At its peak in 2021, Chegg was worth about $14.5 billion. Then its students went to free chatbots instead. The education company blamed AI for the collapse of its business, and by late 2025 its stock was down about 99% with 45% of its staff cut. No recession did that. A tool did, in about three years, to a company that had time to see it coming.

Keep Chegg in view while reading the numbers governments celebrate. In 2024, institutions in the United States produced 40 notable AI models. China produced 15. Europe produced three. Governments spend as if a nation’s future turns on that count, and none of it decides whether AI does a day of real work inside your company.

That is settled far from the scoreboard, in one workflow, by whether anyone rebuilt it around the model. Mostly no one has. A manager needs a line for the board that says the firm is doing something about AI, and finds one. The spend goes in, the adoption is reported, and the number the board watches does not move.

It is not that the tools are weak. Adopting AI and rebuilding the work around it are two different acts, and only the second moves the number. Across the firms I talk to, I see five ways companies take up AI in 2026. The first two rebuild the work, and they are priced for the largest companies. The middle two report adoption while the work runs as before. The fifth rebuilds one workflow at a price a mid-market firm can pay, and on the evidence I see it is the one that pays back.

The two enterprise paths

  1. The builders hire the bench. The largest, most successful corporations build the capability in-house: they hire AI engineers and data scientists, keep the models under their own control, and rebuild their workflows around them. At that scale it works. The price is the bench itself. By Built In’s data, US AI-engineer pay averages about $210,000 in total compensation, so even a two or three person team runs to several hundred thousand dollars a year before compute, and inside a large organization the rebuild still moves at the speed of sign-offs and budget rounds. For a mid-market firm, it is an expensive way to reach a rebuild it could have bought.

  2. The partners bring the lab inside. The second enterprise path is signed rather than staffed. OpenAI sells implementation directly through The Deployment Company, a consulting venture it controls that raised over $4 billion, and Anthropic launched an enterprise services firm with Blackstone, Hellman & Friedman and Goldman Sachs. The same work arrives packaged through the big consultancies, in alliances with the labs and Microsoft: PwC became OpenAI’s first ChatGPT Enterprise reseller and its largest enterprise customer. Either way, forward-deployed engineers arrive from outside and embed in your workflows. It moves faster than hiring a bench, and it places the rebuild, and everything the rebuild teaches, on someone else’s platform. What the embedded engineer carries back to the lab is the subject of a separate piece in this series.

Two paths that never touch the work

  1. The subscribers buy seats and stall. The third group buys the work a tool and hopes the tool will change it. Enterprise Claude, Copilot or ChatGPT seats go to an enthusiast or two, and the workflow underneath stays put. The clean test is the UK Department for Business and Trade’s, which gave about 1,000 Microsoft 365 Copilot licences from October to December 2024 and tracked the 300 who agreed to be measured. They averaged 1.14 Copilot actions a working day and 72% were satisfied, yet the evaluators found no robust evidence the time saved was reaching measurable productivity. People liked the tool. The work was not rebuilt around it, so the number did not move.

  2. The network members wait for head office. The fourth group cannot change its own work, because the AI stack is chosen above it. Franchise operators and member firms of international networks take what head office sends. KPMG committed $2 billion over five years to Microsoft cloud and AI across a network of 265,000 people, and member firms receive it on head office’s timetable, with no guarantee it fits their market. The smaller ones are absorbed by platforms with the capital to move: Baker Tilly and Moss Adams merged in June 2025 in a roughly $7 billion deal backed by the private-equity firm Hellman & Friedman, with capital earmarked for technology.

The first two paths pay to have the work rebuilt, with a bench or with a contract. The middle two report adoption while the work stands still: the subscriber buys tools around it, and the network member waits while head office decides. For a mid-market firm the first two are out of reach at full scale, and the middle two are where budgets go to be reported.

