Convenience was the product

Why AI is disrupting the mid-tier of knowledge work — and what the food industry already taught us

The world’s largest law firm by revenue will hit $11B in valuation this quarter. That is up from $8B in December and from a standing start four years ago. Its customers are 1,300 law firms in 60 countries, including most of the AmLaw 100. Harvey is not a law firm. It sells software that helps them write contracts, run due diligence, and file briefs — the very work that, on the old model, would have been outsourced to a firm’s junior associates.

The question journalists have been asking for two years is whether systems like Harvey are hollowing out BigLaw from below. The evidence says the opposite: BigLaw is absorbing the disruption. Profit-per-lawyer at the AmLaw 100 is up more than 53% since 2019. Tech spend is up 10% year on year. Salaries are rising, not falling. What is happening to law is what has happened to every other industry that faced this pattern before: the top tier is getting richer by absorbing the new tools. The middle tier is where the compression is landing.

This essay is about that pattern. It is not the first time convenience work has collapsed; it is, on the evidence, the third or fourth round. The first was travel agents. Then one-hour photo, retail stockbrokers, video rental stores, tax preparers. Most recently — and the parallel that reveals the most — mid-tier casual dining: TGI Fridays, Applebee’s, Red Lobster, Bar Louie. In each case, the category survived by bifurcating. The bottom became cheap and self-service. The top became premium and experience-led. The middle collapsed.

AI is running the same play against knowledge work. The bottom is already visible: Meta’s Advantage+ ad products reached a $60B annual run-rate by the third quarter of 2025, up from roughly $40B a year earlier, with more than four million advertisers using the generative tools. Lovable, a generative product that lets non-developers build web apps, reached $100M in annual recurring revenue within eight months of launch — though superlatives about “fastest ever” should be read with the caveat that each of the last five years has produced a new “fastest ever” on the same metric. The middle is visible in a different form. In 2025, Accenture announced more than 11,000 layoffs between May and August; WPP reduced its headcount from about 108,000 to 98,655 — a 9% cut, roughly 9,000 jobs eliminated; Omnicom and IPG closed their $13.25B merger in November. The top, as already noted, is growing.

The interesting question is not whether the pattern is happening. The data says it is. The interesting question is what the food-industry version of this story already taught us about how it ends.

Consulting: firing at the middle, billing at the top

Consulting is the purest case, because the data is public. Accenture reported FY25 revenue of roughly $70B with new bookings of $21.3B, of which $5.9B were AI services. Its AI revenue tripled to $2.7B year on year; its AI-and-data headcount grew from about 40,000 in FY23 to 77,000. In the same year, the firm cut more than 11,000 jobs. Its CEO, Julie Sweet, told staff the cuts targeted “people who cannot be reskilled for AI.” The firm is hiring, at volume, at one end of its workforce and firing, at volume, at the other. The generic tier — template decks, commodity research, routine data work — is where the cuts are landing.

McKinsey’s public numbers track the same pattern. The firm’s partners trimmed about 10% of headcount over 18 months; junior-level cuts ran at roughly 25%; back-office at 13%. McKinsey’s own QuantumBlack AI practice grew from about 400 to 1,000 practitioners in two years. About 40% of firm revenue now comes from AI and tech advisory, on the firm’s own reporting. BCG reports 2024 revenue of $13.5B, up 10% year on year; AI consulting was 20% of 2024 revenue, projected to reach 40% by 2026. BCG’s X practice employs roughly 3,000 people across 80-plus cities and is growing toward 5,000. Deloitte Digital did $752.9M in 2024, up 10%.

The pattern is clean: the firms are not shrinking; their generic-work layer is shrinking, and their AI-advisory layer is growing. The people getting laid off are not the people getting hired.

The Indian IT picture is starker and has its own wrinkle. TCS announced 12,000 layoffs in July 2025, with headcount down 19,755 in Q2 FY26; Wipro is down 25,200 from its 2023 peak; Infosys down 12,506; the four largest combined are down roughly 42,000 over two years. TCS cited “skill mismatch” as the reason — not AI. The AI framing is analyst and press, not TCS’s own. The most honest reading is that a large fraction of Indian IT’s pyramid was built to do the kind of routine application-management and testing work that now automates cheaply. Whether you call that AI disruption or scale correction, the junior rungs are where the pressure lands.

