My earlier article You Can’t Drink Data made the argument that data centers are being built in our driest areas, drawing cooling water from strained systems - all the while the data center industry calls its work “water positive”. On August 21st, the U.S. government ordered Arizona, California, and Nevada to reduce their draw from the Colorado River for the next 2 years.[1] The river supplies water to 40 million people across 7 states, 30 tribes, and part of Mexico. Its 2 reservoirs, Lake Mead and Lake Powell, are at record lows, and, if the last 20 years of drought are an indication, we should not expect replenishment.

This is a critical event for all of us. I am not arguing for the us to stop using AI. AI usage inevitable, and it will do wonderful things. What I want to demonstrate here is: when your organization makes decisions on how it will use AI, it is making a decision that has trade-offs with water availability, food supply, and clean energy. That has never been true of marketing and media before. It is true now. And on my priority list, those three come first. AI is fourth.

The real story here is what comes next after losing a river. What follows are 3 dependencies - livestock, crops, and electricity - and each ends with burning more fossil fuel. Losing water creates demand, some would call necessity, to burn more. And burning more fossil fuel is a choice that data centers have already made.

The Herd
Cattle need water - directly, and through irrigated feed and forage that Western ranching runs on. As the water goes, herds shrink; the U.S. cattle herd is already near multi-decade lows, with drought a leading cause. Market economics’ principle of scarcity says that the price of beef will go up, and it has. However, this is really a supply story. When domestic supply contracts, imports fill the gap, and the world’s biggest beef exporters are not next door: Brazil is 1st and Australia 2nd.[2] Every pound of beef that shifts from a U.S. ranch to a freighter from Sao Paulo or Brisbane carries a fossil-fuel bill that we will pay for now and in the future - shipped across oceans, frozen the entire way. In the end, food gets less affordable, and more carbon-intensive, at the same time. Some might say, we should just raise more domestic cattle, but remember there is not enough water.

The Fields
I have not intention of wading into the agribusiness debate here. If you want to know more, I recommend Pulitzer-winner Art Cullen’s reporting on Raccoon River and agricultural run-off. I’ll just stick to the mechanics of fertilizer. Industrial grade fertilizer is nitrogen based to stimulate growth. Fertilizer is not a substitute for water. You cannot spread nitrogen on a dry field and expect the same as a wet-field yield. What happens as irrigated Western acreage shrinks is that production pressure shifts east, onto (hopefully) rain-fed Midwest acres - which introduces more use of synthetic nitrogen to squeeze more yield from the same acres. That nitrogen is made from natural gas - and a lot of it. When gas prices spiked after Russia’s invasion of Ukraine, ammonia prices (from which the nitrogen is derived) roughly tripled, which is how tightly the two are linked.[3] And nitrogen doesn’t stay in the fields. Iowa’s own Department of Natural Resources lists more than 700 rivers, lakes, and wetlands as impaired, roughly 50% of what was assessed.[4] Follow the run-off down the Mississippi and you reach the Gulf’s Dead Zone, a hypoxic area NOAA measures every summer where aquatic life cannot survive.[5] Less water means more fertilizer usage, which means more fossil fuel, which means less swimmable, fishable, and drinkable water somewhere downstream. We solve one water problem and create 3 more.

The Grid
Hydropower supplies about 6% percent of U.S. electricity and about 25% of all renewable energy generated.[6] The Colorado River’s dams are a part of that, and are in the exact region where AI load is growing the fastest. The Hoover Dam’s output has already been derated (i.e. producing below capacity) by Lake Mead’s decline, and Lake Powell has flirted with “minimum power pool” status, the level below which turbines simply stop. Data centers add demand to the same grid at the same moment. Less water means less clean energy, which means more fossil fuel - in the region that can least afford to the heat.

One Big Loop
Let’s tie this all together. Losing water raises the price of food and adds a carbon cost of importing. It pushes toward more fossil fuel-derived fertilizer and produces dirtier water downstream. And it removes clean energy from the grid, where new demand is growing, handing the gap to fossil fuels. Every water loss ends in a decision to burn more, and burning more begets more aridiation. The Southwest is unfortunately living through a drought that started at the turn of the century. We hope it will break, but until then, the region is living in a loop that tightens, a noose.

Where AI Comes In & What I’d Ask of You
AI did not cause the megadrought. And let’s not forget all the amazing things that AI will do for society. But AI is the largest claimant on both our scarce resources in this story - water and electricity. AI is taking these resources with limited resistance. While local governments are starting to ask the right questions and take action, the right choices often compete with need to fund local budgets in dire straits. And for-profit entities are focused on growing their immediate term profits, not whether the communities they serve will be healthy enough to use nor have income to afford their products and services. So here is the ask of you, and it is directed at all of us who use AI, not just those that build data centers: be purposeful: Every AI task has a physical cost - water through cooling systems, and energy through the grid. Some activities will be worth it. Some activities just won’t. The question I used to ask with this technology is: what is it for? Now, I also ask: Is it worth it?

We have lost enough with the loss of the Colorado River. I would not want your river to be next.

A note on how this was written
I wrote this on my own. However, I do not have a human editor nor research assistant, so I used Anthropic’s Haiku 4.5, purposely using a lower grade model to minimize my impact, to review my drafts for structure and grammar and assure my facts and attributions were accurate. I am on a journey and do intend to use less and less AI as I go.

Notes
[1] U.S. Bureau of Reclamation, interim operating plan announced August 21, 2026: mandatory Lower Basin reductions of 1.25 million acre-feet per year for 2027–2028 (Arizona 760,000; California 440,000; Nevada 50,000), plus at least 700,000 acre-feet of additional voluntary conservation; Mexico reduces 250,000 acre-feet under treaty. Percentage impacts per Sarah Porter, Kyl Center for Water Policy, ASU (via CNN, Aug. 21, 2026). See also Washington Post and CBS/AP reporting, Aug. 21–22, 2026.

[2] USDA Foreign Agricultural Service data via Statista, 2025: Brazil is the world’s largest beef exporter (~3.75 million tons forecast for 2025); Australia is second by volume, with India (carabeef) and the United States close behind. The U.S. is simultaneously a top-four exporter and a major importer of lean beef from Australia, Brazil, and New Zealand.

[3] Synthetic ammonia is produced from natural gas via the Haber-Bosch process; fertilizer prices roughly tripled following the 2021–22 natural-gas price shock after Russia’s invasion of Ukraine (World Bank commodity data; USDA ERS fertilizer price reporting).

[4] Iowa Department of Natural Resources, draft 2026 Integrated Report: 723 impaired segments (574 Category 5, 149 Category 4); roughly 51% of assessed river segments impaired. Via Iowa Capital Dispatch, Feb. 17, 2026. From Iowa State University: Waterbodies can be listed as impaired if they do not meet water quality standards for just one of their designated uses. There are ten possible uses, but the most common uses are for recreation, aquatic life, human health, or a combination of all three. Recreation refers to waterbodies where the primary use is swimming or water sports where a high degree of bodily contact with the water, or ingestion is likely. Aquatic life refers to waterbodies where the primary use is protection and propagation of fish, shellfish, and wildlife.

[5] NOAA, annual Gulf of Mexico hypoxic zone (“dead zone”) measurement, driven primarily by nutrient runoff carried down the Mississippi River.

[6] U.S. EIA, Short-Term Energy Outlook, 2026: hydropower expected at ~259 billion kWh, about 6% of U.S. generation; DOE Water Power Technologies Office: hydropower was 27% of U.S. utility-scale renewable generation in 2024. On Hoover derating and Lake Powell minimum power pool: Bureau of Reclamation operational reporting, 2022–2026.

Photo: Colorado River Runs Dry, January 13, 2009, Pete McBride, U.S. Geological Survey (USGS)

The Cost of Energy is the Decoy

When a data center lands on a grid, the public conversation goes quickly to the pocketbook: will my bill go up? Usually, yes. Residential electricity prices have climbed by roughly a third since 2019, ahead of inflation. ^[1] The anger is justified. It is also the half of the story companies are happy to let you have, because the other half is a non-refundable environmental wreckage. While the rate fight absorbs the public’s attention, the supply story unfolds quietly beside it. New power plants are being approved on expedited timelines, under NDAs, without legislated hearings and reviews, and they are overwhelmingly being sourced by fossil fuel.

Start in El Paso

Consider Meta, whose two current projects bracket the scale of what is happening.

Outside El Paso, Meta is building a $10 billion data center that will be powered by a dedicated 366-megawatt gas-fired facility. [2] Run around the clock, that one plant will put out the exhaust of roughly 350,000 additional cars while generating roughly the electricity a city of 300,000 people uses. El Paso itself has 683,000 people. ^[3] This is a simple example.

But El Paso is just the warm-up. In Richland Parish, Louisiana, Meta is building a $50 billion data center across roughly six square miles. A deal The New York Times reported that was negotiated in secret and approved without the administrative-law-judge review such cases normally receive. [4] To serve it, Entergy will deliver enough power to run New Orleans seven times over. ^[4] There are claims of offsets, carbon capture, and customer savings. ^[5][6] But still, it will emit 12 to 13 million metric tons of CO₂ a year — the tailpipes of about 2.7 million additional cars. ^[3] And here is where it gets complicated. The gas will come from Louisiana’s Haynesville Shale. Fracking there uses so much freshwater that the state issued a Ground Water Use Advisory more than a decade ago ^[7], and a magnitude 4.9 earthquake this March — the largest ever recorded on Louisiana soil — is suspected, though not yet established, to be tied to the oil and gas operations. ^[8]

This is not just a Meta story

What Meta illustrates is happening across the industry. Among the major ad-selling platforms, emissions are rising everywhere they are disclosed:

The disclosures themselves deserve credit, and there are good actions being taken. Microsoft paused its purchases of unbundled renewable-energy certificates — accepting a spike in reported emissions rather than keep leaning on paper offsets. ^[10] Google concedes plainly that its climate impact has grown alongside AI, while noting that without its decarbonization work the footprint would be roughly five times larger. ^[12] The outlier is X, which publishes no verified climate reporting at all, and whose xAI affiliate has powered its Memphis-area “Colossus” data centers with fleets of unpermitted gas turbines — as many as 35 at the first site, dozens more across the Mississippi line — now the subject of Clean Air Act litigation brought by the NAACP. ^[13]

But disclosure is not the deepest problem here nor are the offsets bought in places that will never breathe the smog. What is wrong — dead wrong — is the direction. Fossil fuel use and creation ought to be falling. Instead new fossil demand is being created at the scale of our largest cities, several times over, and the official U.S. government forecast says the pattern holds. The Energy Information Administration projects data-center server consumption reaching 818 billion kilowatt-hours by 2050, roughly sixteen times the 2020 level, with natural gas still around 40 percent of U.S. generation in mid-century. ^[14] Since late 2022 — the start of the AI build-out — planned non-renewable capacity additions have grown 71 percent, while planned renewable additions grew 2 percent. ^[15] That shift is not the grid at large, which is adding a record 86 gigawatts of capacity in 2026, more than 90 percent of it solar, storage, and wind, with gas at just 6.3 gigawatts. ^[16] The fossil turn belongs to the data-center build, and it is accelerating: the single largest announced project — a 10-gigawatt data-center campus at Piketon, Ohio, backed by SoftBank — comes with 9.2 gigawatts of on-site gas, enough for five to seven million homes. ^[17]

Why this is this Marketing and Media’s to lead

The companies at the center of the build-out — Alphabet, Meta, Amazon, Microsoft, X — are also among the largest sellers of advertising in the world. The same marketing industry pays them twice: once for media, and again, increasingly, for the AI and cloud capabilities that power their employees’ commercial and R&D work. The reliability of that ad money is what underwrites Wall Street’s conviction in the build-out. The advertising industry is not a bystander to this construction. It is the revenue at its foundation.

The industry has forced supply side transparency before. In 2016 the Association of National Advertisers (ANA) commissioned K2 Intelligence to examine the media supply chain; the report found non-transparent practices pervasive, and by 2019 two-thirds of marketers had rewritten their media-agency contracts. ^[18] In 2023 the ANA went further, putting log-level data under audit and finding that roughly 36 cents of every dollar entering a demand-side platform reaches a consumer. ^[19] The industry asked, in writing and at scale, where the money went, and the market reorganized around their own answers. None of this directs a dollar. It puts a fact on the table and leaves every buyer to decide alone on the future of they will support — the same design that made the media-transparency push both effective and legally boring.

What I’d ask of leaders

Ask the questions. Everyone now asks what an AI capability costs. Almost nobody asks what has to be built for it to exist. If your roadmap assumes a doubling of computing, something physical is being built somewhere to serve it — and right now, it is mostly being built with fossil fuels and aquifers.

Add two fields to the RFP. Where does the power come from, and where does the water come from? It costs nothing to ask, it is legally uncontroversial, and if enough buyers ask, it stops being a question and becomes a specification.

Take it to the trade bodies. Ask the ANA, 4As, and IAB to write energy- and water-sourcing disclosure into its RFP standards. Keep it strictly to disclosure — facts each vendor already holds, no direction of spend.

Where this leaves us

None of this is an argument against data centers. The build-out is coming. The argument is about what powers it, because for the first time in industrial history the choice is real. Sun and wind, firmed with storage, burn no fuel and consume almost no water. ^[21][22] Gas keeps winning these contracts anyway, because it is quick to permit, it runs all night, and current legislation favors it — a preference nobody has yet had to defend against its full cost. What the alternatives lack is not feasibility; the 2026 grid additions settle that. It is demand expressed as specification: buyers, at scale, asking the question in writing.

The advertising industry funds the companies doing the building, contracts with them twice over, and has already proven that a disclosure standard can move an entire supply chain. It knows how to do this. The only decision left is whether to ask before the plants are built, or to audit the damage after. The next RFP is the place to start.

A note on how this was written

The argument here is mine: that the energy debate around data centers is being fought over rates while the supply decision — what gets built, burning what — goes largely unexamined; that marketing money funds the build-out twice over; and that the industry’s own transparency precedent, from K2 to the ANA’s programmatic study, is the template for asking the question. I used Claude to assist in research, check cited sources and to help structure and edit the prose. The judgments are my own.

Notes

1. Residential retail electricity prices rose roughly a third nationally from 2019 to early 2026 (EIA retail price data; the BLS consumer electricity series puts the five-year rise near 39 percent), against cumulative all-items inflation of about 29 percent. Whether data centers drove the national increase is contested — Lawrence Berkeley National Laboratory and EPRI analyses find load growth was not the principal driver through 2024 — but the regional link is direct: PJM’s Independent Market Monitor reported in 2025 that data-center load growth is the primary reason for the interconnection’s record capacity prices, three years running (via PolitiFact, June 12, 2026).

2. El Paso Matters, “El Paso Electric filings detail power plant impact behind Meta’s $10 billion data center,” March 29, 2026; Inside Climate News / The Texas Tribune, “Plan to run El Paso data center on natural gas sparks concern,” January 26, 2026; El Paso Inc., June 2026, on the city’s opposition filing. Source for: the project’s growth from $1.5 billion to $10 billion; the 366-megawatt McCloud facility of 813 modular gas-fired generators (~$473–500 million) on 31 acres adjacent to the site; the initial “bridge period” under which Meta pays all costs; and the City of El Paso’s intervention before the Public Utility Commission of Texas, including City Attorney Karla Nieman’s “half-billion-dollar power plant for a single new customer” statement.

3. My estimates, method in full. Louisiana: 5,200 MW × 8,760 hours × ~75 percent capacity factor ≈ 34 TWh/year — about a third of Louisiana’s ~100 TWh of 2025 generation — multiplied by EIA’s combined-cycle natural gas emission rate of roughly 0.37 metric tons CO2 per MWh ≈ 12–13 million metric tons per year. El Paso: 366 MW of modular simple-cycle generation at a comparable capacity factor ≈ 2.4 TWh/year; at simple-cycle emission rates of roughly 0.6–0.7 tons per MWh ≈ 1.5–1.7 million tons per year. Car equivalence throughout at EPA’s 4.6 metric tons per typical passenger vehicle per year. Neither utility has published an official emissions figure; replace with docket figures if and when they appear. Scope: the Louisiana figure counts only the seven plants announced in March 2026 (5,200 MW; Entergy Louisiana announcement, March 27, 2026). Including the original three combined-cycle units (2,260 MW; Entergy Louisiana, December 2024) raises the total to roughly 17–18 million metric tons a year. Piketon homes equivalence: 9.2 GW at a 65–85 percent capacity factor ≈ 52–68 TWh/year; at EIA’s average household use of ~10.5 MWh ≈ 5–6.5 million homes, and up to ~7 million on instantaneous-load methods.

4. Eli Tan and Maureen Farrell, “How Meta Got Everything It Wanted in a Secret Louisiana Data Center Deal,” The New York Times, July 27, 2026; and Eli Tan, “5 Things to Know About Meta’s Giant Data Center in Louisiana,” July 27–28, 2026. Source for: roughly six square miles; turbines described as enough to power New Orleans seven times over; ten turbines at an estimated $14 billion; the $50 billion project total; more than half of Entergy’s supply; the skipped administrative-law-judge review and Commissioner Davante Lewis’s remarks; Meta as the largest customer in Entergy’s 112-year history; Entergy stock up more than 50 percent; the four-year initial lease term.

5. Power Engineering, December 2025, on Entergy’s efficiency and retirement arguments, the claim that carbon sequestration could offset ~60 percent of new gas emissions, and the plants’ ability to serve all customers if Meta departs.

6. Entergy Louisiana, “Entergy Louisiana announces a new agreement with Meta that will deliver an additional $2B in customer savings,” August 2026: full cost of service; ~$2 billion over 20 years on top of $650 million; support for up to 2,500 MW of additional solar.

7. Louisiana Office of Conservation (Commissioner James H. Welsh), Haynesville reporting for October 2009 through February 2011: average ~4.93 million gallons of water per well; groundwater as the customary source; the Ground Water Use Advisory and frac-water reporting directive. Presented via US EPA.

8. NOLA.com/The Advocate, May 3, 2026, and Houston Chronicle reporting, April 2026, on the March 5, 2026, magnitude 4.9 Haynesville-area earthquake — the largest recorded on Louisiana soil (a 5.3 struck offshore at Grand Isle in 2006) — the USGS and state investigation, prior USGS attributions of near-border quakes to fracking and wastewater disposal, and EPA’s direction to Louisiana regulators on brine injection. The attribution is suspected and under investigation, not established.

9. Amazon, 2025 Amazon Sustainability Report: operational electricity matched 100 percent with renewable-energy contracts for a third consecutive year; absolute greenhouse gas emissions up 16 percent year over year to 81 million metric tons CO2e; the report’s language on the “momentum and complexity” of AI outpacing efficiency gains.

