role: does automating the task eliminate the job, or does it just make the job faster/cheaper and therefore more in demand?
    • Distribution is the moat in a commodity model market. Google and Meta are demonstrating this in real time — adequate models with massive existing surfaces outcompete superior models with no default placement.
    • The accounting headcount chart is a killer data point: headcount grew through adding machines, mainframes, ERP, cloud, and PCs. The null hypothesis for any profession facing automation should be 'more demand, different shape' — not 'fewer people.'
    • Enterprises will take 3-10 years to meaningfully restructure around AI, not 2 weeks. The frictional lag is structural (sales cycles, legacy systems, change management capacity) — not just cultural resistance.
    • Phase-two products — things only possible because of AI, not just old things done faster — haven't been built yet. This is where the real TAM expansion is, and it's largely invisible today.
    Action Items (6)
    • Run a task-vs-job audit on your core product: for each major user workflow, explicitly ask whether the task IS the job (automatable away) or just packaging around the job (price-elastic, will expand demand).
    • Audit your AI vendor strategy through the commodity-infrastructure lens: are you building differentiation at the application layer, or are you dependent on a specific model's current quality advantage that is likely to erode?
    • Map your product's distribution surfaces — where does your AI feature appear in workflows users are already in? Prioritize embedding over standalone AI product launches.
    • Replace percentage-exposure frameworks in your business cases with empirical pilots: instrument a specific workflow with AI, measure actual outcomes, then make headcount and investment calls from real data rather than task decomposition models.
    • Identify one area in your product where you're doing 'the old thing but more' with AI — then run a discovery sprint on what AI makes possible that literally couldn't exist before.
    • If you're in enterprise B2B: build or partner for deployment support capacity. The adoption bottleneck is not feature quality, it's workflow redesign capacity inside customer organizations.
    Skip if: You are already deeply familiar with the commodity-infrastructure vs. application-layer value migration debate and have read Benedict Evans' AI presentations directly. The job displacement and Jevons Paradox sections are standard tech-economics; skip if those are already in your mental model. The AGI redefinition concept is lightweight and mainly rhetorical.