The counter-evidence that this is not just cycle-driven compression is BCG’s own “Build for the Future 2025” report. Only 5% of surveyed firms were what BCG called “AI future-built,” capturing 5× revenue and 3× cost gains; 60% generated no material value. Gartner reports 72% of CIOs are breaking even or losing money on AI as of May 2025. Ethan Mollick, in a Stanford GSB talk in 2025, put it directly: “It used to be that, like, McKinsey or Ernst & Young, they knew everything. They don’t know anything anymore. Nobody has a playbook.” That is not a quote the consulting industry would have tolerated from a professor five years ago. It is one the industry is paying $20B a year to be told.

Marketing and creative: the holding companies break

The advertising holdcos are the clearest mid-tier-collapse story, because they are large enough to show up in public equities. WPP’s 2025 revenue was £13.6B, down from £14.7B in 2024; operating profit fell from £1.3B to £0.4B, margin from 15.0% to 13.0%. Headcount dropped 9% to 98,655. The company deployed 50,000 AI agents through its WPP Open AgentBuilder platform, and freelance usage declined 25% over two years. WPP’s share price fell roughly 60% across 2025. In September, Mark Read — who had run the company since 2018 — was replaced by Cindy Rose, a former Microsoft executive. By February 2025, Publicis had overtaken WPP by market capitalisation, which would have been unthinkable a decade earlier.

Omnicom and Interpublic announced their $13.25B all-stock merger in December 2024; it closed in November 2025. The combined 2024 revenue was $26B. The pre-close layoffs began almost immediately: IPG cut 3,200 jobs in 2025, Omnicom 800 plus a 10% reduction in staff costs. The publicly stated synergy target was raised from $750M to $1.5B over 30 months, of which roughly $1B will come from further job cuts.

Inside the client base, the demand signal matches. Gartner’s 2025 CMO Spend Survey — 402 marketing officers across North America, the UK, and Europe — found marketing budgets at 7.7% of revenue, down from 9.5% three years earlier. Thirty-nine percent of CMOs planned to cut agency budgets. Twenty-two percent reported that generative AI had reduced their reliance on external agencies for creative and strategy work. A separate Typeface survey of US senior marketing leaders put the latter number at 60%.

The canonical case is Klarna — and Klarna is also the cautionary tale, which is why it belongs here rather than in a triumphant footnote. In the first quarter of 2024, Klarna reported that its marketing spend was down 11%, of which 37% was attributable to generative AI — roughly $10M annualised — with external agency spend down 25%, image-production savings of $6M a year, and production cycles compressed from six weeks to seven days. That became the canonical case study for end-customer AI replacing creative work, cited in dozens of subsequent decks. In May 2025, Klarna’s CEO publicly admitted that the AI replacement of customer service “led to lower quality,” and the company began rehiring roughly 700 customer-service roles. Both halves of this story are true. Generative tools produced real savings in creative work; they degraded quality when deployed end-to-end in customer service. The lesson is that the bifurcation is subject-matter-specific: some domains are closer to one-hour photo (extinct at the mid-tier) than others, and the most confident pre-deployment forecasts are often wrong.

Structurally, the picture is the same as in consulting: revenue from generic creative is compressing; revenue from AI-native agency work is growing. Publicis reported that roughly 80% of its connected media is now AI-powered. The market compares a holdco that priced agency labour at its margin to one that priced AI deployment at its margin, and priced them differently.

Legal services: the top absorbs the agents

Legal is the cleanest “top tier absorbs the disruption” case, which is why Harvey is the lede of this essay and not its punchline. US legal services are a roughly $305B market in 2025, projected to reach $488B by 2035 on the middle of the analyst range; global legal is about $1.05T. Estimates diverge 40–60% across Grand View, Precedence, and Mordor, which reflects definitional issues about what counts as legal services, and should be treated as directional.