10. Microsoft, 2026 Environmental Sustainability Report, with prior-year reporting: total Scope 1–3 emissions of 20.3 million metric tons, up 25 percent year over year and a 23.4 percent net increase from the 2020 baseline; the decision to pause purchases of unbundled renewable energy certificates — accepting higher near-term reported emissions — in favor of funding net-new clean power.

11. Meta, 2025 Sustainability Report; Interfaith Center on Corporate Responsibility proxy memorandum, May 2026: total emissions up 38 percent over the company’s baseline; location-based grid-electricity emissions up nearly 150 percent, 2019–2024; roughly 15+ gigawatts of renewable capacity added to global grids over the last decade; investor proxy pressure to reconcile 2030 net-zero targets with fossil-fuel backstops for AI clusters.

12. Google, 2026 Environmental Report; ESG Dive reporting: “ambition-based” emissions up 18 percent year over year and 81 percent against the 2019 baseline; total grid electricity consumption up 37 percent; the company’s acknowledgment that its climate impact has grown alongside the growth of AI, and its estimate that, absent its decarbonization initiatives, the footprint would be roughly five times larger.

13. On X and xAI: Southern Environmental Law Center and Earthjustice, on behalf of the NAACP and its Mississippi State Conference, Clean Air Act litigation, 2025–2026: as many as 35 unpermitted methane gas turbines at the original Colossus site in Memphis (removed after a notice of intent to sue; 15 subsequently permitted); dozens more installed in Southaven, Mississippi, to power Colossus 2 — Reuters-reported records put the count at 59 by mid-2026, roughly double what the company had acknowledged; EPA’s January 2026 determination that such turbines require permits. X publishes no standard, verified corporate sustainability reporting.

14. US EIA, Annual Energy Outlook 2026 (April 2026): data center server consumption reaching 818 billion kWh by 2050, roughly sixteen times the 2020 level; total capacity up 50–90 percent by 2050; natural gas around 40 percent of generation in the baseline at mid-century; nuclear generation essentially flat with its share declining from 17 percent; small modular reactors outside the model.

15. American Action Forum analysis of EIA’s Preliminary Monthly Electric Generator Inventory, December 2022 versus February 2026: net +71 percent planned non-renewable capacity additions versus net +2 percent renewable.

16. US EIA, “New U.S. electric generating capacity expected to reach a record high in 2026,” February 20, 2026: a record 86 GW of planned utility-scale additions — solar 51 percent, battery storage 28 percent, wind 14 percent — with 6.3 GW of natural gas, more than 80 percent of it in Texas, Oklahoma, Ohio, Tennessee, and Florida. EIA’s generator inventory shows the gas wave building behind it: planned combined-cycle additions rise from 3.3 GW in 2026 to roughly 10.6 GW in 2028.

17. DOE/DOC announcements and AP reporting, March 20–21, 2026, on the PORTS Technology Campus at Piketon, Ohio: a 10 GW data center; 9.2 GW of on-site gas at $33.3 billion, funded under the U.S.–Japan Strategic Trade and Investment Agreement.

18. K2 Intelligence for the ANA, “Media Transparency Report,” June 2016; subsequent ANA reporting on two-thirds of marketers updating media-agency contracts by 2019.

19. ANA, “Programmatic Media Supply Chain Transparency Study,” December 2023: roughly 36 cents of each dollar entering a demand-side platform reaches the consumer; roughly $22 billion in available efficiency.

20. University of Oxford, Department of Engineering Science, case study on data center water: Microsoft’s assessment of its San Antonio facility found the true cost of the water roughly eleven times what it was paying.

21. J. Macknick et al., “A review of operational water consumption and withdrawal factors for electricity generating technologies,” Environmental Research Letters 7 (2012): median operational consumption, cooling-tower nuclear 672 gal/MWh (range 581–845); natural gas combined cycle ~200; utility-scale PV ~1; wind ~0; concentrating solar (tower) ~906.

22. On withdrawals: EIA, “U.S. electric power sector continues water efficiency gains” (2021 data: gas combined cycle 2,803 gal/MWh withdrawn; coal 19,185); once-through nuclear withdrawal ranges of ~25,000–60,000 gal/MWh, and nuclear’s ~40 percent share of US thermoelectric withdrawals, per EPRI and World Resources Institute compilations.

I’ve also come to recognize the show as an allegory for the AI enterprise experience. For the past few years, business has been the homeowner: confident, capable enough to be dangerous, letting its people loose on AI tools. The early wins were real. The bills are now arriving. So this article is about the inflection point every DIY story turns on. When do you let your people use AI? When do you build your own? And when do you move to a professional-grade build? Get the timing right and it’s an upgrade. Miss it, and the move-in date keeps drifting.

Give the build-out its due

In 1951, the economist George Stigler argued that young industries integrate vertically because they have no choice. The specialist suppliers don’t exist yet, so the pioneer builds everything itself. [8] Intel is the modern textbook case. For decades it designed and manufactured its own chips, because nobody else could, and that integration was the moat. AMD’s founder had a line for the era: “real men have fabs.” [9] (Fabs are semiconductor manufacturing plants.)

Then, as an industry grows, the market gets big enough to support specialists, and the specialists get better than the in-house version because they see more volume and more variety than any single firm can (I wrote about this dynamic in In-Housing: What to Own, What to Rent). In 1987, TSMC opened as a foundry that designed nothing and manufactured for everyone. Nvidia, the most valuable chip company on earth today, has never owned a fab. AMD sold its fabs in 2009 and thrived. Intel, the pioneer whose integration built the industry, fell behind the specialist and now pays TSMC to make some of its own chips. [9] The fabs were the right answer for forty years, right up until the arithmetic moved. The company that missed the flip became the cautionary tale.

Now fast-forward to AI. In 2023 there was no vendor who could honestly claim expertise, and none worth buying services from. The most sensitive data couldn’t be shared without compromising confidentiality. And an organization had to learn AI by having its own experiences with it. So no, the in-house build-out wasn’t a mistake. It was the fluid phase of a young market: you make, because you can’t yet buy.

Today we are at the end of this early phase, as the receipts start coming in.

The receipts arrive

SG&A, the P&L line where marketing, sales, and the corporate infrastructure live, is the cost AI was supposed to deflate. That was the pitch on every slide: automate the overhead, bank the savings.

What happened was the opposite. In December 2025, The Hackett Group studied the 1,000 largest US public companies and found median SG&A rising from 13.7% to 14.3% of revenue, a five-year high. Sixty-two percent of companies saw SG&A grow as a share of revenue, and 78% couldn’t keep cost growth below the 2.9% inflation rate. [1] Hackett attributes the rise largely to slowing revenue growth, not to AI, and maybe that’s right. But the great AI deflator has now been deployed at historic scale for three years, and the line just hit a five-year high anyway. The deflation is missing. Here is where it went.

1. The tokens ran away first. Uber’s CTO disclosed that the company burned through its entire 2026 AI coding budget in four months. It wasn’t for lack of adoption; 84% of engineers used the tools, and 70% of committed code was AI-originated. The problem was that the COO had to concede, publicly, that token consumption didn’t correlate with useful features shipped. [4] Microsoft told engineers in a major division to stop using an AI coding assistant because of the bills. One company, per Axios, ran up a $500 million AI bill in a single month after nobody thought to set a usage cap. And Nvidia’s Jensen Huang is out there telling everyone a $500,000 engineer should consume $250,000 in tokens a year, while his own VP admits the compute now costs more than the people using it. The company selling the shovels is conceding the shovels cost more than the diggers.

2. The discounts are disappearing. Roughly 95% of enterprise usage still runs on the costliest frontier models, and the sellers lose money at these prices. OpenAI reportedly spends nearly two dollars for every dollar it earns on inference. Anthropic and GitHub both moved enterprise customers to usage-based billing this year, and analysts project bills will rise another 30% to 50% as pricing normalizes. [4] It has to, because the build is no longer just software. Big Tech has announced $740 billion in capital expenditure this year, up 69%, and that money is buying steel, land, chips, power, and water. [4] Software margins were subsidizing the tokens. Brick and mortar doesn’t do subsidies. Every DIY cost model underwritten at 2025 prices sits on a discount that is going away.

3. Companies are getting more selective. MIT found that 95% of enterprise GenAI pilots delivered no measurable P&L impact. More useful, it found the variable that separates the 5%: externally partnered deployments reached production roughly twice as often as internal builds, about 67% versus 33%. [2] S&P Global found the share of companies abandoning most of their AI initiatives jumped from 17% to 42% in a year. [3] Gartner projects more than 40% of agentic AI projects will be canceled before the end of 2027, even as it expects agent-software spending to more than double this year, to $207 billion. [5] Read those two together and you get the strange picture of money accelerating into the failure data.

4. AI does not make all roles less expensive. Forbes, working from MIT’s cost research, reports that AI automation is currently economically viable in roughly 23% of roles. For the other 77%, the humans it replaced were cheaper. [4] In three jobs out of four, companies cut people to fund a machine that costs more than the people did. Read that one twice.

The DIY era, by the numbers

The rehiring nobody announced

The labor market is running the same correction, just more quietly. Forrester’s analysts report a dirty little secret: when clients announce they’re cutting 20%+ of staff for AI, nine times out of ten there is no mature, vetted AI application ready to do those jobs. Most haven’t even started building one. The layoffs were financially driven, and AI was the story told about them. [6]

Fifty-five percent of leaders now say they regret their AI-attributed cuts, and Forrester expects half of those layoffs to be quietly reversed. [7] Robert Half estimates a third of the executives who cut roles for AI have already rehired those people or recruited their replacements. [7] To a degree, this is similar to when we though omni-channel non-personal promotion could replace the impact of the salesperson.

Ford spent this year rehiring hundreds of engineers, with its VP of vehicle hardware conceding they’d undervalued the experience of their most tenured people. [7] IBM restarted entry-level hiring across every business unit, its CHRO asking out loud what happens in three to five years without them: “There’s no pipeline; the well simply dries up.” [7] I published Experience Necessary three weeks before that announcement. I have never been less pleased to be on time. What all of this says to me: the constraint was never headcount. The constraint was judgment, and judgment is what the market for expertise sells.

The wreck premium

The part the show never has to explain is what happens when the pros finally arrive. For the first few days, they don’t build anything. They undo. Often times 25% of the budget goes to this and getting the right inspections in place, so the real work can be done. The same work is paid for twice: once to do it, once to take it apart.

AI has the same arc. Call late and the pros’ first line items are assessment and undoing. The brittle agent stack wired into live workflows that nobody dares turn off. The prompt library only one departed employee ever understood. The shadow processes teams now depend on and will fight to keep. There is even a stat for the machine generating its own demolition: code churn, meaning work written and then thrown away, rose more than 800% under heavy AI adoption. [4]

So here is the crux: every make-versus-buy answer ships with an expiration date, and the date isn’t printed on the label. Make it inside only while that beats buying it. Stigler’s essay adds the clock, because the boundary on when to build or buy moves as the market matures, and 2024’s right answer becomes 2026’s wrong one without anyone making a mistake. Hand off too early and you outsource your own education; you never develop the judgment to be a good client. Hand off too late, which is the common failure right now, and you don’t just pay the DIY running costs longer. You add to the undo bill the pros will eventually hand you. The cost of waiting isn’t flat. It compounds.

Nobody’s job is to declare the phase over. Which is why it has to be yours.

The decision, in three tiers

Not everything moves at once: some jobs you always do yourself, some jobs you need to own with a little help, and some work was never a DIY job at all.

Let people paint. Individual AI use costs little and risks little, and it doubles as your best sensing instrument for where the value actually is. The shadow users are scouts. You never know where the brilliance may come from. Our marketing teams are coming up with ideas that our AI experts are then building on. But be honest about the trade: some of the time that AI saves in calculable ROI needs to repaid in the individuals trial-and-error learning. Treat it as development, and plan for it.

Own the rooms you live in. Keep building where the work is data-sensitive, high-frequency, brand-core, and large enough to keep in-house specialists sharp. Same criteria as in-housing any other organizational capability, because it’s the same test.

Call the pros for the structural work. It’s time when any of the following is true:

• The technology moves faster than your team can learn it. If what your people mastered six months ago is already out of date, and retraining takes a year, the build is decaying faster than you can maintain it.

• The specialists get more practice than you ever will. Expertise here is built on repetition, and a partner deploying AI every week across dozens of companies sees failure modes your team, deploying once a year inside one, never will. That gap is why MIT found external partners reach production twice as often as internal builds.

• The fixed cost is bigger than your worst budget year. Payroll and licenses don’t shrink when the marketing budget does, so a build that only pencils out at peak budget becomes, in a lean year, the cost you can’t cut.

• Checking the machine’s work costs more than the machine saves. When senior people spend more hours reviewing, correcting, and re-prompting AI output than the AI saved by producing it, you’ve automated the easy part of the job and multiplied the hard part.

Once a specialist layer demonstrably beats your build, and the MIT numbers say it does, the burden of proof flips from “why buy?” to “why are we still making this?” The obvious objection is that the tools keep getting easier, so maybe DIY stays viable forever. I’d take that more seriously if the failure data pointed at model quality. It doesn’t. MIT is blunt that the divide comes down to approach and integration, which are exactly the problems experience solves.

And if calling for help feels like retreat, remember what the pros are. The agency is what a market invents when a capability becomes too complex to keep sharp inside a single company. Marketing services didn’t merely survive a moment like this one. They were created by one.

What I’d ask of leaders

Re-run make-versus-buy on your AI stack, and date-stamp the answer. The 2024 call was right in 2024. That’s not the question. Ask whether the specialist market now beats the firm, capability by capability, and write the kill criteria before the review, because sunk cost, pride, and the team you hired will all vote to keep the build alive.

Put the full cost of DIY on one page, including the undo. The token bill is visible. The oversight cost is not: the senior hours spent verifying and re-prompting, the integration work that never ends, the output that gets thrown away. And then there’s the line nobody budgets at all, which is what unwinding the build will cost when the time comes. That number grows every quarter you wait.

When you bring in the pros, change the paper. No open-ended hourly cleanup, and no buying back the old production retainer. Insist on a priced demolition plan, outcome-linked fees, and named kill criteria. And make sure nobody advising the decision gets paid either way.

Pay for judgment like it’s the scarce input, because it is. Production is a commodity now. Cross-client pattern recognition is not. Force the new relationship into the old rate card and you will select for exactly the partner you don’t want, the one still selling hours and tonnage.

Keep the control tower; rent the construction crew; never collect the paintbrushes. Someone inside still has to own the data, the standards, and the evaluation of the partners themselves. Ford and IBM just spent a year relearning that the missing ingredient was experience, not tooling. And individual AI use stays no matter what else moves, because it’s how you’ll spot the next flip before the dashboard does.

Where this leaves us

None of this is an argument against picking up the tools. Every build decision was a bet on a moment, and the moment moves. In 2023 you swung the hammer and did the sawing yourself because nobody worth calling existed yet. That era is over.

Every episode of Help! I Wrecked My House ends the same way: the house is beautiful, the family is relieved, and nobody says out loud what the wreck added to the bill. Business doesn’t get that edit. The number lands in the SG&A line, and we have already seen it. The homeowners’ mistake was never picking up the tools. Their mistake was not recognizing the moment to put them down. For a lot of organizations that moment is now, and unlike the families on the show, you still get to make the call before the ceiling comes in.

A note on how this was written

The argument is mine: that the in-house AI build-out was the right call for an immature market, and Stigler’s economics explain both the build and the flip; that the SG&A, token-cost, and rehiring data mark the flip arriving; that the leadership skill now is timing the handoff to outside expertise, because a late call compounds — the pros must undo before they can build; and that how you pay is the mechanism that selects the partner you deserve. I worked through it in conversation with Claude, Anthropic’s AI, using it to verify the figures against the sources cited below and to help structure and edit the prose. The judgments — and the television habit — are my own. There’s an obvious irony in drafting an essay about the limits of doing it yourself with a machine’s assistance; I’d rather name it than pretend it away.

Notes

1. The Hackett Group, “US SG&A Cost Study and Scorecard,” news release, December 9, 2025. Analysis of the publicly available financial statements of the 1,000 largest US-headquartered public companies, fiscal 2024: median SG&A rose from 13.7% to 14.3% of revenue, a five-year high; 62% of companies saw SG&A rise as a share of revenue; 78% failed to keep cost growth below 2.9% inflation. Hackett attributes the rise largely to slowing revenue growth. One disclosure worth making, in the spirit of cui bono: Hackett describes itself in the same release as “a leading Gen AI consultancy” — the firm benchmarking the cost problem also sells the AI cure. The underlying data come from public financial statements; the framing is theirs.

2. MIT Project NANDA, “The GenAI Divide: State of AI in Business 2025.” Based on 52 executive interviews, surveys of 153 leaders, and analysis of 300 public AI deployments: 95% of enterprise GenAI pilots delivered no measurable P&L impact; externally partnered deployments reached production roughly 67% of the time versus roughly 33% for internal builds; more than half of GenAI budgets went to sales and marketing pilots while measurable ROI concentrated in back-office automation.

3. S&P Global Market Intelligence, “Voice of the Enterprise: AI & Machine Learning,” 2025 survey of more than 1,000 enterprises across North America and Europe (reported March 2025): 42% of companies abandoned most of their AI initiatives, up from 17% a year earlier; on average, 46% of proofs-of-concept were scrapped before reaching production.

4. Jemma Green, “AI Costs More Than The People It Replaced,” Forbes, July 2, 2026. Source for: Uber’s exhausted 2026 AI coding budget, 84% engineer adoption, ~70% AI-originated code, and the COO’s concession that token usage didn’t correlate with features shipped; Microsoft’s directive to stop using an AI coding assistant; the $500 million single-month AI bill (via Axios); Jensen Huang’s token-consumption prescription and Nvidia VP Bryan Catanzaro’s acknowledgment that compute now exceeds payroll for his team; Amazon’s retired usage leaderboard; the ~95% of enterprise usage on the costliest frontier models; OpenAI spending nearly $2 per $1 earned on inference; Anthropic’s and GitHub’s 2026 moves to usage-based billing and analyst projections of 30–50% higher enterprise AI bills at normalized prices; Big Tech’s $740 billion in announced 2026 capital expenditure, up 69%; the Faros AI finding of code churn rising more than 800% under high AI adoption; and the finding, attributed by the column to MIT research, that AI automation is currently economically viable in roughly 23% of roles, with humans cheaper in the remainder. On “tokenmaxxing” more broadly: The Economist, “Companies are scrambling to curtail soaring AI costs,” June 14, 2026. On data centers and water, see my earlier piece “You Can’t Drink Data” (July 21, 2026).

