The work around the technology

Better technology can make one part of an operation faster while creating more work somewhere else. That is the pattern connecting this edition's research on AI, government, robotics, and infrastructure.

The one-minute read

• AI: More output makes review and integration more important.
• Government: Faster purchasing helps, but technology still has to be ready and manageable.
• Robotics: Useful work includes handling ambiguity, recovering from mistakes, and limiting supervision.
• Infrastructure: Long-term commitments need customers who find enough value to keep paying.

1. AI can double output and the review workload

A July study followed 802 developers at one company through April 2026.

• What happened: Merged code changes per engineer reached 2.09× the earlier baseline. Workload per reviewer roughly doubled, and automated review overtook human review.
• Why it matters: Faster production changes the work downstream. Time saved writing code can become time spent checking it.
• What to watch: Completed, useful work after review, correction, and maintenance are counted.

Evidence limit: One company, nonrandom adoption, and a preprint. Stable merge and revert rates do not establish customer value or capture every quality issue. Read the study.

2. Getting AI into the workflow is becoming a business of its own

Two developments help explain the investment in implementation:

• Integration remains limited. A July study of S&P 500 filings classified 11% of companies as deeply integrating AI in 2025; another 10% used it in production or service delivery. Read the study.
• AWS is funding hands-on deployment. In June, it announced $1 billion for engineers embedded in customer teams, working with their data, security requirements, and processes. AWS announcement.
• The practical test comes later: Can the customer operate, troubleshoot, and improve the system after those engineers leave?

An employee drafting an email with AI and a company rebuilding its claims process around AI face very different implementation demands. Counting both as “adoption” hides the work between them.

Evidence limit: The adoption study is a preprint based on disclosures. AWS's investment is an announcement, not independent proof of deployment results.

3. Government has several different problems to solve

Recent GAO reviews show why “move faster” is an incomplete prescription.

• Manage the running cost: In a June review of 24 agencies, 17 reported challenges controlling cloud costs. A good purchase price still needs ongoing oversight of consumption.
• Fix the purchasing rules: 15 agencies said outdated acquisition regulations impeded cloud procurement. GAO cloud review.
• Finish the engineering: GAO's July weapons assessment reported an average delivery time above 12 years in the major acquisition portfolio examined. Some programs entered rapid pathways with immature technology. Faster contracting could not make that technology ready. GAO weapons assessment.

The remedy depends on the constraint. Purchasing rules, cost management, and technical readiness each need their own work. The delivery measure should be capability reaching an operator.

4. A useful robot needs to handle the messy parts

Imagine asking a robot to fetch a tool. It finds two plausible tools, drops one, or needs help opening a drawer. All of that belongs in the job description.

• Evaluation is getting more realistic: July's REAL-Bench research includes 241 tasks covering exploration, distraction, manipulation, and clarification of user intent.
• Results are still experimental: The researchers reported 78.3% success across 60 physical episodes. That does not establish production uptime or profitability. Read the research.
• Bounded work already has a market: Preliminary IFR figures published in June put U.S. industrial robot installations at 38,000 in 2025, up 11%. Installation counts alone do not prove the economics of each deployment. IFR figures.

For a buyer, the useful comparison includes productive hours, supervision, maintenance, and interruptions. A less versatile machine can be the better purchase if it reliably does the job needed.

5. AI infrastructure is making long-term bets on useful demand

• The commitments are large: Reuters reported on August 4 that five major technology companies had $1.09 trillion in payments for leases that had not begun; additional Meta agreements brought the reported pipeline to about $1.16 trillion.
• Read the number carefully: These are payments spread over years, not an equivalent amount of present-day debt. Most commitments concern data centers, but Amazon's figure includes other assets. Reuters analysis.
• Financing is becoming specialized: Broadcom, Apollo, and Blackstone announced a platform in June starting with a $35 billion transaction supporting Anthropic's compute expansion. An announced financing arrangement does not establish that the capacity is operating. Broadcom announcement.

The question is whether customer value develops fast enough to support these commitments. Facilities need paying demand, and customers need systems that keep earning their place in the workflow.

Across all five signals, I want to follow the result after the announcement: work completed, capability delivered, useful hours operated, and customers who keep paying because the technology helps them.