Harvey’s trajectory is the shape of an ambitious legal-tech company in the right moment. ARR went from $50M at the end of 2024 to $100M in August 2025 to $190M in January 2026. The Series F in December closed at an $8B valuation; in March 2026, Sequoia and GIC led a $200M round at $11B. The customer base is 1,300 firms, 100,000 lawyers, 60 countries, the majority of the AmLaw 100. Spellbook raised a $50M Series B at $350M in October 2025 with Khosla and Rabois; Ironclad reports $200M ARR as of January 2026, up 34% year on year; EvenUp closed a Series E at $2B. DocuSign paid $165M for Lexion; Workday acquired Evisort. Thomson Reuters reported 26% of legal organisations using generative AI in 2025, up from 14% in 2024; the ILTA 2025 survey put the figure at 80% among firms of any size and 100% among firms with more than 700 attorneys.

The money has gone to the incumbents. AmLaw 100 profit-per-lawyer is up more than 53% since 2019. Tech spend is up 10% year on year. Salaries are rising. The model — what the legal trade press is beginning to call the Harvey pattern — is that the incumbent becomes a co-developer and distribution partner, not a disruption target. A&O Shearman has 3,500 lawyers on Harvey since the 2022 beta; the firm shares Harvey’s software revenue on co-developed multi-step agents for antitrust filings, cybersecurity, fund formation, and loan review. The counterparty in a multi-billion-dollar merger still wants the name on the door. What they buy now includes software revenue share.

The bottom of legal services is also growing, through a different vector. LegalZoom cited a $48.7B total-addressable-market in its S-1. In March 2024, the National Association of Realtors’ $418M settlement effectively killed the 6% commission rule in US residential real estate; buyer-agent commissions compressed from roughly 3% to 2.55% within months, per Redfin’s measurements. Basic contract review, standard incorporations, residential conveyancing — the volume work — is shifting from mid-tier professionals to end-customer tools. The middle-tier law firm that made its money on commoditised work is in the squeeze. BigLaw is not.

There is one cautionary precedent that belongs here. In 2023, a New York attorney was sanctioned after filing briefs in Mata v. Avianca that included fabricated citations generated by ChatGPT. Two years on, the risk has not gone away. CiteME, a NeurIPS 2024 benchmark, found GPT-4o accurate at attributing claims to source papers only 35.3% of the time, against a human baseline of 69.7%. A firm using generative tools for drafting without a verification layer has a fiduciary problem, not a productivity problem. The top tier can afford the verification layer; the middle tier often cannot, which is another reason the compression lands where it does.

Developer productivity: the contested ground

The claim that AI makes developers dramatically more productive is the one piece of consensus that almost every industry report cites and almost no one has tested carefully. The data is more contested than most accounts admit.

GitHub’s 2022 Copilot randomised controlled trial — 95 developers writing a JavaScript HTTP server — found Copilot users 55.8% faster than controls, with a 95% confidence interval of 21% to 89%, p = 0.0017. The finding is real. It applies to a well-scoped greenfield task performed by unselected developers, over a session-length window. It is the highest-quality early evidence and has been cited more times than any other productivity study on AI coding tools.

METR’s July 2025 randomised controlled trial tried a different population. Sixteen experienced open-source maintainers worked on 246 real tasks in their own mature codebases, with and without Cursor and Claude 3.5 / 3.7 Sonnet. The maintainers expected to be 24% faster with the tools; they reported feeling 20% faster while they worked. They were on average 19% slower. Fewer than 44% of AI suggestions were accepted. The population was experienced; the codebases were mature; the tools were early 2025.

Both results are real. They are not contradictory — they measure different populations on different tasks. An unselected developer on a greenfield task is a different experiment from an experienced maintainer on a mature codebase, and generalising either finding is a mistake. METR itself, in a February 2026 update, reported that its larger August 2025 follow-up estimated participants were 18% faster in repeat use, and that the institute is revising its methodology. The honest conclusion is that the population-level productivity effect of AI coding tools on professional developers is not yet known. Anyone presenting a single headline productivity number should be distrusted by default.

The market has not waited for the answer. Cursor, built by Anysphere, went from $1M ARR in January 2024 to $100M in January 2025 to $500M in June to $1B in November to $2B in February 2026 — doubling roughly every two months for two years, the fastest path to $1B ARR in B2B SaaS history. Cursor reports more than 50% of the Fortune 500 and roughly 70% of the Fortune 1000 as customers. GitHub Copilot has 20 million users and 1.3 million paying seats at roughly $2B ARR. Devin (Cognition) grew from $1M to $73M ARR by June 2025; combined with the Windsurf acquisition, the company reports roughly $150M ARR at a $10.2B valuation. Devin’s self-reported metrics — a 67% pull-request merge rate; an 8×engineering efficiency gain at Nubank on a monolith refactor — are Cognition numbers, unverified by independent testing; Devin’s unassisted SWE-bench score at launch was 13.86%, and independent replication suggests heavy scaffolding is needed.