5. Gartner, June 2025: more than 40% of agentic AI projects projected to be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls. Gartner separately projects AI agent software spending of $207 billion in 2026, up 139% year over year (via Forbes, note 4).

6. J.P. Gownder, “Forecast: AI And Automation Will Take 6% Of US Jobs By 2030,” Forrester, January 13, 2026: 6.1% of US jobs (10.4 million) displaced by AI and automation by 2030; 20% of jobs strongly influenced by generative AI — 3.25x the replacement rate; and the observation that nine out of ten clients announcing AI-replacement layoffs had no mature AI application ready to fill the roles.

7. Bruce Crumley, “55 Percent of Leaders Regret AI Layoffs—and a Major Hiring Reversal Has Begun,” Inc., July 28, 2026: the Forrester survey finding 55% of leaders regret tech-driven staff cuts and the prediction that half of AI-attributed layoffs will be quietly reversed; the Robert Half estimate that a third of executives who eliminated roles for AI have rehired or re-recruited; the resumed hiring at Ford and IBM; and the quoted remarks of Ford VP Charles Poon and IBM CHRO Nickle LaMoreaux.

8. George J. Stigler, “The Division of Labor Is Limited by the Extent of the Market,” Journal of Political Economy, 1951, building on Adam Smith, An Inquiry into the Nature and Causes of the Wealth of Nations, 1776; Ronald H. Coase, “The Nature of the Firm,” Economica, 1937. Coase and Smith are load-bearing in my earlier piece In-Housing: What to Own, What to Rent; Stigler supplies the dynamic version — the boundary of the firm moves as the market matures.

9. On the semiconductor flip: TSMC was founded in 1987 as a pure-play foundry; AMD divested its fabs (into GlobalFoundries) in 2009; Nvidia has been fabless since its 1993 founding; Intel has in recent years outsourced portions of its own chip manufacturing to TSMC. “Real men have fabs” is attributed to AMD founder Jerry Sanders, from the 1990s.

The same region is also one of the more desirable places to build a data center. Land is cheap, labor readily available, and high-speed fiber runs between California and Texas. So the data centers are coming, and they are thirsty. In metro Phoenix, cooling water for data centers is projected to rise 870% — from 385 million gallons a year to 3.7 billion — as announced facilities come online.³ Across major operators, 97% of the water they consume is drawn straight from municipal drinking-water systems.⁴

The industry has a phrase ready for this: “water positive.” Microsoft, Google, Amazon, and Meta have all pledged to be water positive by 2030: to return as much water to their watersheds as they take out, or more.⁵ It is a reassuring phrase, and it is just plain misleading.

What “water positive” means

When a company says water positive, it usually means one of three things.

Avoided usage. More efficient cooling (or conservation somewhere else) less consumption: use X, save Y, X minus Y is the “positive” number. This is accounting, not hydrology. A data center that cools more efficiently is still consuming water that must come from somewhere.

Returning water. Some plants discharge cooling water back into rivers or aquifers and count it against what they drew. But water isn’t fungible across time, temperature, and place. Water pulled from the Colorado during a summer drawdown and returned after usage is not the same water that was taken.

Funding infrastructure. Watershed restoration, rainwater capture, wastewater recycling. In the most generous reading this is offsetting — and offsetting is a familiar shell game. It is the logic that lets a company fund a solar farm in one state, open a gas plant in another, and call the result net zero. It works only for those with the capital to buy absolution.

In the end, none of the three makes water. None resolves the underlying truth: we are putting water-hungry infrastructure into the places with the least water to spare, and drawing it, mostly, from the taps.

The pattern we keep repeating

We have watched innovation outrun regulation and public understanding before. A profitable opportunity appears; early evidence of harm appears with it; the people making money fund friendly science, lobby while the evidence is still “uncertain,” and keep going — through the scarred lungs, the broken metabolisms, the altered atmosphere. Profit outruns precaution, and the damage arrives before the correction does.

• Tobacco. The link to cancer was clear by the early 1950s. The industry’s answer was to manufacture uncertainty; an internal memo put it plainly — doubt was the product.⁷ The delay was counted in decades, and in lives.

• Sugar and processed food. In the 1960s the sugar industry quietly funded research steering blame toward fat and away from sugar, shaping decades of nutrition advice while metabolic disease climbed.⁸ The doubt was bought, not discovered.

• Fossil fuels. Oil companies ran credible internal climate models in the 1970s and knew what burning carbon would do. They funded denial, lobbied, delayed.⁹ We are paying the interest now, in fire seasons and storms and ice that will not come back.

• Social media. The platforms’ own researchers found the products were harming users — teenage girls in particular — and the companies optimized for engagement anyway.¹⁰ The internal documents are not in dispute.

I’ll grant that AI is not tobacco. It carries legitimate upside cigarettes never had, and water — unlike atmospheric carbon — is local and, when handled well, largely recoverable. That should make me more hopeful, not less: this is one of the rare cases where the harm is still preventable, and the tool causing it is also the tool that could undo it.

The paradox

Artificial intelligence could help solve water scarcity. It already does versions of the work: forecasting drought, mapping groundwater, and modeling the climate itself.

The trouble is where the marginal dollar and the marginal gallon actually land: at sharper ad targeting, stickier recommendation loops, synthetic media, and faster trading. I work in marketing. I know what a better banner ad can be worth, and it is not an aquifer. So we have built a tool powerful enough to watch a watershed in real time, and we are mostly using it to sell people things they didn’t know they wanted — while draining the aquifer beneath the transaction.

The capability isn’t the constraint. The choice is.

The choice

So here is the argument. When a private technology runs on a public resource — and drinking water in a drought is about as public as a resource gets — the people who depend on that resource belong inside the decision. Not consulted after the fact. Inside it.

The market built these tools; public provision has its own record of waste and capture; “make AI a public utility” is a slogan that dies on contact with how anything gets funded. All fair — I am not arguing for nationalizing or romanticizing public ownership.

My claim is narrower. Elinor Ostrom won a Nobel for showing that shared resources — fisheries, forests, groundwater basins — survive when the people who depend on them have real standing in how they are governed: monitoring, enforceable rules, and a seat at the table.¹¹ A data center pulling potable water from a stressed basin is extracting from precisely that kind of commons — and right now the community’s “standing” is a public-comment period and a press release with water positive on it.

So the go-forward is neither a ban nor a takeover. Call it water honest instead of water positive: availability is a hard constraint, not an afterthought; binding, audited accounting instead of self-reported offsets; real governance standing for the towns whose taps are on the other end of the pipe; and a public claim on the upside when a facility runs on a public resource. That is a redirection, not a prohibition.

Where this leaves us

None of this is a case against AI or building data centers in the desert. The tools are extraordinary. The case is narrower. We are pointing the most capable problem-solving instrument ever built at the problems that turn a profit, running it on the one input none of us can substitute, in the one place with the least of it left — and calling the arrangement “positive.”

Water is the right place to draw the line. You can live without AI optimized marketing. You cannot live without water. The genie is not going back in the bottle, and I would not want it to. But what we point it at is still ours to decide, and the window to decide is closing while the reservoirs fall.

The question I keep returning to in everything I’ve written about technology is: what is it for? Here, for once, the answer ought to be easy. Before we drain another aquifer proving we can generate marketing experience, it is worth holding onto the one thing the dashboards can’t measure and the ledgers can’t offset: you can’t drink data.

About the article’s title
The term “You can’t drink data” came from the National Congress of American Indians conference last fall, as activists interrupted an AI panel by chanting "You can't drink data!". Reported by David W. Chen, New York Times, July 9, 2026: “Big Tech Is Now Targeting Native American Land for Massive Data Centers”.

A note on how this was written

The argument here is mine: that “water positive” is accounting, not hydrology; that AI’s water draw follows the profit-over-precaution pattern of tobacco, sugar, and carbon while remaining, unusually, preventable; and that a technology running on a public common owes the affected community real governance standing. Rather ironically, I worked through it with Claude, Anthropic’s AI, using it to connect the case to the economics and history and to verify the water data and sources. The judgments are my own; the factual claims are checked against the reporting cited. Given the subject, it would be strange not to say that it was used to assist me.

Notes

1. A. Park Williams et al., “Rapid intensification of the emerging southwestern North American megadrought in 2020–2021,” Nature Climate Change, 2022 — the region’s driest 22-year stretch in at least 1,200 years.

2. U.S. Bureau of Reclamation, Colorado River Basin shortage determinations, 2021 onward; ~40 million is Reclamation’s standard population estimate for the basin.

3. Ceres, “Drained by Data: The Cumulative Impact of Data Centers on Regional Water Stress,” September 2025 — Phoenix-area direct cooling water projected to rise from ~385 million to ~3.7 billion gallons a year as announced facilities come online.

4. Vanessa Privette et al., “Data Centers’ Water Footprint: The Need for More Transparency,” AGU Advances, 2026, citing Bluefield Research, 2025 — over 97% of water consumed by major data-center operators drawn from municipal drinking-water systems.

5. “Water positive by 2030” commitments: Microsoft (2020), Google (120% replenishment), Amazon Web Services (2022), and Meta — see each company’s sustainability reporting.

6. The scale counterpoint — worst-case data-center draw is roughly 1% of Phoenix residential use, agriculture the dominant user — per the same Ceres analysis and Grist reporting, 2026.

7. Brown & Williamson internal memo, 1969 (”Doubt is our product…”); the industry’s 1953 “Frank Statement to Cigarette Smokers.”

8. Cristin Kearns, Laura Schmidt, and Stanton Glantz, “Sugar Industry and Coronary Heart Disease Research,” JAMA Internal Medicine, 2016 — on industry funding that shifted blame from sugar to fat.

9. InsideClimate News, “Exxon: The Road Not Taken,” 2015; Supran and Oreskes, “Assessing ExxonMobil’s climate change communications (1977–2014),” Environmental Research Letters, 2017.

10. The Wall Street Journal, “The Facebook Files,” 2021, and Frances Haugen’s disclosures on Instagram’s effects on teenage users.

11. Elinor Ostrom, Governing the Commons: The Evolution of Institutions for Collective Action, Cambridge University Press, 1990.

Photo: The Apple Data Center in Mesa, Ariz., in 2017. Jim Todd / Reuters file (NBC News)

The automation was so good, for so long, that it eroded the very capability it existed to support.¹ The better the machine got at the routine, the less prepared its humans were for the rare moment the machine couldn’t handle — which was the only reason the human was still there for.

There is a name for this, and it predates the crash by a quarter century. In 1983 the researcher Lisanne Bainbridge published a short paper called “Ironies of Automation.”² When you automate the easy parts of a job and leave a person to cover the hard parts, you’ve done two things at once: 1) handed the human the most demanding work — the exceptions, the emergencies — and 2) removed the daily practice that kept them sharp enough to do it. The job ends up depending most on human judgment at exactly the moments it has done the most to degrade it.

Fast forward to today, and to agentic AI in marketing — and in consulting, law, and every other white-collar forums. We’re about to run Bainbridge’s experiment across a whole generation of office workers, and almost nobody is talking about the irony of it.

The same trap, wearing a suit

Today, agentic AI is genuinely good at the work that used to belong to the most junior people on the team. Pulling the campaign data together. Draft the first version of the copy. Cleaning the spreadsheet. Building the deck. Chasing down the numbers. Nobody loved this work. The case for handing it to a machine is real, and anyone holding a budget or a stopwatch can tell you why it makes sense to do this.

But that work was doing a second job that never appeared on any invoice. It was how a twenty-three-year-old with a good degree and no judgment evolved into a thirty-five-year-old you’d trust with a client, a crisis, or a decision the data couldn’t make. The grunt work was learning by repetition, and the repetition was a big part of how a person became worth promoting. We’re now removing the repetition and keeping the expectation that experienced people will keep appearing, as if they arrive fully formed from somewhere else.

Chip Conley put the distinction even better. In Wisdom @ Work, taking on Peter Drucker’s idea of the “knowledge worker,” he argues that we’re drowning in knowledge and starving for wisdom — and that wisdom, unlike knowledge, can’t be automated.³ When the founders of Airbnb brought him in, they told him they’d hired him for his knowledge, but what they got was his wisdom. Take a pause there and think for a moment about new hires you’ve brought in for their experience in other industries — what did their real value end up being? Knowledge is the thing a machine now hands you instantly and for free. Wisdom is knowledge that has been paid for — in repetition, in failure, in the slow accumulation of having seen this before. And the price is paid in precisely the entry-level work we’re automating away.

So the proposition of this essay is simple to state and hard to shake. The work we’re handing to the machine is the work that makes people experts — so the better the machine gets, the faster we starve ourselves of the judgment we will need to oversee it.

The apprenticeship was hidden in the grunt work

The junior-level job was a cross-subsidy. When a firm paid a entry-level or junior-level analyst a modest salary to do modest work, it bought two things in one transaction and only ever noticed one of them. It noticed the output — the cleaned data, the reviewed contract, the finished slides — and it paid for that. Riding along inside the same transaction, unpriced and unremarked, was the second product: a person becoming an expert. The firm got its work done. The economy got its next generation of senior people. The training was a byproduct of the work, so nobody had to fund it deliberately, and nobody did.

One transaction, two products — and only one of them was ever on the invoice.

AI severs the two. It does the output better and at a fraction of the cost, so the rational firm stops buying the junior employee’s output — and the training bundled inside that output dies as collateral. Not because anyone decided the next generation didn’t matter. Because the thing that manufactured the next generation was never a line item anyone chose to protect. It was a side effect, and you don’t defend a side effect in a budget meeting.

Four ideas that say this is a real problem

The natural hope is that this is fine — that we’ll simply start people higher up, at the judgment layer, and skip the boring years. Four of the most durable ideas in how humans acquire and hold skill say it isn’t that easy. If you ever had to make this case to a room, these are the four principles to explain.

1. The knowledge that matters can’t be downloaded (Polanyi).

In 1966 Michael Polanyi compressed a lifetime of thinking into one sentence: we know more than we can tell.⁴ The deepest expertise is tacit — it lives in the expert’s instincts and can’t be fully written down, because the expert themselves can’t fully articulate it. The explicit, codified kind of knowledge is exactly what an AI holds in abundance and hands to anyone. The tacit kind — the read on a room, the sense that a deal is going sideways, the feel for which number in the model is lying — transfers only by doing the work under the eye of someone who already has it. That transfer was the apprenticeship. Remove the doing, and the tacit knowledge has no way to move.

2. There’s no express elevator to expertise (Dreyfus).

Hubert and Stuart Dreyfus spent the 1980s mapping how a person travels from novice to expert, and found five stages — novice, advanced beginner, competent, proficient, expert — that run in order and can’t be leapt.⁵ More pointedly, they argued that genuine expertise doesn’t work by following rules faster; the expert has stopped consulting rules and simply sees what the situation requires. That intuitive seeing is the summit, and it’s built only by climbing through the earlier stages one repetition at a time. The machine can follow rules better than any novice; it can’t hand anyone the seeing, and it can’t make the stages go faster.

3. Automation erodes the judgment it depends on (Bainbridge).

This is the pivot, and it’s why the problem runs deeper than a hiring pipeline. It would be bad enough if the only issue were that no new experts are coming through. Bainbridge’s irony says the damage runs in both directions at once. Automate the routine work and you also decay the experts you already have. Senior judgment isn’t a trophy earned once and kept forever; it’s a muscle, kept in tune by contact with the actual work. Hand all of that contact to the machine — the analysis, the drafting, the first pass on everything — and the senior’s own fluency withers, exactly as the pilots’ manual-flying did behind the autopilot. And here’s the cruel timing: the whole justification for keeping the human in the loop is that they’ll catch what the machine gets wrong. Catching it requires the sharpness that only comes from doing the work the machine has taken. Dependence on judgment goes up as the supply of judgment goes down. That is not a bug in the deployment. That is the shape of the trap.

Bainbridge’s loop: the pipeline empties and the existing experts rust — at the same time, for the same reason.

4. The market won’t fix this on its own (Becker).

The comforting reply is that if firms genuinely need experienced people, they’ll invest in producing them. Gary Becker explained sixty years ago why that fails for exactly this kind of training.⁶ His Human Capital drew a line between two kinds of skill. Firm-specific skill — how this company does things — a firm will happily fund, because it captures the return. General skill — transferable judgment a person could take to any competitor — a firm is loath to fund, because the moment the worker is valuable they can leave, and a rival that spent nothing on training simply hires them away. Firms therefore systematically underinvest in portable expertise. The only reason it got produced at all was the cross-subsidy: the training came free, bundled inside junior-level labor the firm wanted anyway. Break the bundle and you’re left with a form of human capital every firm needs, no firm will pay to create, and every firm would rather poach than build.

That’s a textbook collective-action failure, and it resolves the way they always do. Each firm’s saving is immediate, legible, and lands in this year’s numbers. The cost — a hollowed-out generation of experienced people — lands a decade out, is diffuse, and belongs to no one. No executive is ever fired for it. Everyone poaches; no one trains; the shared pool everyone drinks from slowly empties, because nobody’s job was to refill it.

The one place that solved this didn’t do it through the market. The German and Swiss dual apprenticeship systems produce transferable expertise at scale precisely because the cost is shared — firms, industry chambers, and the state co-fund the training, so no company carries a burden a rival can free-ride on.⁷ White-collar work never built anything like that, because it never had to; the junior-level job did it for free. It’s worth noticing that the blue-collar trades kept their apprenticeship ladder — formal, funded, and deliberate. An electrician still becomes an electrician the slow way. It may turn out the knowledge workers are the ones who let their apprenticeship rot, precisely because theirs was invisible and unpaid and easy to lose without noticing.

What the numbers actually say

First, the good news: the apocalyptic framing — the warning that AI could erase half of all entry-level white-collar jobs within a few years — hasn’t shown up in the aggregate data.⁸ White-collar employment has grown. There are more software developers than in 2022, more radiologists, and, tellingly, more paralegals — the professions everyone nominated for extinction.⁹ If you’re looking for a collapse, it isn’t in the totals.

What’s happening is subtler, and worse in a way that trips no alarms: the damage is landing specifically on the on-ramp. A Harvard working paper analyzing résumé and job-posting data across tens of millions of workers found that at firms actively adopting AI, junior-level hiring has fallen sharply since 2023 — entry-level employment down roughly nine percent within six quarters of adoption, relative to firms that didn’t adopt — while senior headcount at the same firms kept growing.¹⁰ A Stanford analysis of payroll records found a roughly sixteen percent relative decline in early-career employment in the most AI-exposed occupations since late 2022, concentrated in the fields — software, customer service — that used to absorb armies of junior-level hires.¹¹ Economists at Indeed have a name for the milder version: “experience creep,” employers demanding more experience for the jobs that used to provide it.¹² The perverse endpoint is a market where landing your first job requires proof that you’ve already had one.