The fifth path

  1. The self-starters rebuild one workflow now. The fifth group does what the middle two avoid and the first two overpay for. Without a big budget, it picks one workflow and rebuilds it around a scoped system, bought rather than built. Mid-market firms are already moving: in RSM’s 2025 survey, 91% of the US middle market already used generative AI, but only 25% called it integrated into core operations, and that gap is the work of rebuilding. Buying also reaches the finish line more often than building: a preliminary, non-peer-reviewed 2025 report from MIT’s Project NANDA, whose method drew criticism, found bought AI reached deployment about 67% of the time, roughly twice the rate of in-house builds. Read with that caveat, for a firm without an AI bench, a scoped rebuild is the likelier route to a workflow that actually changed.

The case for waiting, and what it misses

There is a sixth choice, which is no path at all: wait. The commonest version inside companies is the modernization schedule, finish the CRM replacement or the data migration first, take up AI properly in two or three years on better models. The economics look supportive. The cost of running a GPT-3.5-level model through inference fell about 280 times between late 2022 and late 2024, from roughly $20 to about $0.07 per million tokens. Why rebuild on today’s model when next year’s is better and cheaper?

Because the price fall reaches everyone at once, it gives the waiter no advantage. What the waiter lacks is what the mover has been building all along: the rebuilt workflow, and the organizational time it took. Gartner clocked an average of eight months from prototype to production, with only 48% of projects arriving at all, and the budget and legal rounds around that time do not shrink because the model got cheaper. That time cannot be bought later, and the firm that starts in 2027 will not have it.

And the wait has a measured price. BCG’s September 2025 assessment placed about 60% of companies in a laggard cohort with minimal gains, against leaders growing revenue about 1.7 times as fast as the laggards, with 3.6 times their three-year shareholder return. That does not prove AI caused the gap, but the gap is measured. Chegg, at the top of this piece, was a company on a schedule, and the market did not wait for it.

Place yourself on the map

So before you call anyone, place yourself on the map. A bench of your own puts you in the first group, a lab or alliance contract in the second, seats in the third, a head-office timetable in the fourth. Only the fifth is a rebuilt workflow scoped to one firm, and a rebuilt workflow is easy to fake, so it is worth a hard test. An owner runs it on their firm, an advisory partner on a client, a diligence team on a target that credits AI with margin. Four specific answers mean a workflow was rebuilt. Four vague ones mean it was not.

  1. The workflow and its number. One workflow and the single number the rebuild was meant to move, in a sentence from the person who runs it. “Adopt AI” is not a workflow.
  2. The data and where it lives. The system runs where the client’s data already sits, and no outside model trains on it. Files leaving the building is a risk taken on purpose.
  3. The expert who checks it. A named person who knows the work reviews each output before anyone acts on it. Fluent and wrong is the failure that looks finished.
  4. The before and after. What the number read before the rebuild, and what it reads now. If nothing was measured beforehand, the rebuild is a story with no number behind it.

The scoreboard will keep moving. The United States counted 59 models to China’s 35 in 2025, and the next tally will be larger. None of it will do a day of work inside your company. That comes down to one workflow, and whether you rebuilt it around the model, hired someone to, or only bought a seat. Name the one workflow you would rebuild, and the number it should move, and you are on the fifth path.

The fifth path lives or dies on how tightly that one workflow is scoped, which is why an enthusiast with a chatbot cannot stand in for it. That is the subject of the next article, Why can’t you vibe-code a business solution?

If you are weighing a first rebuild, see what a scoped deployment looks like for your firm.