The benchmark leaderboards tell their own conflicted story. SWE-bench Verified scores climbed from 49% (Claude 3.5 Sonnet, October 2024) to 77.2% (Sonnet 4, October 2025) to roughly 87–94% on top of the leaderboard by April 2026. Those numbers were the headline of 2025. They are less impressive on inspection. In September 2025, Scale AI released SWE-bench Pro, an uncontaminated, multi-language version of the benchmark requiring changes of at least ten lines of code. The same Claude Opus 4.7 that scores 87.6% on Verified scores 45.9% on Pro. OpenAI’s own audit of Verified found that frontier models could reproduce verbatim gold patches on some tasks, and 161 of 500 Verified tasks are one or two lines of code. Verified is saturated; Pro is the harder measurement, and it says capability is closer to 46% than 94%.

End-customer tools: the bottom grows

The other end of the barbell is louder and newer. Meta’s Advantage+ end-to-end AI ad products reached a $60B annual run-rate by the third quarter of 2025, per Mark Zuckerberg on the earnings call; more than four million advertisers used the generative creative tools, up from about one million six months earlier. Thirty-five percent of US retail advertising spend ran through Advantage+ in Q2 2025. This is the most concrete single proof that end-customer self-service is replacing a slice of professional media-buying work at scale.

Lovable reached $100M ARR in eight months post-launch, targeting non-technical founders who want to build web applications. Replit went from $10M to $100M ARR in nine months after launching Replit Agent. Cursor’s trajectory is above; its users are developers, not laypeople, but the product functions as a convenience-tier substitute for work that mid-tier contract developers would have billed for five years ago. n8n — a workflow-automation platform targeting end-customers who want to build their own agent systems — raised $180M at a $2.5B valuation in October 2025, with 230,000 active users and 3,000-plus enterprise customers; YipitData’s panel shows n8n’s mid-market customer count grew roughly 10× over the year ending January 2026, with about 80% of new customers migrating from Zapier.

The signal at the individual-founder level is the most interesting and the most easily overstated. Stripe Atlas reported that AI startups rose from 15% of its founders in January 2023 to 42% in 2025; its top-100 AI companies hit $1M ARR in a median of 11.5 months, four months faster than the fastest-growing SaaS of the subscription boom. The Stripe Indie Founder Report found 44% of profitable SaaS products are now run by a single founder, double the 2018 share.

In September 2024, Sam Altman told Alexis Ohanian that his tech-CEO friends had started a betting pool for the first year a one-person billion-dollar company emerges. As of early 2026, no company qualifies. The closest public examples — Pieter Levels’s Photo AI at roughly $1.6M ARR, David Bressler’s Formula Bot at $2.8M, Ivan Kutskir’s Photopea at $2.4M, Nick Dobos’s BoredHumans at roughly $8.8M — are two orders of magnitude short. The distribution has shifted; the tail claim is unverified.

The food industry already ran this play

The parallel that reveals the most about what happens next is not a technology parallel. It is a restaurant parallel.

US restaurant sales went from roughly $580B in 2010 to $1.1T in 2023 to a projected $1.5T in 2025. The restaurant count grew from about 600,000 in 2010 to somewhere between 700,000 and one million by 2024, depending on definitions; employment rose from 14.2M to 15.9M. Food-away-from-home hit a record 58.9% of total food spending in 2024. Per-capita real food-at-home spending declined 2.3% in 2022 and 3.1% in 2023. Americans were cooking less at home in real terms even as the restaurant industry grew. The category, in aggregate, did not shrink. It restructured.