The on-ramp, by the numbers

And leaders are choosing this on purpose — the survey numbers above aren’t a side effect; they’re policy.¹³ You can watch it most clearly where white-collar apprenticeship was most formalized: consulting. The firms’ own AI tools now do most of the junior-analyst research and slide-building that used to be a first-year’s entire existence, and some firms have hired former consultants specifically to train the machine on entry-level work.¹⁴ The pyramid is being sawn off at the base by the people who climbed it. Where the future partners come from is a question the deck doesn’t answer.

There’s a case that I could be wrong

Three good logical objections to my point-of-view do deserve exploration.

1. This is an old panic in new dress. Socrates thought writing would destroy memory. Every generation is sure the new tool will produce a soft generation, and every generation turns out mostly fine — it just develops skills its parents didn’t value. Maybe “wisdom” gets redefined, and what we’re mourning is one flavor of it.

2. Maybe the ladder doesn’t break; it compresses. This is the strongest objection, and the one I’d bet on shaping the good outcomes. Maybe the new junior-level colleague starts at the judgment layer on day one, with AI doing the floor, and — supervised well — accumulates the repetitions that matter faster than we did, because the machine has cleared the drudgery from the calendar. The flight simulator is the encouraging precedent: it didn’t deskill pilots, it became the fastest way ever invented to build the skill, letting a trainee live through a hundred emergencies that reality serves up once a career. AI could be that — a sparring partner that manufactures repetition on demand — rather than the autopilot that quietly removes it. Which of the two it becomes isn’t a property of the technology. It’s a choice about deployment.

3. A lot of the grunt work was just grunt work — it built tolerance for tedium, not judgment, and good riddance to it. True. But that’s a reason to be more deliberate, not less: if the accidental training is gone and only some of it mattered, keeping the part that mattered now has to be done on purpose, by someone, or it won’t happen at all.

Even if all three are true, the problem doesn’t go away. The industrial machine that used to manufacture judgment is being switched off on cost grounds, and nobody is deciding what replaces it. That’s the worry — not that we chose wrong, but that we didn’t choose.

What I’d ask of leaders

1. Treat apprenticeship as a system you design, not a byproduct you assume. For a century the junior-level job trained the person for free. It won’t anymore. If you want experienced people in ten years, someone has to build the thing that makes them — the rotations, the exposure to real judgment calls, the repetition against hard problems — on purpose, and fund it as what it is: an investment, not an overhead line to trim.

2. Move your junior people up the ladder, not out of the building. The firms that will look smart in a decade aren’t the ones cutting early-career hiring hardest; they’re the ones putting young people on real judgment work sooner, with AI doing the floor beneath them. Several large employers have already reversed course and gone back to hiring new graduates aggressively, reasoning that a company that stops taking in young talent is, in one firm’s own phrase, starving itself of its future.¹⁵

3. Measure the pipeline, not just the savings. The money you save by cutting a junior-level program shows up this quarter; the capability gap it opens is invisible and years away, so it never makes the dashboard. Put it on the dashboard. If the only things you can point to are dollars saved and assets produced, you haven’t built a leaner organization. You’ve quietly borrowed against your own future and called it efficiency.

4. Give the talent pipeline a real owner with real authority. A training system that everyone is responsible for is one that no one defends when budget pressure arrives. Name the person. Give them the standing to protect early-career development against the reasonable, relentless pressure to cut it. Without that authority, every good intention here is a suggestion.

5. Treat this as bigger than your own firm. Becker’s logic is brutal and correct: no single company can afford to be the one that trains the whole industry’s talent for its competitors to poach. The real fix is partly collective — the unglamorous machinery of shared standards, funded apprenticeships, and industry-level commitment that the trades never abandoned and the professions never built. That isn’t a memo you can write alone. But it starts with leaders willing to say the quiet part: the market, left alone, won’t produce the experienced people the market is about to need.

Where this leaves us

The pilots of that Airbus weren’t stupid, and they weren’t lazy. They were failed by a system that had let a skill atrophy and then demanded it back in an instant, at the one moment it was gone. That’s the shape of the thing I’m afraid of, scaled up from a cockpit to a generation. Not a dramatic collapse — the totals will look fine for years, the dashboards will stay green — but a slow, invisible thinning of the one capability everything else now rests on, discovered only when we reach for the experienced colleague and find we stopped making them somewhere back down the line.

The repetition is still available. The young people are still willing. Knowledge is cheaper and more abundant than at any point in human history. The only thing genuinely at risk is the wisdom — the part you can’t download, that has to be earned one step at a time. It’s at risk not because AI is too powerful, but because we’re too content to let AI do the very work that used to make us worth having. The ladder is still standing. Someone has to decide, on purpose and against the arithmetic, to keep sending people up it.

A note on how this was written

The argument here is mine: that entry-level work has a hidden cross-subsidy that produces the next generation of experienced people as a free byproduct; that automating it severs the training from the work and leaves no means to replace it; that Bainbridge’s irony guarantees the damage runs in both directions, hollowing out the pipeline and decaying the experts we already have; and that Becker’s economics make this a structural certainty rather than a choice any single firm can reverse. I worked through it in conversation with Claude, Anthropic’s AI, using it to connect the case to the underlying theory and to find and check the industry data and sources. The judgments are my own, and the factual claims are checked against the reporting cited below. The knowledge came from the machine, easily and for free. Whether any of this adds up to wisdom is the one thing the machine couldn’t tell me, and the one thing worth arguing about ;)

Notes

1. Air France Flight 447, 1 June 2009; BEA (Bureau d’Enquêtes et d’Analyses) final report, 2012. The accident was multi-causal; the automation-dependency and skill-erosion thread is one prominent strand of the analysis, not the sole cause. As a business case: Nick Oliver, Thomas Calvard, and Kristina Potočnik, “Cognition, Technology, and Organizational Limits: Lessons from the Air France 447 Disaster,” Organization Science, 2017, summarized in Harvard Business Review, September 2017 (“The Tragic Crash of Flight AF447 Shows the Unlikely but Catastrophic Consequences of Automation”); and Jamie O’Brien, “Mystery over the Atlantic: The Tragic Fate of Air France Flight 447,” The CASE Journal, Vol. 15, No. 1, 2019.

2. Lisanne Bainbridge, “Ironies of Automation,” Automatica, Vol. 19, No. 6, 1983. The foundational statement that automating routine tasks both burdens operators with the hardest work and strips them of the practice needed to do it.

3. Chip Conley, Wisdom @ Work: The Making of a Modern Elder, Currency, 2018. On the distinction between knowledge and wisdom and the claim that wisdom cannot be automated; builds on Peter Drucker’s coinage of the “knowledge worker” in Landmarks of Tomorrow, 1959. The “hired me for my knowledge … got my wisdom” line is Conley’s account of the Airbnb founders.

4. Michael Polanyi, The Tacit Dimension, 1966 (and Personal Knowledge, 1958). “We know more than we can tell” — the argument that the deepest expertise is tacit and cannot be fully codified.

5. Hubert L. Dreyfus and Stuart E. Dreyfus, Mind Over Machine, Free Press, 1986, building on their five-stage model of skill acquisition (1980). On the developmental, unskippable progression from novice to expert and the non-rule-based character of expert intuition.

6. Gary S. Becker, Human Capital, Columbia University Press for the NBER, 1964. On the distinction between general and firm-specific human capital and why firms underinvest in transferable skill.

7. On the German and Swiss dual apprenticeship systems: the vocational-training model in which firms, industry chambers, and the state co-fund transferable skill, structurally solving the free-rider problem Becker’s analysis predicts.

8. The “half of entry-level white-collar jobs” warning is Anthropic CEO Dario Amodei, first widely reported in May 2025 and subsequently softened by Amodei and other AI-company leaders.

9. On aggregate white-collar growth (roughly three million white-collar jobs added since 2022; 7% more software developers, 10% more radiologists, 21% more paralegals): Washington Monthly, May 2026.

10. Seyed Hosseini and Guy Lichtinger (Harvard), working paper, 2026, analyzing résumé and job-posting data covering 66 million workers across more than 280,000 U.S. firms, 2015–2025: at AI-adopting firms, entry-level employment fell roughly 9% within six quarters of adoption relative to non-adopters, while senior employment continued to grow; junior-level hiring declined sharply from 2023. Reported in Forbes, May 2026.

11. Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen, “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence,” Stanford Digital Economy Lab, November 13, 2025. Analysis of ADP payroll records: a ~16% relative decline in employment among early-career workers in the most AI-exposed occupations since late 2022, concentrated in software development and customer service.

12. Laura Ullrich, lead economist at Indeed, on “experience creep,” via Washington Monthly, May 2026.

13. British Standards Institution survey of 850 business leaders across seven countries, via Fast Company, May 2026: 39% had already reduced or cut entry-level roles due to AI; 43% expected to in 2026.

14. On consulting: futureofconsulting.ai (January 2026) on in-house tools (e.g., McKinsey’s Lilli, BCG’s Deckster) performing the bulk of junior-analyst research and slide work; TheStreet (May 2026) on roughly 150 former consultants hired to train AI on entry-level tasks.

15. On employers reversing course (IBM, Reddit, Dropbox, Cloudflare, LinkedIn expanding early-career hiring; PwC recommitting and warning that cutting it risks “starving your organization of its future”): Fast Company, May 2026.

For about forty years, the safest bet in corporate technology was that the people at the top would be the last to understand it. Employees smuggled the first PCs past an IT department that still thought real computing happened in a climate-controlled room down the hall. The internet arrived the same way, from below and outside, while the executive floor waited to be convinced. The smartphone is the cleanest case: it walked in through the front door in people’s pockets, against the explicit wishes of the departments meant to control it.

So there’s something odd about the present moment. With AI, for the first time in living memory, the boss got there first. The CEO can’t stop talking about it; the board has it at the top of the agenda; the mandate is coming down the building, not seeping up it. After four decades of leadership-as-laggard, the corner office has decided to lead from the front.

That’s the good news. The rest of this piece is about why it isn’t all good news — and why the way the C-suite chose to lead may be the costliest part of it.

The pattern that always held

The internet was a grassroots phenomenon before it was a boardroom one, built by startups while incumbents watched. The skepticism reached the very top: Bill Gates underrated the web in the first edition of The Road Ahead in 1995 before reversing hard, and the era’s most-quoted artifact is a 1995 Newsweek essay explaining, with great confidence, why the internet would “fail”.¹ The people paid to see around corners didn’t.

The smartphone gave the dynamic a name: the consumerization of IT. The best technology a person touched was now the one in their own pocket, and IT departments that dismissed early iPhones as insecure toys lost, because the demand came from everywhere at once — an emergent, bottom-up process driven by employees and treated as an afterthought.² Businesses had spent decades driving consumer technology. Now it was driving them.

Everett Rogers named the shape of this in 1962 — innovators, early adopters, the majorities, laggards — and on every wave since, the C-suite behaved like the late majority.³ Clayton Christensen explained why in 1997, and it isn’t stupidity: disruption enters at the edges, where it doesn’t yet serve your best customers, so ignoring it is the locally rational move right up until it isn’t.⁴ The boss missed the last several waves because of where those waves first appeared. Hold onto that.

Rogers’s curve: the C-suite lagged every prior wave — and, on AI, jumped to the front.

What’s different this time

The data on AI is not subtle: this is a top-down wave by any measure. BCG’s 2026 AI Radar finds nearly three-quarters of CEOs are their company’s chief decision-maker on AI, with the value-generators personally spending eight-plus hours a week on it.⁵ The World Economic Forum calls AI “the CEO’s mandate,” and half of surveyed CEOs think their job depends on getting it right.⁶ The Conference Board found AI named the single largest 2026 investment priority, ahead of product innovation.⁷ The people at the top are driving.

Where the numbers point

And yet the bottom-up wave happened too. Employees brought AI to work on personal accounts ahead of any policy — shadow AI — with three-quarters of knowledge workers already using it, most bringing their own tools while their leaders had no plan;⁸ by multiple 2026 estimates the majority of workplace ChatGPT use still runs through unsanctioned accounts.⁹ So the difference isn’t that AI skipped the bottom-up stage. It’s that, for the first time, the top didn’t sit behind it. The boss didn’t get dragged along, the way he was with the smartphone. He grabbed the wheel. But why?

Why they didn’t wait

Not because executives got wise. Three structural things changed.

First, AI changed where the disruption lands. The internet entered as a channel, the smartphone as a device — both filed under IT and delegated down. AI won’t file that way; it’s sold, by boosters and doomsayers alike, as a rewrite of the operating model, the cost base, the headcount. The last waves were additive; AI is framed as substitutive — exactly the altitude the corner office has always watched. It stopped entering from below, so leadership stopped missing it.

Second, the markets grade this in real time. CEOs answer for their AI narrative on every earnings call, to analysts who treat AI maturity as proof that management is awake. The survival instinct in the WEF data isn’t about the technology; it’s about what happens to a CEO who looks slow in front of the people who can remove him.

Third, and most human: memory of the last time. Everyone now positioned to lead this wave watched the Kodaks and Blockbusters die for arriving late. The corner office is early partly because it was late last time and got burned — so much of the leadership we’re applauding is overcorrection, not foresight. And fear and desire produce very different strategies.

The doctrine they reached for

There’s a fourth reason, and it’s fashion, not structure. The pressures above explain why leaders felt they had to act. They don’t explain why so many took the identical posture: seize control, decide from the top, move without waiting for the building to agree.

That posture has a name now. Paul Graham’s September 2024 essay “Founder Mode,” built on a talk by Airbnb’s Brian Chesky, spread through boardrooms faster than anything written about management in a decade.¹² The argument: “hire good people and give them room” is a trap, and great leaders stay hands-on and skip levels the way Steve Jobs did. Graham was openly contemptuous of the management layer it routes around, casting seasoned executives as people skilled mainly at lying upward. The heroes are the founder-kings — Jobs, Musk, Huang — who held the wheel and never let go.

There’s a real version, and I’ve relied on it: I praised Chesky in an earlier piece precisely because he had the product knowledge and the authority to change Airbnb’s model and make it stick. Done by its namesakes, founder mode isn’t about command. It’s about a leader who knows the product so deeply that proximity is an asset. The command is downstream of the knowledge. What spread, though, wasn’t the knowledge. It was the command — a generation of non-founder CEOs keeping Jobs’s willingness to overrule the room and skipping the decades he spent earning the right to.

And it arrived inside a wider current that makes command feel like courage. The leadership style ascendant this decade, in politics and in Silicon Valley’s founder-worship alike, is the same silhouette: the strongman. The lone decider who trusts his gut over the experts, treats institutions and process as obstacles to bulldoze, reads consultation as weakness and dissent as disloyalty, and staffs the room with people who already agree. It flatters whoever sits at the top of an org chart — it’s fast, it’s legible, it’s heroic — and it has acquired its corporate scripture: the founder-kings entrenched behind dual-class shares that make them answerable to no one, the essays that wave off the professional managers as liars, the celebrated executive who boasts about overruling his own engineers. Strip away the rare genuine product genius and what actually spreads is the pose: concentrate the power, distrust the building, decide alone.

Here is the trouble with running a company like a strongman, and it’s the trouble strongmen have always had. The model’s signature move — sealing yourself off from the people who might tell you you’re wrong — is the precise move that guarantees you learn it too late. The yes-men optimize for access, the dissenters go quiet or leave, and the leader mistakes the silence for agreement. It feels like strength and functions like blindness. Strongmen don’t fall because they’re stupid. They fall because, by design, they’re the last to know.

Friedrich Hayek explained the mechanism eighty years ago, and not about politics. In 1945 he asked why centrally planned economies fail even when the planners are brilliant, and the answer was knowledge: the information a complex system needs — what’s actually happening, in this market, with this customer, now — lives in no single place. It’s dispersed, fragmentary, often tacit, held by the many people closest to the facts, and no central authority can gather it fast enough to beat the distributed system that already holds it. The planner’s failure isn’t a failure of intelligence; it’s a failure of access.¹³ The strongman in the corner office commits the central planner’s error — at the worst possible moment.

Who’s selling the mandate

And ask who profits from the urgency, because the loudest voices telling CEOs to seize AI from the top are not disinterested. I made this point about in-housing: when your adviser also sells what they’re advising, scrutinize the advice. Here it’s true twice over.

The consultancies have bet their next decade on AI transformation — well over ten billion dollars since 2023 — and BCG has told investors it expects AI work to reach roughly forty percent of revenue by 2026.¹⁵ A firm with that much riding on adoption is not a neutral narrator of whether you should adopt; and some of the most-cited research urging the C-suite to lead the charge is published by the very firms that sell the engagement to lead it.

The stranger incentive is underneath. AI is doing to consulting what it does to marketing production: leveling it. Junior analysts synthesizing research and building decks is exactly what a model now does in seconds — by one estimate the firms’ own tools already do eighty percent of that work, and roughly a hundred and fifty ex-consultants have been hired to train AI to do entry-level consulting, paid to teach the machine to replace the bottom of their own pyramid.¹⁵ So the expertise is worth less every quarter, and the new thing to sell is the transformation itself — the most billable possible version of the story. None of this makes the advice wrong. It makes it interested. And a handful of firms selling the same frameworks to every rival is, as I’ve argued, a machine for making everyone’s strategy look alike, now at the speed of software.

Leading isn’t the same as understanding

And leading is not the same as understanding what you’ve mandated. The research that shows CEOs owning AI shows comprehension lagging the enthusiasm: Pearl Meyer found that while 90% of boards say the C-suite owns AI, the C-suite itself splinters four ways on who actually does — and that the message reaching employees is usually just “start using it,” the quiet part (we’re not sure where) left off;¹⁰ BCG’s own follow-up is titled around the gap, aligned in theory and divided in practice.¹¹ The leadership isn’t fake; it’s lopsided — rich in conviction, patchier in comprehension, shot through with a performative streak the job-security numbers make hard to deny. A CEO who declared AI the priority and stood up a transformation has led in every visible sense. Whether he knows what it’s for is the question the dashboard doesn’t answer.

The companies that listened

It helps to remember what the other way looked like, because the bottom-up waves didn’t merely happen to companies. The winners watched what their own people and users were doing, and followed it.