The short version

  • Adopting AI and rebuilding the work around it are two different acts, and only the second moves the profit-and-loss number. Of the five paths companies take in 2026, the first two rebuild the work at enterprise price, the middle two never touch it, and the fifth is the scoped rebuild priced for the mid-market.
  • The first enterprise path hires the bench: US AI engineers average about $210,000 in total compensation (Built In), and the rebuild still moves at the speed of sign-offs. The second signs the lab: OpenAI’s Deployment Company raised over $4 billion, Anthropic launched an enterprise services firm with Blackstone, Hellman & Friedman and Goldman Sachs, and PwC resells ChatGPT Enterprise. Forward-deployed engineers embed in the client’s workflows, and the rebuild lands on someone else’s platform.
  • The third path hands out seats. In the UK Department for Business and Trade’s late-2024 Copilot trial, 1,000 licences and 300 tracked users gave 1.14 actions a day and 72% satisfaction, yet no robust evidence of a productivity gain. The fourth waits for a head-office platform (KPMG committed $2 billion across 265,000 people).
  • The fifth rebuilds one workflow around a scoped, bought system. In RSM’s 2025 survey 91% of the US middle market already used generative AI but only 25% called it integrated into core operations, and a preliminary, non-peer-reviewed MIT Project NANDA report found bought AI reached deployment about twice as often as in-house builds, roughly 67% of the time.
  • Waiting is not a path: inference prices fell about 280 times from late 2022 to late 2024 (Stanford HAI), the same for everyone, while BCG measured laggards against leaders at about 1.7 times slower revenue growth, and Gartner clocked eight months from prototype to production with only 48% of projects arriving.

Sources

  1. Forbes. Chegg Stock Down 99%. October 29, 2025. https://www.forbes.com/sites/petercohan/2025/10/29/chegg-stock-down-99-learn-whether-ai-45-layoffs-make-chgg-a-buy/
  2. Stanford HAI. AI Index Report 2025. April 2025. Notable AI models by country in 2024, and inference-cost decline. https://hai.stanford.edu/ai-index/2025-ai-index-report
  3. Built In. AI Engineer Salary in the US. Data for 2025 and 2026. https://builtin.com/salaries/us/ai-engineer
  4. CIO Dive. OpenAI’s Deployment Company raises over $4B for AI consulting and integration. May 11, 2026. https://www.ciodive.com/news/openai-deployment-company-4-billion-ai-consulting-integration/819942/
  5. Blackstone. Anthropic Partners with Blackstone, Hellman & Friedman and Goldman Sachs to Launch Enterprise AI Services Firm. May 4, 2026. https://www.blackstone.com/news/press/anthropic-partners-with-blackstone-hellman-friedman-and-goldman-sachs-to-launch-enterprise-ai-services-firm/
  6. CNBC. PwC to become OpenAI’s first reseller and largest enterprise user. May 29, 2024. https://www.cnbc.com/2024/05/29/pwc-to-become-openais-first-reseller-and-largest-enterprise-user.html
  7. UK Department for Business and Trade. Microsoft 365 Copilot Evaluation. Published 2025, trial October to December 2024. https://assets.publishing.service.gov.uk/media/68adbe409e1cebdd2c96a19d/dbt-microsoft-365-copilot-evaluation.pdf
  8. Microsoft and KPMG. Agreement announcement. July 11, 2023. https://news.microsoft.com/source/2023/07/11/kpmg-and-microsoft-enter-landmark-agreement-to-put-ai-at-the-forefront-of-professional-services/
  9. Journal of Accountancy. Baker Tilly, Moss Adams merge to create 6th-largest US CPA firm. April 2025. https://www.journalofaccountancy.com/news/2025/apr/baker-tilly-moss-adams-merge-to-create-6th-largest-us-cpa-firm/
  10. RSM US. 2025 Middle Market AI Survey. June 2025. https://rsmus.com/newsroom/2025/middle-market-firms-rapidly-embracing-generative-ai-but-expertise-gaps-pose-risks-rsm-2025-ai-survey.html
  11. MIT Project NANDA. The GenAI Divide: State of AI in Business 2025 (preliminary, non-peer-reviewed). July 2025. https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf
  12. Gartner. Survey press release on generative AI deployment (48% from prototype to production, average eight months). May 7, 2024. https://www.gartner.com/en/newsroom/press-releases/2024-05-07-gartner-survey-finds-generative-ai-is-now-the-most-frequently-deployed-ai-solution-in-organizations
  13. BCG. AI Leaders Outpace Laggards in Revenue Growth and Cost Savings. September 30, 2025. https://www.bcg.com/press/30september2025-ai-leaders-outpace-laggards-revenue-growth-cost-savings
  14. Stanford HAI. AI Index Report 2026, Research and Development chapter. 2026. Notable AI models by country in 2025. https://hai.stanford.edu/ai-index/2026-ai-index-report/research-and-development