Inside the aggregate, the mid-tier collapsed. TGI Fridays operated about 600 US units in 2008; 384 in 2020; 163 when it filed for Chapter 11 in November 2024; 85 by April 2025. Red Lobster filed for Chapter 11 in May 2024, closed 130-plus locations, was acquired by Fortress Investment Group with a $70M injection. Applebee’s went from about 2,000 units in 2008 to 1,567 at the end of 2025, with 300 closures since 2017 and 20-to-35 more guided for the following year. Hooters, Bar Louie, Buca di Beppo, and On the Border all filed in 2024 or 2025. Business-and-industry foodservice — the cafeteria work that serves captive corporate demand — “has been the slowest segment to recover from the pandemic due to slow return-to-office trends,” in one industry trade’s phrasing. Aramark’s recovery came from sports, entertainment, and education. The corporate cafeteria never came back.

Above the middle, fine dining grew. US fine-dining revenue went from $14.6B in 2019 to $16.7B in 2024 to a projected $17.2B in 2025; unit counts rose from 4,618 in 2023 to 4,771 in 2025, a 2.8% compound annual growth rate. Global fine dining was $166.9B in 2024, projected to reach $243.2B by 2030 — a 6.5% CAGR. Michelin three-star restaurants grew from 100 in 2010 to 157 in August 2025; the guide expanded to Argentina, Vietnam, Malaysia, and several new European and North American markets in 2025 alone. Atomix’s tasting menu in New York went from $175 in 2018 to $395 in 2024. The New York omakase count went from a handful in the mid-2010s to roughly 140 by 2024; the most exclusive now require $10,000 memberships. At the close of 2024, René Redzepi announced the end of Noma’s regular-service model, publicly declaring fine-dining labour economics unsustainable at $500–800 per person. That is a labour-cost problem, not a demand problem.

The meal kits are the cautionary tale. Blue Apron went public in 2017 at roughly a $1.9B valuation; it was acquired by Wonder Group in November 2023 for about $103M — a 95% destruction of public-market value. HelloFresh peaked at €7.6B in revenue in 2022; orders fell 4% in 2024; the stock is down 93% from its 2021 peak, and the company abandoned its €10B revenue target and is pivoting to ready-to-eat meals. Global meal-kit penetration is roughly 0.3% of households. The product was sold as a restaurant replacement. It was a niche.

Delivery is the real structural disruption. Delivery’s share of total consumer foodservice went from 9% in 2019 to 21% in 2024 on Euromonitor’s count; it is projected to reach 24% by 2029. Roughly 75% of US restaurant traffic is now off-premises — takeout, drive-thru, or delivery. DoorDash holds 60–67% of the US market, Uber Eats 23–26%, Grubhub 6.2%. Grubhub sold to Wonder in 2024 for roughly $650M — a fraction of the $7.3B Just Eat paid for it in 2020.

The speculative layer collapsed. CloudKitchens raised at a $15B valuation in 2021; Kitchen United shut down in December 2023; Wendy’s cancelled a 700-unit Reef deal; Reef itself closed facilities in New York, Portland, and Philadelphia. More than $3B of ghost-kitchen funding from 2020 to 2022 produced limited operating substance. Instant Brands — whose Instant Pot multi-cookers sold $758M in 2020 — filed for Chapter 11 in June 2023 after sales fell to $344M in 2022, a 55% drop in two years. Home-cooking hardware as a category did not survive the return of restaurant spending.

What worked: experience-driven tasting concepts, hyper-specialised single-product formats, delivery-native quick-service, and premium grocery prepared foods. Chipotle’s digital sales went from under 10% in 2018 to roughly 35% post-2020. Trader Joe’s opened 34 new stores in 2024, triple its 2023 pace; Whole Foods holds more than 500 stores; Erewhon’s $20 celebrity smoothies have become a cultural reference point. Starbucks grew from about 11,000 US stores in 2010 to 16,864 in fiscal 2025. Specialty coffee is an $83.6B global market growing at 11–12% CAGR. Home espresso grew massively, and Starbucks grew anyway. The value was ritual, location, and speed. It was not the drink.

The barbell holds everywhere

The food-industry pattern is not an analogy. It is the same pattern, run in a different category a decade or two earlier. The pattern holds everywhere convenience has been unbundled by technology.

The timing is remarkably consistent. From first technology introduction to full industry restructuring runs 13–18 years. Kodak introduced digital imaging in 1994; the 1-hour photo extinction was complete by 2013. Napster launched in 1999; the streaming transition was done by 2015. Expedia launched in 1996; Blockbuster filed for bankruptcy in 2010; Netflix, founded in 1997, became larger than Blockbuster had ever been within the same window.