Slack is the cleanest case. It was never conceived as a product — it was the internal chat tool a small studio, Tiny Speck, built to talk to itself while making a game, Glitch. The game failed. The tool, the thing the team couldn’t stop using, was the real discovery, and the founders had the honesty to pivot the whole company; Salesforce later bought it for nearly twenty-eight billion dollars.¹⁶ The same founder had done it once before: Flickr began as a photo feature inside an earlier failed game. Twitter started as a side project inside a podcasting company whose main business collapsed. Gmail came out of one engineer’s internal experiment. All of them were shaped by use before they were ever designed for a market. None arrived as a mandate; each was a signal from the edge someone near the top was paying enough attention to catch.

Even AWS, now the most profitable thing Amazon does, began as a 2003 engineering memo — two infrastructure people, solving Amazon’s own scaling pain, noting the platform they wanted could be sold to outsiders. Bezos backed it. The telling detail: the team built it in South Africa, and one of them later said the distance was the point — he spent much of his time, in his own words, trying to stay out of the CEO’s eyeline.¹⁷ The breakthrough needed room from the top, not command from it.

The mobile wave shows the same thing inside a big, cautious incumbent. When employees started bringing their own phones to work, most IT departments tried to ban them. Intel did the opposite: its security chief reasoned a ban would only drive the behavior underground, so the company chose to “run to the risk in order to shape it,” moving the point of trust to the employee. The result, by Intel’s own accounting, was on the order of five million hours of productivity a year — because it enabled what its people were already doing instead of fighting it.¹⁸ Hold that example; it is the closest thing we have to a playbook for shadow AI.

In none of these did leadership invent the answer. The answer was already in the building, or in users’ hands, and leadership’s entire job was to notice it and back it. That is the exact capability a top-down mandate switches off.

The trouble with being pushed instead of pulled

A bottom-up wave has a brake built into it. When a technology is pulled in by the people who use it, it earns its place one workflow at a time — users keep what helps and quietly drop what doesn’t. That slow, distributed selection is a form of governance, and the prior waves got it for free precisely because the boss wasn’t in charge of them.

A top-down mandate skips that test entirely. It arrives as a directive to adopt, often with a number attached, before anyone has answered the only question that matters — what is this for? — and a mandate measured on adoption is blind to whether the adoption is worth anything. It counts seats, hours saved, assets produced. It cannot count whether any of it moved the business.

If that sounds familiar, it should: it’s the same machine I’ve described from five other angles — the empty Strategy room in Making More ≠ Making Better, the unreinvested savings in AI Raised the Floor, the cost program that dies on an untouched power structure in Beware the Culture You Have Created, the category converging on one answer in The Convergence Trap, the fixed-cost studio that becomes a volume pump in In-Housing. I kept calling it a marketing problem. The deeper cause is upstream: the push itself. The last two waves were pulled in by people who wanted them; AI is being pushed down by people afraid of missing it — and a directive born of fear, measured by adoption, with no bottom-up brake, reliably produces motion that looks like progress and isn’t. The reversal we’re applauding removed the safety mechanism the last two waves quietly relied on.

What to do when you don’t know

So here is the view I’d defend, plainly. There is a kind of decision a leader should make alone, fast, from the top: the one where the answer is known and the job is conviction — kill this line, back that bet, hold this standard. Authority is for the known. But AI strategy in 2026 is not the known. Nobody — not the CEO, not the consultant, not the model — reliably knows which uses will compound and which will be dead weight in eighteen months.

And uncertainty inverts the rule. The harder it is to know the answer in advance, the worse it is to centralize the call — because what you need under uncertainty isn’t a single confident decider, it’s the largest possible array of accurate sense-points, and those live at the edge, with the people closest to the customer, the workflow, the data, the failure. Centralizing the decision throws away your best instrument at the exact moment you need it most.

This is Hayek turned into a management instruction, and the military turned it into doctrine a century before the economists named it: mission command. The commander sets the intent — what we’re trying to achieve and why — and pushes the actual decisions down to the people with eyes on the ground, because a plan dictated from headquarters is always slower and dumber than the judgment of whoever can see the terrain.¹⁴ Intent from the top; decisions at the edge; signal flowing back up faster than it flows down. That is the opposite of a mandate, and the right shape for a moment when the map is still being drawn.

It also dissolves the false choice the founder-mode debate sets up. The answer to the absent leader was never the omniscient one. A leader can be intensely engaged — close to the detail, demanding, present — and still treat the organization as a sensing instrument rather than an audience for decisions already made. Engagement and humility about the unknown are not opposites. The strongest way to lead this wave is to be everywhere in the question and nowhere near pretending to own the answer.

What I’d ask of leaders

Decide alone only where you actually know. Reserve the unilateral call for decisions where the answer is clear and the job is conviction. For everything genuinely uncertain — most of AI strategy right now — your job isn’t to supply the answer. It’s to build the thing that finds it.

Treat your people as the sensor array, not the audience. Your edge under uncertainty is that hundreds of smart, trusted people touch the market and the work every day — an instrument no dashboard can match. Set the intent, push decisions to whoever can see the terrain, and spend your energy metabolizing what comes back instead of broadcasting what you’ve already concluded.

Shape the shadow AI; don’t ban it. The unsanctioned use in your building is the most honest signal you have about where value is. Banning it doesn’t stop it; it blinds you to it. Run to it the way Intel ran to BYOD, and treat the people doing it as scouts who found something, not risks to contain.

Treat your own urgency — and your consultant’s — as a warning light. If you’re moving because everyone else is, or because the firm selling the transformation says the window is closing, that’s the herd, not a strategy, and in AI the herd walks everyone to the same place. Ask what the deck isn’t pointing at, and who profits if you follow it.

Measure belief, not adoption. The dashboards that say AI is working count seats, hours, and output. Put the question that decides whether a brand wins — are we moving what the customer believes? — back on the report, or you’ll keep optimizing the thing that’s easy to count and losing the thing that’s hard to fake.

Where this leaves us

For the first time in forty years, the people with authority to act on a generational technology are engaged from the start. That’s genuinely good; the prior waves never had it. The danger is only in what they do with the engagement — whether they spend it being the decider or the conductor.

The bottom-up waves were saved partly by the fact that the boss wasn’t in charge of them: the drag of distributed adoption forced a selection on what was worth keeping. This wave has no such brake, so the discipline has to be chosen rather than inherited — and the discipline isn’t complicated, only hard. Under real uncertainty, stop trying to be the smartest single point in the system and start being the thing that makes the whole system smart: set the intent, trust the edge, listen harder than you decide.

The companies that won the internet and the smartphone didn’t out-mandate their rivals. They out-listened them — caught the signal their own people were already sending and had the nerve to follow it. That option is still on the table. It just asks a leader to believe the most valuable thing in the building might not be sitting at the top of it, and to lead, for once, by asking.

A note on how this was written

The argument is mine: that the C-suite lagged every major technology wave because disruption entered from the edges; that AI reversed this by appearing to threaten the core; that the centralized, strongman posture is the wrong instrument under genuine uncertainty, where Hayek’s dispersed knowledge and the military’s mission command both point to the edge rather than the top; that the consultancies amplifying the mandate are doubly interested; and that the firms which won the last two waves did it by listening to their own people, not out-deciding rivals. I worked through it with Claude, Anthropic’s AI, using it to connect the case to the theory and to find and check the company histories, survey data, and sources. The judgments are mine; the factual claims are checked against the reporting cited below. Given that the subject is what happens when leaders stop asking what a technology is for, it would be a poor piece to be coy about where the assist came from.

Notes

1. Bill Gates underplayed the web in the first edition of The Road Ahead (Viking, 1995) before his May 1995 “Internet Tidal Wave” memo reversed Microsoft’s posture; the “internet will fail” view is Clifford Stoll, Newsweek, February 1995.

2. On consumerization of IT and BYOD as an employee-driven, bottom-up process IT initially resisted: TechTarget, “IT consumerization”; “IT Consumerization” (ResearchGate, 2020); “Bring your own device” (Wikipedia).

3. Everett M. Rogers, Diffusion of Innovations, Free Press, 1962. The adopter categories — innovators, early adopters, early and late majority, laggards.

4. Clayton M. Christensen, The Innovator’s Dilemma, Harvard Business School Press, 1997. On why incumbent leadership rationally under-reacts to disruptive technologies that enter at the low end or periphery.

5. BCG, “AI Radar 2026: As AI Investments Surge, CEOs Take the Lead,” 2026. 72% (nearly three-quarters) of CEOs are their company’s chief AI decision-maker; CEOs investing 8+ hours a week are likelier to generate value. Survey of ~640 CEOs and ~2,360 leaders.

6. World Economic Forum, “CEOs are all in on AI but anxieties remain,” January 2026, on BCG data. AI as a top-down endeavour and “the CEO’s mandate”; ~half of CEOs believe their job stability depends on getting AI right.

7. The Conference Board, 2026 C-Suite Outlook Survey, January 2026. ~43% named AI and technology the leading 2026 investment priority, ahead of product/service innovation and customer experience.

8. Microsoft and LinkedIn, 2024 Work Trend Index (~31,000 people, 31 countries). ~75% of knowledge workers use AI at work; ~78% of users bring their own tools (“BYOAI”); ~60% of leaders lacked an implementation plan.

9. On shadow AI: Cyberhaven (2025) and 2026 surveys (e.g., Salesforce) finding most workplace generative-AI use runs through personal accounts; Samsung’s 2023 ChatGPT ban after engineers uploaded source code is the canonical cautionary case.

10. Fortune, “Boards say the C-suite owns AI strategy. The C-suite doesn’t agree,” April 22, 2026, on a Pearl Meyer survey of 108 executives and board members (fielded February–March 2026): 90% of boards place AI ownership with the C-suite, while C-suite responses split 32/27/22/17 across the executive team, individual business-unit leaders, the layer just below the C-suite, and functional heads.

11. BCG, “CEOs and Boards Are Aligned on AI in Theory, but Divided in Practice,” May 2026. Survey of 625 CEOs and board members.

12. Paul Graham, Founder Mode, September 2024 (paulgraham.com), prompted by Brian Chesky’s Y Combinator talk; contrasts hands-on “founder mode” with delegating “manager mode,” citing Jobs, Musk, and Jensen Huang. Chesky later clarified the point was deep involvement and product expertise, not swagger.

13. Friedrich A. Hayek, “The Use of Knowledge in Society,” American Economic Review, 1945. The “knowledge problem”: the information needed for good decisions is dispersed, local, and tacit, which is why centralized planning underperforms distributed systems.

14. Mission command (Auftragstaktik) traces to the 19th-century Prussian army under Helmuth von Moltke the Elder and is codified in modern doctrine (e.g., U.S. Army ADP 6-0, Mission Command): commander’s intent with decentralized execution and disciplined initiative at the edge.

15. On consulting and AI: futureofconsulting.ai (January 2026) on $10B+ of AI investment since 2023 and in-house tools (McKinsey’s Lilli, BCG’s Deckster) doing ~80% of junior-analyst research and slide work; TheStreet (May 2026) on MBB rethinking fees, BCG projecting ~40% of revenue from AI work by 2026, and ~150 former MBB consultants hired to train AI on entry-level tasks; Fast Company (December 2025) on McKinsey layoffs and the commoditization of analytical advantage.

16. On products shaped by use before market: Slack began as Tiny Speck’s internal tool during the game Glitch and was acquired by Salesforce for ~$27.7B (2021); Flickr, Twitter (from Odeo), and Gmail followed the same pattern. Company histories and contemporaneous reporting.

17. AWS: the 2003 Benjamin Black–Chris Pinkham infrastructure memo proposing selling compute as a service; Pinkham led EC2’s build in South Africa. Brad Stone, The Everything Store (2013), and Seattle Times / Network World reporting; Pinkham’s remark about welcoming distance from the CEO is from Stone.

18. Intel BYOD: Intel’s annual IT report via Computerworld and InfoWorld (2013) — ~5 million hours of annual productivity gains and ~57 minutes saved per employee per day; CIO Kim Stevenson on moving the trust point to the user, and CISO Malcolm Harkins on choosing to “run to the risk in order to shape it” rather than ban employee devices (program began 2009–2010).

Builds on my earlier pieces “AI Raised the Floor. The Work That Matters Now Lives at the Ceiling” (May 26, 2026), “Marketing Costs & Controls: Beware the Culture You Have Created” (June 9, 2026), “The Convergence Trap” (June 10, 2026), “Making More ≠ Making Better” (June 17, 2026), and “In-Housing: What to Own, What to Rent” (June 24, 2026)

I want to be clear-eyed about this rather than tribal, because both sides of the table are getting the argument half-right. In-housing is not a fad to be sneered at, and it is not a transformation that wins by default. It’s a make-versus-buy decision, taken capability by capability, and most of the heat in the debate comes from treating it as an identity instead. So this is written for both rooms: for in-house leaders deciding what to own, and for agencies deciding what they’re actually for. The honest version helps both.

Because this is not a new experiment. We have run the in-house media trial before, at scale, and we already have results. For every brand that tells a clean story of double-digit savings, there’s another that quietly reversed course a year or two later, having learned that operating a media function is harder than it looks on a slide. [1] The trouble is that the current wave is being underwritten on the shallowest of the three reasons — cost — and it’s carrying a fixed cost into a market that has not yet had a genuinely hard year.

The one good reason, and it isn’t cost

Start by giving the case its due, because the strongest argument for in-house media has nothing to do with saving money.

In 2016, the ANA commissioned K2 Intelligence to look under the hood of the U.S. media-buying ecosystem, and the report landed like a grenade: undisclosed rebates and other non-transparent practices were, in K2’s account, pervasive. [2] By 2019, two-thirds of marketers had rewritten their media contracts in response. Many went further and pulled buying inside, not to be cheaper but to be able to see. That instinct has only been vindicated. The ANA’s December 2023 supply-chain study found that of every dollar entering a demand-side platform, roughly 36 cents actually reaches the consumer, with around $22 billion in efficiency sitting unclaimed in the murk between spend and impact. [3] A decade after K2, the industry’s own verdict is that the problem hasn’t been solved so much as moved — into programmatic, into walled gardens, and now into agentic AI trading systems the holding companies have every incentive to keep opaque. [4]

So when a brand says it wants control of its media, it is pointing at something real, and any agency that hears that as a threat rather than a brief has misread the moment. That is the defensible core of the whole movement. The trouble is that control and a fully staffed in-house buying desk are not the same thing, and the conflation is where the money gets lost.

Give in-house its due

Before the cautions, the credit — because it’s earned. The ANA’s 2026 In-House Agency Fact Book is a portrait of in-house done well, and it’s genuinely impressive. These aren’t skeleton crews pushing out banner ads. Bank of America’s Enterprise Creative Solutions runs 60,000-plus jobs a year; Citi’s Red Arc clears 4,000; CVS Health’s Heart Haus and Eli Lilly’s Lilly Marketing Studios each field over a hundred people. [5] Award-winning, high-volume creative and content operations sit close to the business, move fast, and know the brand cold — which is exactly the work in-house should own. No outside partner will ever know your brand the way the people who live inside it do.

And the smart version isn’t all-or-nothing. In the same Fact Book, 60 percent of brands with an in-house agency still keep an external agency of record, and most of these teams handle creative, content, and production while routing other work out. [5] That’s the clear-eyed model the loudest voices on both sides tend to miss: in-housing at its best isn’t replacement, it’s a portfolio decision — own the high-frequency, brand-critical, data-sensitive work; buy the rest. The brands in that book mostly aren’t choosing in-house or agency. They’re deciding, capability by capability, which is which. The argument that follows isn’t against in-house. It’s about the two places the current wave tends to overreach: media, and fixed cost.

The question in-housing keeps skipping

There’s a piece of economics from 1937 that answers this better than any deck I’ve seen. Ronald Coase asked a deceptively simple question: if markets are so efficient, why do firms exist at all? Why not just buy everything you need, transaction by transaction, on the open market? His answer was that using the market has costs — searching, negotiating, contracting, coordinating — and a firm internalizes an activity only when doing it inside is cheaper than buying it outside. The boundary of the firm sits exactly where those two costs meet. [6]

In-housing is nothing more than redrawing that boundary. Which means the real question was never “can we save money?” It’s “is the market genuinely more expensive than the firm for this particular activity?” And the honest answer depends on three things the cost case waves away. They happen to be the three places I think the current wave is fooling itself.

Crack one: expertise is built across clients, not inside one

Adam Smith gave us the constraint in 1776: the division of labour is limited by the extent of the market. [7] A specialist can only stay a specialist if there is enough varied work to keep them specialized. That is the entire reason the agency model exists. A good media practitioner stays sharp by seeing many categories, many campaigns, and every platform shift the moment it lands, because they’re touching all of it across a roster of clients. Pull that person inside a single brand and you’ve narrowed their world to one P&L, one category, one data set. The frontier keeps moving; their exposure to it doesn’t. Within a couple of years you have either an expert going slowly stale, or an expert who leaves for somewhere they can keep growing. Neither is the capability you thought you bought. You can replicate an agency’s org chart inside your building. You cannot easily replicate the thing that made the agency good, which is the breadth of work flowing through it.

Crack two: a fixed cost, funded by a volatile budget

This is the one I’d underline twice, because it’s the crack that opens quietly and then all at once. Cutting the agency fee feels like savings, but it’s really a swap: you trade a variable, scalable cost for a fixed one you now carry alone. An agency spreads its tools, talent, training, and buying leverage across dozens of clients, and it flexes — you scale the retainer, renegotiate, or in the worst case don’t renew. An in-house team is payroll, leases, and licenses that have to be paid every month whether the work is there or not.

The ANA’s own Fact Book shows how fixed this really is. In-house agencies are funded as cost centers, chargebacks, and subsidized allocations — standing overhead, not pay-for-use. [5] And they are young: 67 percent of the in-house agencies profiled were established in the last ten years, which means most were stood up during the longest bull market in memory and have never been through a real downturn. [5] A fixed cost built in good times is a very different animal in bad ones.

