A 15-minute paper guide · arXiv:2607.08920v1

AI adoption is not the same as using ChatGPT.

This paper asks whether large companies have truly rebuilt products and operations around AI—and whether the financial benefits are visible yet.

The short answer

Adoption is accelerating, but deep integration is still rare. Higher profits appear only at the mature end, while broad productivity gains have not shown up yet.

Start the guide
01

Start with the right question

Is AI being discussed—or actually used to run the business?

The researchers are not counting every employee who opens an AI assistant. They want evidence that AI performs tasks inside repeatable business processes.

That could mean ranking content, detecting fraud, sorting packages, serving customers, discovering drugs, or managing a supply chain.

Mental model: This is a study of organizational transformation, not chatbot popularity.
02

Understand the measurement

Five steps from AI talk to AI dependence

GPT-5-mini classified AI-related paragraphs from each annual filing using this rubric. The authors manually reviewed the results.

  1. 1

    No adoption

    Everest

    AI appears only as an industry, regulatory, or competitive risk.

    The company knows AI exists, but shows no evidence of using it.
  2. 2

    Exploring

    Best Buy

    The company is training people or building the capacity to use AI.

    It is preparing, but AI is not embedded in normal operations.
  3. 3

    Pilot

    Fiserv

    AI supports selected products or processes such as service or fraud detection.

    Some teams use it, but the business does not yet depend on it.
  4. 4

    Production

    FedEx

    AI operates in real workflows with an expected cost or revenue impact.

    It has moved beyond testing and now does real work every day.
  5. 5

    Deep integration

    Meta

    AI is central across products, operations, strategy, and financial performance.

    Removing AI would materially change how the company works and competes.
The dividing line: Scores 1–3 mean discussion, preparation, or limited use. Scores 4–5 mean production use or deep integration.
03

See the adoption gap

The AI boom is real—but narrower than it sounds

21% of firms reached production use or deep integration in 2025.

That includes 11% at the deepest level.

18% provided no evidence of current adoption.

Software & services70%deep integration
Advanced adopters⅔were technology firms
Most non-tech sectors3typically still in pilot

So “companies are adopting AI” hides two different realities: a small technology group is moving aggressively, while much of the rest of the economy is still testing.

04

Learn the central idea

Why AI can hurt before it helps

New technology rarely drops cleanly into an old organization. Companies pay the costs first and may receive the benefits later.

Costs arrive first

Data cleanup, infrastructure, model integration, training, workflow redesign, and failed experiments.

Benefits need maturity

Accumulated learning, automation, better allocation, and redesigned processes can eventually improve margins.

Buying the technology is an expense. Reorganizing around it is the investment.
05

Read the results correctly

What changed—and what did not

Profitability

A J-curve relationship

Early-stage adopters had lower margins, while deeply integrated firms had higher margins. The pattern was especially strong outside technology.

Productivity

No clear gain detected

Revenue per employee did not rise consistently with AI adoption. Task-level improvements may not yet have changed the whole company.

Capital spending

No broad relationship

Most firms rent AI through cloud services and APIs. The huge infrastructure bills belong mainly to a few technology giants.

Employment

No broad contraction yet

Total headcount did not show mass shrinkage. Job types may still be changing beneath the company-wide totals.

Market valuation

Higher among tech adopters

Technology firms with more advanced adoption tended to have higher Tobin’s Q, a forward-looking valuation measure.

Why can task-level gains disappear?
AI makes one step fasterAnother team still waitsCompany output barely changes
06

Keep the most important warning

Correlation is not causation

Story A

AI improves the company

Integration reduces costs, improves decisions, and eventually raises profit.

or
Story B

Strong companies adopt AI

Well-managed, innovative, well-funded firms can adopt earlier and more deeply.

The paper cannot fully separate these stories. Its results describe relationships, not proof that AI caused the outcomes.

07

Make the paper yours

Can you explain it now?

Try answering before revealing each explanation.

Your one-minute explanation

“The paper uses company filings to separate AI talk from real operational adoption. By 2025, advanced adoption was growing quickly but remained concentrated in technology. Profitability showed a J-curve: early integration was costly, while deep integration was associated with higher margins. The researchers did not yet find broad productivity, capital-spending, or employment effects—and none of the relationships prove causation.”