The employment decline is a function of the regulatory moat. Where there is no moat — photo, video rental — extinction at the middle runs 85–95%. Where there is a light moat — travel agents, bookkeepers — decline runs 40–70%. Where there is a strong regulatory moat — lawyers, CPAs, licensed realtors — the outcome is fee compression and 10–30% headcount pressure, not extinction. Agencies and consultancies have only the light moat: reputation, scale, client relationships. No one needs a licence to practise strategy. The expected structural outcome at the middle tier is a 40–70% employment compression, not the 10–30% regulated-profession outcome.

Ben Thompson’s Aggregation Theory is the strongest available explanatory frame. Aggregators commoditise suppliers, own the demand side, and destroy the intermediaries. The new middleman — Booking, Spotify, Intuit, LegalZoom, Meta’s ad stack — ends up larger than the one it replaced, serving orders of magnitude more customers at lower per-unit cost. Consumer surplus is real. It concentrates at the aggregator.

Counter-evidence: are agents actually good enough?

The strongest objection to the mid-tier-collapse thesis is that the replacement agents are not yet good enough to do the work. The evidence on this is less reassuring than headline benchmark scores suggest, and more reassuring than RCT field trials alone would suggest.

TheAgentCompany, the Carnegie Mellon benchmark released in December 2024, is the most realistic measurement available of whether an agent can replace a junior knowledge worker end-to-end. Claude 3.5 Sonnet, the frontier model at release, completed 24% of real office workflows autonomously. Gemini 2.0 Flash completed 11.4%. Llama 7.4%. The mid-2025 rerun put the best agent at about 34% full completion, 39% partial. The answer to “can an agent replace a junior office worker today” is, in the most honest available measurement: not at 90%; not at 50%; currently about one task in three.

SWE-bench Pro, discussed above, puts frontier coding capability at 46% on a harder, uncontaminated harness. TheAgentCompany puts end-to-end office-task completion at 34%. Gartner projects that more than 40% of agentic AI projects will be cancelled by the end of 2027 for unclear ROI; 72% of CIOs say they are breaking even or losing money on AI. BCG’s “Build for the Future” put AI future-built firms at 5% of the surveyed population. This is not a picture of uniform capability uplift.

The qualification is that the mid-tier collapse is not contingent on agents replacing every junior worker. It is contingent on agents replacing enough of the routine work that the firm’s cost structure has to be rebuilt. Even at 34% end-to-end completion, agents are doing measurable work that previously sat on a billable rate card. Even at 46% on hard-coding benchmarks, they are substituting for a non-trivial share of what used to be junior-associate-billable hours. At 35% of US retail ad spend, Advantage+ is replacing a substantial share of what used to be agency media-buying revenue. The threshold for structural compression is not 100% capability. It is a capability level high enough that the old per-seat economics stop clearing. That threshold has been crossed in more than one sector.

Closing

Every prior convenience category has followed the same arc. Travel agents did not disappear; the mid-tier storefront did. One-hour photo did not survive the smartphone; the professional studio did. Blockbuster died; Netflix was larger within a decade than Blockbuster had ever been. The middle is where the pattern hits first and hardest.

What survived above the middle in each case was not effort, or prestige, or intelligence. It was the thing a customer could not produce for themselves. The Michelin three-star restaurant survived not because the food was technically better than what is available in an Applebee’s — much of the technical work in fine-dining is within the reach of a skilled home cook with a sous-vide and four hours. It survived because the experience could not be assembled at home. The travel agent who arranges a complex multi-country itinerary, with language support and on-the-ground fixers, survived because Expedia could not replicate what her network offered. The corporate lawyer at an AmLaw 100 firm survives because, when the merger of a multi-billion-dollar business is at stake, the counterparty wants the name on the door, not the draft contract.

In knowledge work, the question is what that thing will turn out to be. The answer is likely to vary by field. Some fields — accounting, basic contract review, SMB advertising — will look more like one-hour photo than like travel agency. Others, the fields in which the value lies in accountability, in taste, in relational trust, in the kind of judgement that is not reducible to a prompt, will look more like fine dining. What knowledge workers inside the pattern should ask themselves, on Monday morning, is not whether their category will be disrupted. It is whether the part of their work that a customer can now buy on the margin is the part that was paying the rent.