Now set that against the budget that funds it, which is one of the most volatile lines in the entire business. Gartner’s CMO Spend Survey puts average marketing budgets at about 11 percent of revenue in the years just before the pandemic; in 2021 they cratered to 6.4 percent, the lowest the survey has ever recorded, then snapped back to 9.5 percent in 2022 — and the swings didn’t stop there. Budgets fell roughly 15 percent in a single year into 2024 and have settled at about 7.8 percent of revenue in 2026, still around a third below the old norm, with 56 percent of CMOs saying they lack the budget to deliver this year’s strategy. [8] The largest companies move hardest of all — enterprises above $2 billion in revenue bottomed at 5.7 percent in 2021 — and Gartner now warns that lingering volatility is raising the odds of mid-year cuts, with 39 percent of CMOs planning to trim agency budgets and 39 percent planning labor reductions, in the same breath. [8]

Individual advertisers move even harder than the average, and rarely on a schedule they control. Temu spent 2024 as one of the largest digital advertisers in the country — the single biggest on X — and then, after new tariffs and the end of the “de minimis” import rule wrecked its economics, cut U.S. ad spend by roughly 95 percent year over year, dropping to the 51st-largest advertiser on the platform by early 2026, with comparable cuts on YouTube and TikTok. [9] It isn’t a new pattern, either: in 2020, Coca-Cola cut its advertising about 35 percent in a single year, a far steeper drop than its revenue decline. [9] The point is narrow and hard to dodge: a marketing budget can move that far, that fast, and usually because of something — a tariff, a recession, a bad quarter — the marketing team didn’t choose. A retainer follows it down. A fully loaded in-house team does not; it sits there at full cost while the revenue that justified it shrinks, until the only way to move the line is layoffs. Which is exactly why, in a downturn, the team you built to save money becomes the cost you can’t easily cut, in the function that was supposed to be the savings. That’s the bill coming due, and it’s arithmetic, not pessimism.

Crack three: speed is a process, not a postcode

The agility argument assumes that proximity creates speed — that work goes faster simply because the people doing it now sit down the hall. It doesn’t. Speed comes from strong process: clear briefs, fast decision rights, a tested workflow. That’s portable; a good agency carries it with them, and a badly run in-house team will be slower than the agency it replaced, because it inherits your procurement, your legal review, and your internal approvals along with the work. Moving people inside the building changes their badge, not their throughput. If your process is the bottleneck, in-housing imports the bottleneck and adds a payroll line on top of it.

Why media, specifically, is the hardest thing to bring home

The ANA calls media “the final frontier” of in-housing, and its own Fact Book shows why the phrase is earned. That book is overwhelmingly a catalogue of creative, content, and production studios; dedicated in-house media operations are the rare exception. The two clearest examples, Lenovo’s Global Media Center of Excellence and PepsiCo’s Digital Activation Team, are small — roughly 25 and 6 people — stood up only in 2021, and run alongside retained external agencies rather than instead of them. [5] Even among brands proud enough of their in-house shops to be profiled, media buying is the thing most of them still don’t fully own.

There are good reasons. Media is always-on, not project-based, so the team can never quiet down. The terrain churns constantly — retail media networks, CTV, the walled gardens, and now AI-driven buying — which makes the expertise problem from crack one most acute here, because the half-life of what you know is measured in months. And it rewards scale in negotiation in a way creative never did.

It also walks into a trap I described in my last piece. A fixed in-house team has a metabolism: it has to stay busy to justify its headcount, so output becomes the measure, and the function quietly converts from buying well to buying more (Making More ≠ Making Better). You set out to gain control of your media and end up with a volume machine that reports activity up the chain while the questions that actually move the business go unasked. The dashboard stays green. The advantage erodes underneath it.

Follow the money — and notice who’s holding the pen

Here’s the part that gets too little airtime. When a C-suite decides to in-house, who advised them? Increasingly, it’s the management consultancies — and they are not disinterested parties.

Over the past fifteen years, the consultancies spent billions buying their way into marketing services, chasing a brand-services market worth north of half a trillion dollars. Accenture built Accenture Song — for six straight years Ad Age’s largest digital agency network — through more than forty acquisitions, including Droga5 in 2019; Deloitte has Deloitte Digital, IBM has iX, PwC has its own digital arm. [10] They work top-down, selling “business transformation” to the C-suite, which is precisely the altitude from which the in-housing decision gets made, while the holding companies work bottom-up from creative and media execution. And the consultancies don’t just recommend the move. Several will help you stand up your in-house trading desk and then run it for you as a managed service. [11]

Sit with that. The same firm can advise you to leave your agency, sell you the transformation program to do it, and supply the competing capability you replace the agency with. That is not a conspiracy; it’s a structural conflict, and a quiet one. When the people advising you toward a decision are also positioned to profit from it three different ways, the advice deserves more scrutiny than it usually gets. Cui bono is not a cynical question. It’s due diligence.

So: worth it, payback, and when to cut

Is it worth it? Run the Coase test, activity by activity, not as ideology. Bring inside what is genuinely core, high-frequency, and data-sensitive and large enough to keep specialists sharp — that’s the rare combination where the firm beats the market, and it’s exactly where the best in-house shops in the Fact Book are winning. Keep buying what is episodic, frontier-dependent, or sub-scale, because that’s where an agency’s breadth and shared cost base will always win. And separate the two things the pitch fuses: the control you need and the buying you can rent. For many brands the right answer is a small internal control tower — the visibility, the data ownership, the contractual teeth, the head of media the ANA says any $50M-plus advertiser should have — sitting on top of media still executed outside. That captures the one good reason without taking on the fixed cost.

When should it pay back? The cost saving — cutting the agency margin — shows up in year one. It is also the shallowest, most seductive version of the win, and the one I’d trust least. The capability return is what you’re really underwriting, and it should show up in two to three years on business metrics: better-informed buying, faster learning loops, demonstrable lift. If at 24 to 36 months the only things you can point to are dollars saved and assets produced, you didn’t build a capability. You built a cheaper factory, and you should say so plainly.

When do you cut it? This is the hardest question, and it’s the reason this keeps getting deferred. Cutting an agency is a non-renewal — an email and a transition plan. Cutting an in-house team is a restructuring, with layoffs, politics, and sunk cost pulling against it. So in-house units calcify: kept alive long past their usefulness because killing them is expensive and embarrassing. The honest triggers to cut are simple and worth naming out loud — the savings stopped materializing, the team has fallen behind the frontier, your best people have left, or the budget got cut and the fixed cost is now a millstone. The trick is that you have to write those kill criteria down before you build, because you will not have the stomach to write them afterward.

If you run an in-house team

Know your lane, and defend it honestly. Your edge is brand knowledge, speed on high-frequency work, and proximity to the business — own that without apology. But size your fixed cost to your trough budget, not your peak, because the peak is the thing that disappears first in a bad year. Use agencies for what you’re structurally sub-scale on: media buying and programmatic (which, per the ANA’s own data, you’re probably routing out already), frontier expertise that needs more reps than one brand can supply, surge capacity, and the outside read that keeps you from slowly converging on your own assumptions. Needing an agency for those things isn’t a failure of the in-house model. It’s the model working as designed.

If you run an agency

Stop defending production. That work has gone in-house and to AI, and it isn’t coming back; treating its loss as the headline is how you lose the rest of the conversation. The opportunity is everything the Fact Book shows brands are not bringing in-house: media at scale, strategy, frontier capability, cross-category expertise, and the judgment to see what a single brand’s data can’t show it. The question a brand leader is really asking is no longer “can you produce efficiently?” — plenty of teams can, inside and out. It’s “can you bring us the thing we can’t build for ourselves?” Answer that, and in-housing isn’t your threat. It’s the thing that clears the field of everything you shouldn’t have been doing anyway.

Where this leaves us

None of this is an argument against in-housing, and none of it is a defense of agencies as they are. For the right activity, at the right scale, for the right reason, bringing media home is a sound and sometimes overdue move — and the transparency case alone justifies a hard look. The argument is narrower. Most of the current wave wasn’t underwritten on that case. It was sold on cost, agility, and control as a slogan, by parties who benefit from the sale, and it’s carrying a fixed cost into a market that has not yet had a genuinely hard year.

When that year comes, it will sort the builds that were capability investments from the ones that were cost arbitrage dressed up as transformation. The first kind will earn their keep. The second kind will become the line nobody can cut, in a function that was supposed to be the savings. That outcome isn’t a prediction so much as an arithmetic certainty waiting for a trigger.

The oldest question in economics still works. Before you bring it home, ask honestly whether the firm really beats the market for this particular thing — and make sure the person answering doesn’t get paid either way.

A note on how this was written

The argument here is mine: that the case for in-house media rests on a genuine reason (transparency and control) that the current wave has conflated with a weak one (cost); that the economics turn on scale, on expertise built across clients, and on process rather than proximity; that a fixed in-house cost funded by a volatile marketing budget is a structural mismatch that a downturn will expose; and that the management consultancies advising the in-housing decision are not neutral parties. I worked through it in conversation with Claude, Anthropic’s AI, using it to connect the case to the underlying economics, to find and check the industry data and sources, and to help structure and edit the prose. The judgments are my own, and the factual claims are checked against the reporting cited below. This piece builds on three earlier ones, linked where they’re load-bearing.

Notes

1. Adweek, “Taking Media In-House Is Not a Case of One Size Fits All,” December 2019. On the pattern of in-housing successes matched by quiet reversals, and on EA’s data-and-responsiveness rationale for taking digital buying in-house.

2. K2 Intelligence for the ANA, “Media Transparency Report,” June 2016. Found non-transparent practices, including undisclosed cash rebates, to be pervasive in a sample of the U.S. media-buying ecosystem; by 2019 the ANA reported that two-thirds of marketers had updated their media-agency contracts in response (MediaPost, April 2021).

3. ANA, “Programmatic Media Supply Chain Transparency Study,” December 2023. Analyzing log-level data from 21 major advertisers, the study found that roughly 36 cents of every dollar entering a DSP reaches the consumer, with an estimated $22 billion in efficiency available to client-side marketers; best practices could lift the effective figure to 50 cents or more.

4. Campaign US, “One Step Forward, Two Steps Back: Media Transparency 10 Years After the ANA’s K2 Report,” June 2026. On the persistence of the transparency problem in programmatic, walled gardens, and agentic AI trading, and on ANA’s Bill Duggan arguing that advertisers spending $50M+ on media need a dedicated internal head of media.

5. ANA, “In-House Agency Fact Book,” January 2026. Source for: 67% of profiled in-house agencies established in the last 10 years; 60% of brands with an in-house agency also keep an external agency of record; the prevalence of cost-center, chargeback, and subsidized funding models; volume figures (e.g., Bank of America’s ECS at 60,000+ jobs/year, Citi’s Red Arc at 4,000); and the small, recently established, agency-supplemented in-house media units at Lenovo (Global Media Center of Excellence, ~25 FTE) and PepsiCo (Digital Activation Team, ~6 FTE). The ANA’s “final frontier” characterization of media is from its 2023 in-house agency research.

6. Ronald H. Coase, “The Nature of the Firm,” Economica, 1937. The foundational treatment of transaction costs and the boundary of the firm — the make-versus-buy question.

7. Adam Smith, An Inquiry into the Nature and Causes of the Wealth of Nations, 1776. “The division of labour is limited by the extent of the market” — the constraint on how specialized a function can stay.

8. Gartner CMO Spend Survey, 2021 through 2026 editions. Average marketing budgets ran ~11% of revenue in the four years before the pandemic, fell to a record-low 6.4% in 2021 (5.7% for enterprises above $2B in revenue), rebounded to 9.5% in 2022, dropped ~15% into 2024, and sat at 7.7% in 2025 and 7.8% in 2026 — still roughly a third below the pre-pandemic norm. The 2025 edition warned that volatility raises the likelihood of in-year cuts (39% of CMOs planning agency-budget reductions, 39% planning labor reductions); the 2026 edition found 56% of CMOs say they lack the budget to deliver their strategy.

9. Temu: Digiday / Sensor Tower (June 2026) and CNBC / Sensor Tower (June 2025). Following 2025 U.S. tariffs and the end of the “de minimis” import exemption, Temu cut U.S. ad spend roughly 95% year over year, falling from the single largest advertiser on X (Jan–May 2025) to the 51st (Jan–May 2026), with cuts of ~74% on YouTube and TikTok. Coca-Cola precedent: per its 2020 annual report (Campaign / WARC, March 2021), advertising expense fell ~35% to ~$2.8 billion in 2020 from ~$4.25 billion in 2019 — its lowest since 2007 — against an 11% revenue decline.

10. Accenture newsroom (April–May 2019); Ad Age Agency Report; and The Drum (April 2022). Accenture Interactive — rebranded Accenture Song in 2022, consolidating 40+ companies under one P&L — was ranked Ad Age’s largest digital agency network for six consecutive years and was built through more than forty acquisitions, including Droga5 in 2019. On the broader land-grab (Deloitte Digital, IBM iX, PwC) and the $500B+ brand-services market: AdExchanger (2017) and RockWater (August 2025).

11. AdExchanger, “Where Consultancies Play in the Media-Buying Space,” 2017. On consultancies helping clients launch in-house trading desks and running that media as a managed service, and helping clients bring programmatic in-house.

Builds on my earlier pieces “AI Raised the Floor. The Work That Matters Now Lives at the Ceiling” (May 26, 2026), “Marketing Costs & Controls: Beware the Culture You Have Created” (June 9, 2026), and “Making More ≠ Making Better” (June 17, 2026).

If you want a single diagnostic for where marketing is heading, try asking your team the question that used to start every campaign: what do we want our audience to believe? In a lot of organizations that question now lands with an awkward pause. Not because the team is weak. Because the building has quietly been reorganized around a different question: how much more can be we do for customers because of AI?

I’ve written pieces about this before. That AI has reduced barriers to production and will push the work that matters up to a higher level. That the same tools, fed the same data, are pulling every brand toward the same answers. That cost saving marketing programs almost never stick. This article here is about the thing those arguments have in common, which is the structure we’re building to hold all of it. The in-house studio, the content operation, the centralized production function: we are constructing them at speed, and almost all of them are volume-oriented constructs. We are building a faster way to make more, at the exact moment more stopped being the thing that wins.

What we’re building

The case for in-housing is typically made on cost. Bring production in, cut the agency line, capture the savings AI now makes possible. It’s tidy, and it’s easy to say up the chain. I understand the appeal; anyone holding a budget would.

But look at what you get when you build it. You don’t get a cheaper version of strategy. You get a studio, and a studio has a metabolism (same holds true for in-house media – but that is for another time). It has to stay busy to justify its headcount. Busy means output, so output becomes the measure, and once output is the measure the machine runs on its own: budget pressure leads to an in-house build, the build creates a fixed cost that has to be fed, feeding it makes volume the KPI, and volume crowds out the only thing that was ever scarce, which is the time to ask why. By the end you have a marketing function that has become an operations function, and nobody chose that in a meeting.

Nobody designs this. It is organic - each step a reasonable response to the one before it.

The numbers describe the same drift. Three-quarters of content professionals say AI has increased the volume they produce.¹ Nine in ten plan to increase output further, and nearly half expect three to five times more. Meanwhile the share of marketers who believe their own strategy is effective sits at sixty-one percent.² One strategist now carries roughly eighty-three percent more accounts than five years ago.³ We are pouring capacity into making more and thinning out the people who decide what’s worth making. The Observatory International’s Stuart Pocock put the risk plainly: the danger is producing stuff to meet an expectation rather than a need.⁴

I made the case in an earlier article that when production falls to near-zero cost, competition moves up to the strategic ceiling. Here’s the uncomfortable corollary. In-housing invests at the floor. We are building capacity in the one layer AI just commoditized and calling it a transformation.

Why it won’t fix itself

The natural objection is that this is a transition problem. The efficiency is real, the savings are real, and surely once the dust settles teams will bank some of that freed up time and put it back into thinking. I’d like to believe it. Experience tells me otherwise. And the economists back me up, and they’ve said so since 1865!

That year William Stanley Jevons noticed something strange about coal. More efficient steam engines were supposed to reduce coal consumption. Instead, consumption climbed, because efficiency made steam power cheap enough to use everywhere, and the everywhere swamped the savings.⁵ The pattern has held for a century and a half. Efficient lighting didn’t lower energy use; we just lit more of the world. Fuel-efficient cars didn’t cut fuel demand; we bought bigger ones and drove them further. When DeepSeek made frontier AI radically cheaper at the start of 2025, the market briefly assumed the world would need fewer chips. The opposite happened, and Satya Nadella named it within hours: Jevons strikes again, and as AI gets cheaper its use goes up, not down.⁶

The loop feeds itself: cheaper assets pull more demand, and the time AI frees up is spent making more.

Marketing is not exempt. The mechanism is almost insultingly simple. AI lowers the cost per asset. Lower cost per asset raises demand for assets, because now you can afford the variant, the test, the localized cut, the just-in-case version. Demand rises faster than efficiency, so total workload goes up, and the freed capacity gets spent before anyone decides to spend it. The thirty percent of production time AI gave back doesn’t become strategy time. It becomes more production. The machine eats its own dividend.

So, the in-house studio isn’t a neutral container you can later fill with better thinking. It’s a demand pump. The more efficient you make it, the more it asks of you, and the busier the people who were meant to govern it become. This is the part most transformation plans miss. You’re not just installing a tool. You’re installing an appetite.

Motion that looks like progress

What makes this hard to catch is that the dashboards stay green. We measure what the operation is good at producing, which is operational metrics. Roughly half of CMOs report AI’s value as time saved; forty percent as cost saved; a quarter as more content capacity.⁷ All real, all worth having. But notice what isn’t on that list: brand positioning, competitive differentiation, long-term demand, audience insight, market-share trajectory. The things that decide whether a brand wins are absent from the report that says the AI investment is working.

Follow the money and the same story repeats. Budgets have shifted toward paid media, up sharply year on year, while the lines that fund strategic input, labor and agencies, decline.⁸ Nearly six in ten CMOs say they don’t have the budget to execute their strategy, even as overall spend holds steady at a flat share of revenue.⁹ And the proof gap is widening: nine in ten marketers now use AI daily, but the share who can demonstrate a return has fallen, not risen.¹⁰ Customer retention and brand-value measures hit two-year lows in 2025, in the same window AI adoption roughly doubled.¹¹ The pattern is consistent enough that eMarketer summarized the whole period as productivity gains that never showed up in engagement or results.¹²

The gap between motion and outcomes

That is the indictment, and it’s worth stating without flinching. More activity, weaker brand outcomes. The factory got bigger and the results got softer. Motion is easy to mistake for progress, especially when the motion is fast and the charts point up. But a machine running efficiently toward no particular destination is not a strategy. It’s a very good way to get lost on schedule.

There’s a sharper edge to it, which I got into when I wrote about convergence. A volume machine fed by AI doesn’t only waste effort. It actively takes away difference. The more competent, on-trend, algorithmically shaped content you push into the market, the more you sound like everyone else doing the same thing with the same tools. So the construct we’re praising for its throughput is, underneath, eroding the one asset that earns a premium. We are scaling our way toward sameness and reading it off the dashboard as productivity.

