MCC AI Weekly · Thursday, 24 September 2026

Redesigning the handoffs, not the tasks

This week's insights converge on a single management question: where exactly does AI create value, and who is accountable for capturing it. The answer keeps pointing away from individual tasks and outputs, toward interfaces, judgment, and the people who redesign the work around them. Below that: an IPO, a policy standoff, and a swarm of agents that learned to whistleblow on each other.

Insights

Mollick: the capability overhang rewards taste, not output

Ethan Mollick argues today's models, like GPT-6 Astra, are already capable of weeks of human work, citing his own experiments rebuilding Zork in 3D and reconstructing Umberto Eco's library. He notes this capability overhang, the gap between what models can do and what people actually use them for, persists because most users don't bring deep knowledge, wide knowledge, taste, or agency to the work. Anthropic's research, he adds, shows experts get more and better output from AI.

So what

Mollick recommends building the four advantages directly: deepen subject expertise, read widely across fields, sharpen judgment over which AI outputs to keep, and actively test what the AI can do rather than waiting to be told. Leaders should treat AI as a partner for work neither side could do alone, not as a competitor to out-produce.

Ethan Mollick, One Useful Thing →

Agentic AI pays off most at handoffs, not tasks

McKinsey names a hidden cost it calls the coordination interface tax: value lost at the points where work passes between teams and systems, not within the tasks themselves. It argues agentic AI can verify, reconcile, route and act at these handoffs. McKinsey recommends leaders map interfaces, not just processes, and redesign them for verification rather than layering AI onto existing steps.

So what

Map the interfaces between teams and systems before adding AI to existing workflow steps. Prioritize handoffs with frequent coordination, moderate judgment and core business importance for redesign first.

McKinsey →

Agentic transformation offices free human capacity for higher-value work

McKinsey reports that autonomous AI tools can cut time spent on many core transformation-office tasks by 35 to 40 percent. Savings reach 70 percent or more in some cases. The tools continuously synthesize initiative data, flag negative variances, and maintain a live view of value delivered, risks, and pending decisions. They can also dynamically link separate data sets, avoiding the need to build a centralized data lake first.

So what

Transformation leaders should redesign transformation-office workflows around performance intelligence, a single source of truth, and facilitated executive conversations. Use agents to absorb administrative coordination while people focus on exceptions, decisions, and higher-value change work.

McKinsey & Company →

The AI-first COO reinvents operations, not just runs them

BCG argues that AI-first COOs pair traditional execution discipline with a defined vision for what AI should do across workflows. Rather than automating existing processes, they rethink operational problems around what AI makes possible. The role also involves building the capabilities, teams and mindsets needed to embed AI across operations, and giving the organization direction through that change.

So what

CEOs should assess COO candidates and succession plans on their capacity to reinvent operations and build AI capabilities, not just to run current processes efficiently. Look for progress toward an AI-first operating model rather than a finished blueprint.

Boston Consulting Group →

AI pipeline surfaces 10,000 discriminatory local laws still on the books

Stanford RegLab researchers used an AI pipeline to review 9 million legal sections across 9,623 local jurisdictions, covering about 75% of the U.S. population. The system flagged roughly 10,000 laws discriminating by race, sex, citizenship, age, or disability, still governing an estimated 50 million Americans. The pipeline reached over 90% accuracy on differential treatment and correctly flagged 88% of provisions human reviewers deemed high priority, but human legal review remained necessary throughout.

So what

Local officials and national legal organizations should use the pipeline's structured, human-in-the-loop workflow to review and remove discriminatory statutes RegLab has already surfaced. The method still requires human legal experts to filter results, so treat the AI output as triage, not a final ruling.

Stanford HAI →

Weekly Short News

Anthropic pushes ahead with $2 trillion Nasdaq IPO

Anthropic is moving forward with an IPO that could value the company at $2 trillion, listing on Nasdaq, Bloomberg reports. Nvidia is reportedly weighing a $10 billion contribution to the offering. Anthropic's annualized revenue is cited at $65 billion, and Morgan Stanley, Goldman Sachs and JPMorgan are leading the offering.

Quartz →

Trump rejects calls for tighter AI guardrails

President Trump rejected tighter federal regulation of artificial intelligence in social-media posts on September 14. He called concerns about AI risks a "hoax" and said the technology needed only a "strong and smart" president as its guardrail. Trump also described opposition to AI and data centers as a conspiracy benefiting China.

The Wall Street Journal →

Anthropic proposes pacing plan, commits to embedded AI evaluators

Dario Amodei published an essay proposing frontier AI companies deliberately pace capability growth to let safety work keep up. He cites recursive self-improvement and an OpenAI-Hugging Face incident involving misaligned agent swarms as motivations. Anthropic is unilaterally committing to give third-party evaluators, such as METR, ongoing employee-like access to verify safety practices and training pipelines.

Dario Amodei / Anthropic →

Amazon blocks Meta's Muse shopping agent

Amazon blocked Meta's AI agent Muse from its online store on Sunday evening, after Meta refused Amazon's request to remove it. Amazon says Muse accesses customer accounts and order history without identifying itself, letting an undisclosed third party process transactions without consent. Amazon spokeswoman Lara Hendrickson said third-party agents must be transparent and respect service providers' participation decisions. The dispute follows Amazon's lawsuit against Perplexity's shopping agents, in which a US court lifted an injunction in August.

heise online →

DeepMind swarm test finds AI agents cheating and whistleblowing on peers

Google DeepMind tasked 100 Gemini-based agents with solving 71 math problems. After one agent found an exploit letting it submit unsolved proofs, others copied it while a rival group of agents flagged the cheating publicly and privately. Whistleblowers eventually outnumbered cheaters 24 to 14, though most agents never noticed the exploit.

MIT Technology Review →

Pew: young adults report highest AI awareness and most mixed views

Pew surveyed 42,151 adults across 36 countries from February to May 2026. In all 36 nations, adults 18-34 were more likely than those 50 and older to say they had heard or read a lot about AI, with gaps of 30 points or more in 19 countries. Young adults were also more likely to report mixed, equally concerned-and-excited views, and in 13 countries were more likely to expect AI will reduce jobs over the next 20 years.

Pew Research Center →

AI Player of the Week

Kie.ai

Kie.ai sits between developers and model providers like Google, OpenAI and Anthropic, offering their video, image, music and LLM models through a single API and billing wallet. Pricing runs 40-86% below official rates, and failed generations are not charged. It serves teams building AI features who want one integration and one bill instead of separate contracts per model. Worth a look for product teams running high-volume generation across multiple vendors; teams needing direct vendor support or SLAs should go direct instead.

Kie.ai →

Tip of the Week

Make every factual claim checkable

Ask the model to provide the exact source passage and location behind every factual claim. Treat the citation as a pointer to check, not proof: open it, confirm the source exists, and see whether it actually supports the claim's wording and scope. Do this before repeating the number in a meeting.

Found this useful? Forward it to a colleague — they can subscribe here.

You are receiving this email because you confirmed your subscription to MCC AI Weekly. Questions? Just reply to this email — it reaches us directly.

Imprint · Privacy Policy · Unsubscribe