The fix that isn’t more process

The obvious response is governance. Stand up a framework, write the standards, appoint someone to coordinate, put a filter on the front door. I believe in the filter, and I’ll come back to it. First I want to be honest about why most governance dies.

Governance fails for three reasons, and they tend to arrive together. The people who should make the value calls are the same people buried in production, so governance becomes a meeting nobody can afford to attend. Strategy and operations don’t share a language; one counts impact, the other counts throughput, and every prioritization turns into a negotiation between two rulebooks. And the moment leadership asks for more, a well-built framework folds, because nobody in the room has the standing to say stop.

This is the same wall cost-reduction programs hit. A coordination layer dropped onto an untouched power structure gets politely ignored by the people with real authority, who keep the parts that help them hit their number and route around the rest. The harder the center pushes, the more ingeniously the edges evade, and you end up paying for the governance and the workarounds both. The lesson, which I keep relearning, is that you can’t fix what the culture produces without touching the culture, and culture lives in incentives and authority, not in frameworks and tooling.

So, the version of governance that survives contact with a busy production organization is narrow and has teeth. It rests on one question, applied to every brief before anyone resources it: does this work move an audience from one belief to another, or does it just fill a channel? A request that names a specific belief shift, tied to a real priority, aimed at a defined audience, measurable in behavior, passes. A request that exists to keep a calendar full, or because it’s fast and cheap to make with AI, doesn’t. That filter does the work of a committee in a single sentence.

But a filter needs an owner, and the owner needs power. Not a new team. One named individual, reporting to the marketing lead rather than to operations, who chairs a short weekly gate where new work either earns a brief or doesn’t, and who holds genuine authority to stop, redirect, or defer. Governance without that authority is a suggestion. And authority without air cover is theater, which is why the load-bearing piece is the leadership contract: when the value lead stops a piece of work, leadership backs the call, with no route-arounds and no just-this-once. The contract also disarms the thing that breaks every framework, the reflex to ask for more, by making volume a trade rather than a default. Every new ask names what stops to make room for it. Ask for more, and you have to say what you’re giving up. That one rule is what keeps the appetite from quietly winning.

What I’d ask of leaders

Audit the productivity dividend. AI gave you back time. Find out where it went. If it flowed straight back into more production, the machine ate it, and you funded an operations expansion while believing you were freeing your strategists.

Protect strategy time as a hard line. If your best thinkers spend most of the week on output, you have an operations team, not a marketing team, whatever the org chart says. Carving out and defending that time is now a leadership act, not a scheduling preference.

Make the value question someone’s real job. Name the person, give them the gate, and give them the authority to stop work. Then, when they use it, back them in public. The role only works if the stop button is real.

Make “do more” a trade. The most useful sentence a leadership team can adopt is that every additional volume request names what it displaces. It costs nothing, and it defuses the whole Jevons dynamic.

Measure belief, not just busyness. Put strategic outcomes back on the report that decides whether AI is moving brand preference/share of consideration – is it bringing about a real shift in what the target audience thinks. If those never appear, you’ll keep optimizing the thing that’s easy to count and losing the thing that’s hard to fake.

Where this leaves us

None of this is an argument against AI, or against efficiency, or even against in-housing as such. The tools are remarkable, the savings are real, and you should take them. The argument is narrower and, I think, harder. We are building structures whose native unit is volume, at a moment when volume is the cheapest and least differentiating thing in the market, and we’re doing it faster than we’re deciding what the volume is for.

AI didn’t hand marketers superpowers. It handed them more rope, and a lot of organizations have used it to build an elaborate machine that runs beautifully toward no particular place. The factory keeps getting bigger. The blueprint is the part that still must come from a person willing to sit in the empty room and answer the question nobody has time for. The good news is the room is still there, and the question still works. Someone just has to walk back into it.

A note on how this was written

The argument is mine: that in-housing and the content-operation build-out are volume-oriented constructs, that the Jevons dynamic guarantees AI’s efficiency dividend gets spent on more output unless leadership governs for value, and that no governance layer survives a busy production culture unless it reaches real authority and incentive. The piece draws together three of my own talks and connects them to my earlier writing. I worked through it with Claude, Anthropic’s AI, using it to structure and edit the prose, build the diagrams, and check the figures against the sources cited below. The judgments are my own. I think it’s worth being plain about where AI helped, given that the subject is what happens when we stop being plain about it.

Notes

1. Canto & Ascend2, Content Marketing & AI survey, 2026 (≈75% of content professionals report AI has increased their output); 10x, 2025 (≈91% plan to increase content output, with nearly half anticipating a 3–5× increase).

2. Digital Marketing Institute, 2024: roughly 61% of marketers consider their marketing strategy effective. A comparable figure appears in HubSpot’s 2024 reporting.

3. Fluency, 2025: workload analysis indicating strategists now carry roughly 83% more accounts than five years earlier.

4. Stuart Pocock, Co-Founder, The Observatory International, on the risk of producing content to meet an internal expectation rather than a genuine audience need.

5. William Stanley Jevons, The Coal Question, 1865. The original statement of what is now called the Jevons paradox: gains in the efficiency of a resource tend to increase, not decrease, total consumption of it.

6. On the broad pattern (lighting, vehicle miles, compute) see standard treatments of the Jevons / rebound effect. The DeepSeek episode and Satya Nadella’s “Jevons paradox strikes again” remark date to late January 2025, following DeepSeek’s release of a low-cost frontier-class model and the brief sell-off in AI hardware.

7. Gartner CMO Spend Survey, 2025. Reported sources of AI value: time efficiency (≈49%), cost efficiency (≈40%), increased content capacity (≈27%), reduced agency spend (≈22%). Strategic outcomes such as positioning strength and brand-equity growth were largely absent from how value was reported.

8. Gartner CMO Spend Survey, 2025. 2025 budget allocation showed paid media rising to roughly 31% (up sharply year on year), with labor and agency lines declining.

9. Gartner CMO Spend Survey, 2025: ≈59% of CMOs report lacking the budget to execute their strategy; total marketing budgets held flat at roughly 7.7% of revenue for a third consecutive year.

10. Jasper, 2026: ≈91% of marketers actively use AI in daily work, while the share able to prove AI ROI fell to ≈41% (from ≈49%).

11. The CMO Survey (Spring 2025): customer-retention and brand-value measures reached two-year lows in 2025, during a period in which AI adoption roughly doubled.

12. eMarketer, 2026, summarizing the period: productivity improvements have not translated into stronger engagement or measurable business results.

Builds on my earlier pieces “AI Raised the Floor. The Work That Matters Now Lives at the Ceiling” (May 26, 2026), “Marketing Costs & Controls: Beware the Culture You Have Created” (June 9, 2026), and “AI: Bought to Stand Out, Built to Blend In” (June 10, 2026).

This isn’t a knock on AI. It’s a description of what the machine does. AI makes its recommendations by finding patterns in past data and projecting them forward. So when every competitor in a category feeds it overlapping data, tunes it against the same benchmarks, and leans on the same published research¹, the answers come back looking alike. Feed similar inputs to any algorithm and you get similar outputs. That’s arithmetic, not a bug.

And it lands hardest on the one thing that matters most. Differentiation, the real kind that holds up and earns a premium, is by definition a move away from the norm. It’s the position nobody else is standing in, the idea nobody else has claimed. AI, trained on what already worked across the whole field, is built to recommend the norm. It pulls toward the center of the pattern, never away from it.

The tool that promises differentiation is wired to produce its opposite.

How the sameness happens

It helps to trace the path, because the mechanism is simple once you see it.

A brand turns on an AI strategy tool, or hires an agency running one. The tool takes in what’s available: category research, positioning maps, audience data, campaign histories, social listening, search trends. It finds patterns, surfaces openings, makes recommendations. The brand acts.

Now picture every serious competitor doing the same thing at the same time. The inputs overlap, because much of the research comes from the same published sources. The frameworks overlap, because the major tools share methods underneath. The metrics they optimize against are the industry’s standard ones. Run that process in parallel across a whole category and you get a landscape where everyone is being steered toward the same open spaces, the same audiences, the same messaging.

The white space AI finds doesn’t stay white. A gap that every competitor’s algorithm spots at the same moment is a fishing tip texted to every boat in the harbor: accurate, and worthless by the time you get there. That’s the part the dashboards don’t show. The recommendation isn’t wrong. It’s right for everyone at once, which is the one way a right answer loses its value. What’s left is a category where the brands all sound alike, not because anyone ran out of ideas, but because the instrument they’re all holding keeps handing them the same answer.

And the loop feeds itself. The convergent work this process produces goes into market, performs, and becomes the data the next round of analysis learns from. The model is learning from a world the model is busy flattening. Every cycle, the center of the pattern gets denser and the edges get thinner.

Les Binet and Peter Field saw the shape of this a decade before AI made it sharp.² Working through the IPA’s databank of hundreds of campaigns, they found that marketing actually does two different jobs. One is short-term activation: the promotion, the offer, the targeted push that produces a quick sales bump and then fades. The other is long-term brand-building: the broad, emotional work that moves slowly and compounds over years. The campaigns that performed best put roughly sixty percent of their money behind the slow brand-building, because that is what creates durable advantage, even though almost none of its payoff shows up in this quarter’s numbers.

Data-optimized marketing does the reverse. It chases the activation it can measure now and starves the brand-building it can’t, which is exactly the work that produces lasting difference: the emotional pull, the distinctive identity, the consistent voice that Byron Sharp calls mental availability.³ AI inherits that bias by design. It recommends what the data says works. And what the data says works is what everyone is already doing.

The focus groups got there first

If there’s comfort in any of this, it’s that the industry has run a version of the experiment before.

Go back to the research-driven years of the 1970s and ’80s. As more brands hired the same research firms, ran the same focus groups, and tested against the same panels, the conclusions converged. Categories filled with brands making nearly identical claims to nearly identical audiences in nearly identical ways, all of them convinced, with the full authority of the data, that this was the right place to be.

The brands that broke out were the ones with the nerve to go against what the research said. Volkswagen’s “Think Small” ran straight at a consensus that Americans wanted big, powerful, status cars. Apple’s “1984” and “Think Different” ignored every convention the data said computer advertising should follow. None of that came from the numbers. It came from a conviction the numbers of the day would never have backed.

Volkswagen, “Think Small.” Doyle Dane Bernbach, 1959.

The investor Howard Marks has spent a career making the same point about any competitive field: you cannot move with the crowd and expect to beat it.⁴ To outperform, he argues, a position has to be two things at once, different from the consensus and right, and the catch is that being different from the consensus means holding a view that looks wrong to almost everyone until it pays off. The look of wrongness and the fact of differentiation are the same thing. A position that looks obviously right to everyone, including an AI trained on everyone’s data, is a position everyone is already walking toward.

Jack Trout and Al Ries had been making the related case in marketing terms since the 1970s. They were the ad men who gave the industry the idea of “positioning,” first in a 1972 series of Advertising Age articles and then in the 1981 book that grew out of them, and their argument was that in an over-communicated market you don’t win by having a better product, you win by owning a clear, distinct position in the customer’s mind. Differentiation lives in perception, not in the product.⁵

The AI version is the old research problem at much higher speed and scale. Same dynamic. The fix is the same too, except it now takes more nerve to apply, because the convergent answer doesn’t arrive as a focus-group summary. It arrives wearing the authority of the algorithm.

The split that’s already starting

You can already see the market dividing in two, even if nobody’s named it yet.⁶

On one side are the brands and agencies that have wired AI through the whole strategy stack, not just execution and optimization but audience insight, positioning, the campaign idea itself. They move fast. They produce at volume. They also keep arriving, in aggregate, at the same conclusions, and you can hear it in the work.

On the other side is a smaller group that drew a line on purpose. AI runs execution, optimization, the analytical heavy lifting. Human judgment, helped by AI but not handed over to it, owns positioning, brand, and the thinking that has to be genuinely new. These brands aren’t winning the speed race at the production layer. They’re winning where it counts, because the positions they’re building aren’t the ones every rival’s algorithm is finding at the same moment.

The Kantar BrandZ data points the same way.⁷ Across categories, brand equity, and the pricing power and resilience that come with it, is concentrating in the brands that score highest on distinctiveness. The ones gaining ground aren’t the ones with the most content or the slickest AI. They’re the ones with the clearest, most consistent, most emotionally resonant identity. That isn’t an output. It’s a human achievement.

What the differentiated brands do differently

The trap isn’t fate. It’s the sum of a few specific choices, and the choices can go the other way. Here is what the evidence says the brands escaping it are actually doing.

Draw a hard line between what AI decides and what people decide. Not every strategic call is the same kind of problem. Give AI the optimization, the execution, the audience intelligence, the pressure-testing of human strategy. Don’t give it the positioning, the creative platform, or the category-defining bets that only pay off when they’re genuinely new. Drawing that line and holding it is the foundational discipline, because if you don’t, AI’s territory expands on its own until it reaches the exact decisions where its limits do the most damage.

Treat algorithmic consensus as a warning light, not a green light. When the AI points confidently at a territory, when the white space is mapped and the angle is obvious, the right reaction is suspicion, not relief. If your AI can see it, so can everyone else’s, and the value of that territory is already draining away. The useful question isn’t “should we go here?” It’s “what is the algorithm not recommending, and why?” Mark Ritson has made the narrower version of this point for years: the decisions that need real originality are exactly the ones data-derived insight handles worst.⁸ McKinsey’s own adoption research shows strategic outputs converging across peer sets, which is the trap turning up in the data.⁹

Own the insight AI can’t commoditize. Underneath everything, this is a data problem. When everyone pulls from the same published research and the same platform behavior, AI hands everyone the same strategy. The antidote is proprietary: primary research, first-party data, real audience relationships, cultural reads that live in no published set. That’s where differentiation begins. Brands that build their own insight aren’t just feeding the machine better inputs; they’re building a moat a competitor can’t cross by buying the same software. The Content Marketing Institute has tracked the flip side of this, the steady decline in distinctiveness of AI-assisted content and the rising premium on original perspective.¹⁰

Rebuild the nerve to back a call the data won’t. Part of this is technology and part of it is courage. Years of data culture have built organizations where a strategy needs a data justification to survive, where a recommendation analysis can’t support starts out at a disadvantage no matter how good it is. That’s not irrational. (Nobody ever got fired for following the data, which is precisely the problem.) But it makes the genuinely differentiating positions, which are by definition the ones the data doesn’t yet support, the hardest ones to get approved. The Adobe and Oxford Economics work on CMO priorities found the same thing from another angle: what separates the AI leaders from the laggards isn’t tool access, it’s strategic alignment.¹¹ Building the capacity to make and defend a data-divergent call, and protecting the people who generate those calls, matters as much as any deployment decision you’ll make.

Ask your agency for the thing convergence makes scarce. The strategic creative relationship has never mattered more, and convergence is the reason. A partner who brings real category knowledge, a cultural read, creative conviction, and the willingness to put forward a position your data won’t obviously support is offering the one thing AI can’t. The question for a brand leader looking at the roster isn’t “can they produce efficiently?” Plenty can. It’s “can they see what our data can’t show us?” For partners whose value has mostly been production, that’s a real and human shift, and it deserves candor on both sides.

Where this leaves us

The competitive landscape of the next decade won’t be decided by who deploys AI most completely. It’ll be decided by who understands its limits most clearly. AI is a powerful instrument for executing a distinctive position. It’s a poor one for finding it. And in the hands of an entire industry at once, it’s an instrument for sanding down the distinctions that are already there, a little more with every cycle, while every dashboard reports that things are going well.

David Ogilvy said the most important word in advertising is “insight,” the deep human read on what a person actually wants and fears and values, the thing no surface analysis can fake.¹² He was right when he said it and he’s still right. AI can process information at a scale Ogilvy couldn’t have pictured. What it can’t do is generate the kind of understanding that makes a brand genuinely hard to replace in the minds of the people it serves.

The brands that get out of the convergence trap won’t be the ones using AI less. They’ll be the ones using it with the clearest possible sense of exactly what it can’t do.

A note on how this was written

The argument here is mine: that industry-wide AI adoption is structurally biased toward strategic sameness, and that the brands which stay differentiated will be the ones that fence off human judgment for the decisions AI is built to get wrong. I worked through it in conversation with Claude, Anthropic’s AI, using it to connect the argument to the existing research, to find and check the examples and sources, and to help structure and edit the prose. The judgments are my own, and the factual claims are checked against the work cited in the notes. I think it’s worth being plain about where AI helped, not least because it’s the whole subject of the piece.

Notes

1. IAB, “State of Data 2025: The Now, the Near, and the Next Evolution of AI for Media Campaigns,” 2025. On the concentration of AI tooling and overlapping training-data sources across holding companies and major brands. iab.com/news/iab-state-of-data-report-2025

2. Les Binet and Peter Field, “The Long and the Short of It,” Institute of Practitioners in Advertising (IPA), 2013. On the tension between short-term performance optimization and long-term brand building, and the underinvestment in brand equity that data-led optimization tends to produce.

3. Byron Sharp, “How Brands Grow,” Oxford University Press, 2010. On mental availability, distinctive brand assets, and the drivers of market-share growth.

4. Howard Marks, “The Most Important Thing: Uncommon Sense for the Thoughtful Investor,” Columbia Business School Publishing, 2011. On the concept of “second-level thinking” and the argument that outperforming the consensus requires being both non-consensus and correct.

5. Jack Trout and Al Ries, “Positioning: The Battle for Your Mind,” McGraw-Hill, 1981, building on their 1972 “Positioning Era” series in Advertising Age. The foundational argument that differentiation is a matter of perception rather than product superiority.

6. EY, “How AI Is Reshaping the Future of Marketing,” EY Studio+ Future of Marketing Report, 2026. On the emerging split between brands investing in AI-assisted execution and those investing in AI-informed but human-led differentiation. ey.com/en_us/insights/cmo/how-ai-is-reshaping-the-future-of-marketing

7. Kantar BrandZ Global Report, 2025. On the concentration of brand equity and the relationship between distinctiveness and pricing power across more than 100 categories.

8. Mark Ritson, Marketing Week columns and Melbourne Business School executive education, 2024. On the limits of data-derived insight in positioning decisions that require genuine originality.

9. McKinsey & Company, “The State of AI in 2025,” April 2025. On convergence in AI-assisted strategic outputs across competitive peer sets. mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai

10. Content Marketing Institute, “B2B Content Marketing Benchmarks, Budgets, and Trends,” 2025. On the declining distinctiveness of AI-assisted content and the rising premium on original research and proprietary perspective.

11. Adobe / Oxford Economics, “2026 AI and Digital Trends Report: CMO Priorities for AI,” April 2026. On the finding that strategic alignment, not tool access, is the primary differentiator between AI leaders and laggards. business.adobe.com/resources/reports/cmo-digital-trends.html

12. David Ogilvy, “Confessions of an Advertising Man,” Atheneum, 1963. On the relationship between genuine consumer insight and advertising effectiveness.

The instinct is sound, and the timing is no accident. While a company is growing it tolerates a remarkable amount of operational slack, and for the most part it is right to. The duplicated tools, the overlapping agencies, the two teams quietly building the same thing: all of it is affordable while the top line is climbing, and more than affordable, it feels like the price of moving quickly. Nobody examines the machinery too closely while the numbers are good. The slack accumulates, year after year, and goes unremarked, because growth pays for it and hides it at the same time.

When growth flattens, the same slack that looked last year like the price of speed looks this year like waste, and management turns on it. That is the natural order of things, and it is where most cost programs begin. The trouble is what comes next, because this is a far harder problem to solve than it first appears. This is a reversal of culture. This is a family you don’t recognize.

What they find, once they go looking, is real and has a name. (and the “they” is typically an outside consultancy). The media consultancy Mediasense, recently called it marketing’s invisible tax: a loading of roughly 10% on all the non-working investment, spent on coordinating, reconciling, re-briefing, and waiting.¹ Multiple ways of work. Fragmented tools. Measurement that won’t reconcile. Agencies pulling in different directions. The instinct is to read all of this as an operating-model problem, so the remedy is an operating-model remedy: governance, shared standards, a single source of truth, decision rights, someone appointed to orchestrate the ecosystem. Every piece of it is sensible.

And almost none of it will take. Not because the design is wrong, but because the company has misread what kind of problem it has.

The problem isn’t the ecosystem

The fragmentation isn’t the disease. It’s a symptom. The company built it on purpose, even if no one would say so out loud, by spending a decade rewarding a particular kind of person: the P&L owner who acted like an entrepreneur: moved fast, made their own calls, owned their own number, and didn’t wait for the permission. That culture is what produced the growth. It is also what produced the sprawl, because a building full of people optimizing hard for their own outcomes will, of course, buy their own tools, hire their own agencies, and build their own version of everything. The messy ecosystem is the exhaust of an entrepreneurial culture running hot. You cannot change the exhaust without touching the engine.

The theory saw this coming, and the theory is right. Larry E. Greiner described it back in 1972.² The autonomy that drives a company’s growth eventually produces a crisis of control, and the cure, coordination, is not a tune-up. It is a transfer of power away from the people who had been running their own show. That is a real fight, with real losers, and companies flinch from it.

Donald N. Sull, in 1999, explained the flinch and what it costs.³ Companies in trouble respond with energy poured into the old reflexes. And here is the part that matters: the core reflex of an entrepreneurial culture is finding the clever way around an obstacle. It is the exact skill that won them the market. So when you hand that culture a new set of ecosystem rules, you have handed a room full of expert problem-solvers a fresh problem to solve, and the problem is you. They will adopt the parts of the new system that help them hit their number and engineer their way around the parts that don’t. Not out of malice. Out of habit, and incentive.

Who actually calls the shots

Which brings up the thing the redesign almost always gets wrong. The new ecosystem, the governance and the standards and the orchestration role, is designed and typically owned by people who sit outside the P&L. They carry responsibility for the system and hold no authority over the people who have to live in it. The P&L owners still control the budget. They still carry the target. They are still told, in every way that counts, that hitting the number is their job. So when a standard they didn’t create and has no proven impact slows them down, or simply doesn’t move their number, they ignore it, and nobody with real power stops them, because the people imposing the change don’t own their results and can’t touch their incentives.

And there is a deeper reason they ignore it. The P&L owners are growth people. They were chosen, promoted, and paid for the ability to make a number go up, not for fluency in cost discipline or patience with centralized control. Asking them to carry a cost program is asking them to work against their own instinct and their own incentives. They will keep doing the thing they are good at, which is chasing growth, by the only method they trust, which is moving fast on their own terms. And the organization has always supported them doing that. Heck, that is how the people in real power got there themselves. To them, the cost program is simply one more constraint on growth to be managed around. The very people with the authority to make the change are the people least inclined to want it.

What you get is selective compliance, which is the most corrosive outcome of all. The shared data platform gets used where it helps and bypassed where it doesn’t. The briefing standard is honored by the teams it suits and quietly skipped by the teams it slows. The organization doesn’t reject the change cleanly. It absorbs the convenient parts and discards the rest. And because the entrepreneurial instinct is to re-route around friction (like GPS around a traffic snarl), the harder the center pushes, the more ingeniously the edges evade. The company ends up moving in the opposite direction from the one the redesign intended, now carrying the cost of the governance layer on top of the cost of the workarounds. It set out to cut costs and added them.

What an honest fix would cost

The honest version of this is much harder, which is why so few fully succeed. You cannot fix the ecosystem without fixing the culture, and you cannot fix the culture with an email decree or an offsite (though the latter lasts a week or two). The culture is held in place by what the P&L owners are measured and rewarded on, and by who holds authority over what. Changing it means a genuine restructuring: the incentives, the accountabilities, the reporting lines, the definition of meeting your annual plan. If the people who have to change are the P&L owners, then the change has to land on the P&L, with teeth, sponsored by someone who actually controls those numbers. Force a coordination layer onto the side of an untouched power structure and you are rearranging the plumbing while the people who control the water carry on as they please.

This is the unglamorous truth the new operating-model walking deck skirt, because restructuring authority creates winners and losers and political pain, while standing up a center of excellence feels like progress and offends no one. One of those changes behavior. The other mostly generates documents.

The ones who actually changed

The companies that came through this had the authority to act and used it on the real levers. When P&G’s pressure came, Marc Pritchard didn’t run a side governance project. With the standing of a global chief brand officer, with trust and authority of executive peers, he went at the contracts, the standards, and the spend, personally telling partners to clean up the “murky at best, fraudulent at worst” supply chain or lose the business, then cutting spend and showing that sales didn’t suffer.⁴ Authority, applied to the mechanism itself.

Airbnb is the sharper case, because Brian Chesky is the founder and chief executive, and when he changed the marketing model he changed the structure that held it rather than the rules around its edges. He dismantled the conventional siloed setup, folded marketing into a model he personally owned, and reset what the function was for.⁵ It stuck because the person with authority over the P&L made the change to the P&L, instead of asking people who answered to a different scorecard to please cooperate.

So what

None of this means autonomy is the enemy. The entrepreneurial culture is an asset, and an organization optimized into total compliance is one that has forgotten how to move. The point is narrower and harder than that. You don’t get to keep the culture exactly as it is and also fix what the culture produced. Something real has to give, and it has to give at the level of power and incentive, not process and tooling.

Until it does, the cost program is a suggestion, and the people it is aimed at are the most resourceful in the building and the least interested in the goal. They will take what helps them and leave the rest, and the company will look up a year later to find the structure unchanged, the workarounds multiplied, the costs no lower, and the drift pointed firmly the wrong way, wondering how a plan everyone nodded along to never quite made it into the work.

A note on how this was written

The core argument here is mine: that the inefficiency companies tolerate while growing is a product of culture, not operating model, and that no cost or ecosystem redesign survives contact with an entrepreneurial P&L culture unless the restructuring reaches the incentives and authority of the P&L owners themselves, who are built for growth rather than cost control. I developed the piece in conversation with Claude, Anthropic’s AI, using it to connect the argument to existing management theory, to surface and check the company examples, and to help structure and edit the prose. The judgments are my own, and the factual claims are verified against the reporting cited in the notes.

Notes

1. MediaSense, “Marketing’s Invisible Tax — The Cost of Poor Ecosystem Design,” 26 May 2026 (Siobhán Woodrow). The estimate of a ~10% loading on non-working investment is drawn from this piece. https://www.media-sense.com/2026/05/26/marketings-invisible-tax-the-cost-of-poor-ecosystem-design/

2. Larry E. Greiner, “Evolution and Revolution as Organizations Grow,” Harvard Business Review, July–August 1972; reissued May–June 1998. The phases of delegation and coordination, and the crises of control and red tape, are his.

3. Donald N. Sull, “Why Good Companies Go Bad,” Harvard Business Review, July–August 1999. The idea of “active inertia” is developed further in his book Revival of the Fittest (Harvard Business School Press, 2003).

4. Marc Pritchard, Chief Brand Officer, P&G, keynote at the IAB Annual Leadership Meeting, 29 January 2017; reporting in The Drum, CNBC and Marketing Week (Jan–Feb 2017). Pritchard later said an early cut to digital spend produced no measurable drop in sales. https://www.thedrum.com/news/2017/01/30/pg-review-all-agency-contracts-2017-four-step-plan-bring-transparency-media-supply

5. Airbnb Q1 2021 earnings call and shareholder letter; CEO Brian Chesky; reporting in Marketing Week and Campaign (2021–2023). Airbnb cut performance marketing, held roughly 90% of traffic as direct or unpaid near 2019 levels, and made the shift from performance to brand permanent. https://www.marketingweek.com/airbnb-performance-marketing/

The numbers tell part of it. The AI marketing sector reached $47.3 billion in 2025 and is growing at a 36.6% CAGR toward $107.5 billion by 2028.[1] In 2025, roughly a fifth of all video ad creative was built or enhanced with generative AI, and buyers expect that to approach 40% by 2026.[2] Yet only about a third of organizations have fully integrated AI across the campaign lifecycle.[3] A powerful production engine has been put in our hands, and most of us are still working out where we mean to take it.

There’s a space between the volume of what we can now produce and knowing why we are producing, and that space is where brand value is quietly eroding . It’s why I think the conventional wisdom about AI reducing the need for agencies has it backwards.

We’ve been here before

This isn’t the first time a new tool has made production nearly free, and the pattern each time has been reassuring. When desktop publishing arrived in the late 1980s, the expectation was that professional designers would be displaced. PageMaker and QuarkXPress put layout on every desk. The opposite happened. The volume of designed material exploded, a flood of competent but forgettable work hit the market, and the value of real creative direction went up. The designer’s job didn’t disappear. It moved from execution toward concept.

The content marketing era of 2012 to 2018 serves as another reminder. Once publishing platforms made content effectively free, nearly every brand set out to become a media company. The glut got so large that organic reach fell, email engagement softened, and attention became the scarcest thing in the ecosystem. The brands that came through it (HubSpot, Patagonia, Red Bull) didn’t win by publishing more. They won because they’d built coherent narratives that gave their content gravity. Production was cheap. The thinking behind it was not.

The Shopify-era direct-to-consumer boom told the same story. When launching an e-commerce brand cost almost nothing, thousands of new brands poured into every category. The ones that struggled usually had beautiful storefronts and thin brand thinking. The ones that lasted won on story, community, and the experience they designed.

In each case, as production costs fell toward zero, the value moved up to the strategic and creative layer. We’re seeing the same pattern now, only faster and at a far bigger scale.

The trap hiding in the cost math

The in-house argument rests on what I’d call the linear cost model: cut production cost, and you cut the total marketing investment by the same proportion. It’s a tidy idea that is easy to explain up the line. It also misses something structural about how campaigns behave and customers attention is drawn.

Every campaign has a strategic carrying capacity, a ceiling on how much content it can hold before volume starts to dilute the core idea instead of amplifying it. Push more content through a platform that isn’t strategically deep and you don’t multiply impact. You mostly speed up incoherence. So AI hasn’t only raised the floor for production. It has steepened the slope above it. When everyone can produce competent work cheaply, competent work stops setting you apart. The work isn’t bad. It just looks like everyone else’s. The competition moves up to the ceiling, where strategic creative thinking lives.

The IAB’s own research makes the gap plain: even with broad adoption, only about 30% of agencies, brands, and publishers have fully integrated AI across the media campaign lifecycle (see note 3). The capability has run ahead of the strategy, which is a normal place to be this early.

Where the gap shows up

What the holding companies are doing

If you want to see where value is moving, watch where the money goes. The holding companies are betting on strategic depth.

Publicis is the clearest example. Its acquisitions of Epsilon and LiveRamp, and the build-out of its CoreAI platform are a bet that the premium to come from connecting creativity, data, and strategy, not from producing more throughput. The results show it: Publicis grew 5.7% organically in Q3 2025, its eleventh straight quarter of outperformance, while much of the sector was flat or shrinking.[6] By the company’s own account, about three-quarters of its model now runs on AI, and its leaders keep making the same point: AI is there to extend strong creative and strategic minds, not replace them.

Major clients are reaching the same conclusion. Heineken’s May 2026 global agency review left the company with fewer agencies and deeper ties to the ones it kept. Global creative was concentrated across Publicis, WPP, and Stagwell, and Dentsu held onto global media. Heineken’s chief commercial officer put it succinctly: the aim was “moving to fewer, better and bigger agency partners.”[7] Consolidation used to broadly be a procurement-driven cost decision. It’s a rightfully become a quality decision, in a market where production is no longer the bottleneck.

What this does to the market

AI has removed the production barrier to entry in content marketing. Almost any organization, whatever its budget or headcount, can now produce content at volume. That’s the definition of a commodity market: near-zero marginal cost and near-infinite supply.

In a commodity market, the only durable advantage is being different. When every competitor can produce similar work at a similar cost, the deciding questions are human ones: what to say, who to say it to, and in service of what idea. That’s judgment, that’s differentiation, not output.

The Content Marketing Institute’s 2026 expert panel echoes this direction: toward brand voice, narrative clarity, and creative intention over sheer volume.[8] And the evidence keeps reinforcing that work grounded in human insight beats fully automated output on engagement and traffic. The reason isn’t that AI writes badly. It’s that without direction, its output looks like everything else in the feed.

Recommendations for Leaders

Take stock of your campaign’s carrying capacity before you scale production. Before you use AI to add volume, ask honestly whether the idea underneath the campaign can carry it. A strong platform has depth. It can travel across formats, channels, and audiences without contradicting itself. A thinner one gets exposed faster at scale. The question isn’t “how much can we make?” It’s “how much can this idea carry?”

Resist letting production savings become the whole ROI story. The savings from AI-assisted production are real, and you should take them. What matters is where the freed budget and capacity go next: into sharper positioning, deeper audience understanding, and stronger creative strategy. Savings that don’t get reinvested in strategy tend to buy volume without impact, and that’s a costlier problem to unwind than the one AI was meant to solve.

Ask more of your agency partners, and ask the right thing. If AI is carrying more of the execution, the value a good partner brings has shifted. The question is no longer “can they produce?” Now it’s “can they think alongside us?” Partners who bring real category expertise, cultural perspective, and the ability to build campaign platforms with genuine depth are more valuable now. And for partners whose contribution has been mostly production, this is a real and human transition, one that deserves candor and care on both sides.

Treat brand coherence as something you actively steward. At high volume, coherence doesn’t take care of itself. The more work you have in market at once, the greater the risk of dilution, tonal drift, and quiet contradiction. That takes real stewardship: clear standards, editorial oversight, and a layer of human judgment. The organizations doing this well have made brand stewardship a capability, not an afterthought.

Make differentiation the goal. When production barriers are near zero, the only lasting advantage is differentiation: original research, a real point of view, creative work that’s hard to copy no matter what tools made it. The brands that define their categories over the next five years will use AI to amplify differentiated thinking, not to cover for the lack of it.

Where this leaves us

AI was sold, in part, as a way to need agencies less. The evidence points the other way: the historical pattern, the IAB data, the holding-company results, and the choices big clients are making. What AI has done is sharpen an old question. What is a strategic creative partnership for?

The question for leaders isn’t whether to use AI. That’s settled. It’s whether the strategy underneath the deployment is strong enough to carry the volume AI makes possible. For a lot of organizations, it isn’t yet, and that’s understandable at this moment. That gap, between what we can produce and what we’ve thought through, is where brand value will leak and be lost: quietly, at scale, one competent but forgettable piece at a time. The good news is that it’s also where the most valuable work is waiting, for anyone willing to do it.

[1]AI marketing market size: SEO.com / Market Research, “50+ AI Marketing Statistics,” May 2025. The $47.3B figure appears across aggregators citing Statista / MarketsandMarkets; the 36.6% CAGR toward $107.5B by 2028 is consistent across sources.

[2]IAB, “2025 Digital Video Ad Spend & Strategy Full Report,” July 2025. Roughly 22% of video ad creative was AI-built or enhanced, projected toward ~40% by 2026; a majority of advertisers reported using or planning to use generative AI for video.

[3]IAB, “State of Data 2025: The Now, The Near, and The Next Evolution of AI for Media Campaigns,” March 2025. Survey of 500+ subject-matter experts across buy- and sell-sides; only ~30% of agencies, brands, and publishers have fully integrated AI across the media campaign lifecycle, leaving roughly half without a complete strategic roadmap.

[4]Marketing Week, “State of Marketing Leadership 2025” (cited via Marketing Agent Blog, December 2025): while ~83% of CMOs report having an AI strategy, only ~29% report consistent execution across business units. Corroborated by ABM Alliance, “CMO and Marketing Leaders Challenges Using AI in 2026,” April 2026.

[5]Averi.ai (citing 2025 research) reports human-written content generating 5.44x more traffic over five months and 41% longer session durations versus fully AI-generated content.

[6]Publicis Groupe, Q3 2025 Revenue Statement, October 14, 2025: +5.7% organic growth (its 11th consecutive quarter of outperformance); FY25 organic-growth guidance raised to +5.0%–+5.5%; roughly three-quarters of the group’s model described by the company as “AI-powered.” CEO Arthur Sadoun has characterized AI as a present-day driver of the group’s growth. Sources: publicisgroupe.com; Adweek; Marketing Dive, October 14, 2025.

[7]Heineken, press release, May 8, 2026, “HEINEKEN appoints new global marketing agencies to drive next phase of brand growth.” Global creative consolidated across Publicis, WPP, and Stagwell; Dentsu retained for global media. CCO Bram Westenbrink described the shift as “moving to fewer, better and bigger agency partners,” part of Heineken’s broader commercial transformation. Sources: theheinekencompany.com; Marketing Dive; Campaign.

[8]Content Marketing Institute, “42 Experts Reveal Top Content Marketing Trends for 2026,” December 2025. Expert consensus emphasized brand voice alignment, narrative clarity, and creative intention over sheer volume. contentmarketinginstitute.com.

A note on how this was written

The strategic frameworks, hypotheses, and conclusions in this post reflect my own original analysis. I used Claude (Anthropic) to assist with research synthesis, drafting, and editing - specifically to help adapt a longer internal point-of-view document into this format. I believe in being transparent about where AI played a role, not least because it’s directly relevant to the argument I’m making here: the thinking is mine; the assist was real.

Leave a comment