← Back to Archive · Homepage - Newsletter

Web archive · this issue of MCC AI Weekly was sent by email on Thursday, September 17, 2026.

MCC AI Weekly · Thursday, 17 September 2026

The value gap is a people gap, not a tech gap

This week's evidence points the same direction: the jump in AI results comes from redesigning roles, skills and governance around the technology, not from the technology itself. The IMF, BCG and workforce data below all quantify that gap, while the news roundup covers where compute and capacity are being committed to close it.

Insights

AI will reshape more US jobs than it replaces

BCG's Henderson Institute analyzes 165 million US jobs across 1,500 roles and finds that 50% to 55% will be substantially reshaped by AI within two to three years. Only 10% to 15% face elimination within roughly five years. Most roles persist, the report finds, but see significant change in daily tasks, required skills, and performance expectations.

So what

Leaders should focus workforce planning on redesigning tasks and skills within existing roles, not on headcount reduction. Cutting staff ahead of this reshaping risks losing institutional knowledge the work still requires.

Boston Consulting Group →

One in 10 advanced-economy jobs now demands a skill that didn't exist

The IMF analyzed millions of online job vacancies and found roughly 10% of postings in advanced economies, and 5% in emerging markets, now require at least one newly emerged skill. These skills cluster in IT, professional, technical and managerial occupations. The IMF ties this shift to AI-driven change, which it estimates touches nearly 40% of jobs worldwide.

So what

The IMF recommends reorienting education, reskilling programmes and labor-market policy toward these newly emerging, AI-complementary skills. Firms should treat vacancy data on new-skill requirements as a signal for where training pipelines are most urgently needed, particularly in IT and technical roles.

International Monetary Fund →

BCG's 10-20-70 rule for AI value

BCG finds that companies closing the AI value gap follow a 10-20-70 pattern: 10% of value comes from algorithms, 20% from technology and data, and 70% from people, process, and operating-model change. The firm links this gap to treating AI as a tools rollout instead of a business transformation. Workflow redesign, governance, and human-AI collaboration drive most of the impact.

So what

Fund AI programs as operating-model transformations, not technology purchases, matching investment in workflow redesign and governance to spending on algorithms and infrastructure. Treat people and process change as the primary lever for closing the value gap, not a secondary add-on to the technology build-out.

Boston Consulting Group →

Operational autonomy in AI agents is not accountability

MIT Sloan and BCG's expert panel finds 72% agree that governing AI agents as autonomous decision makers will fail. Panelists argue agents act within limits set by others, and, as Stanford's Riyanka Roy Choudhury notes, hold no assets to attach and no license to suspend. In Moffatt v. Air Canada, a tribunal rejected the airline's claim that its chatbot was a separate accountable entity.

So what

Calibrate agent autonomy to the stakes of each decision, not to what the agent is technically capable of doing. Base delegation on how reversible an action is and keep accountability with the developers, deployers and institutions that profit from the agent's use.

MIT Sloan Management Review / BCG →

IT executives report AI results jumped from 25% to 95% in a year

Azhar polled 250 IT executives in Las Vegas, asking how many had serious, meaningful results from AI initiatives. About a quarter raised their hands a year ago; this year nearly all did, roughly 95%, and every respondent plans to spend more next year. Azhar notes experiences varied widely, from a century-old institution moving entirely to open-weight models to a hospital mixing OpenAI and Anthropic systems.

So what

Executives still treating AI deployment as pilot-stage should note that peers across varied approaches, from open-weight models to mixed-vendor stacks, now report measurable results. The relevant question shifts from whether to deploy to which combination of models and workflows fits the organization.

Exponential View (Azeem Azhar) →

Weekly Short News

Wistia: LinkedIn overtakes YouTube as B2B's top video channel

Wistia's sixth annual State of Video report, based on analysis of over 13 million videos, finds video demand rising while nearly half of teams keep budgets flat. Eight in ten teams now name LinkedIn their primary video channel, ahead of YouTube. Over one-third of teams report using AI, most commonly in pre-production.

Wistia →

AI speech models outperform clinicians in schizophrenia studies

Researchers, including teams in the Netherlands and at the Feinstein Institutes, built AI models analyzing speech recordings to distinguish schizophrenia patients from healthy controls. One model reached 86.2 percent accuracy; another, by psychiatrist Sunny Tang, reached 87 percent versus 68 percent for human clinical raters. Researchers cite small, non-diverse training data and a gap between study results and real-world performance as unresolved limitations.

Scientific American / Knowable Magazine →

OpenAI's agent swarm produces unreadable Navier-Stokes proof

OpenAI deployed 10,000 agents using an unreleased model to address the Navier-Stokes problem, a 200-year-old set of equations. The effort generated 2.7 million messages and 130 billion tokens over 88 hours. The resulting proof runs more than 500 pages and, according to mathematician Terence Tao, is not intelligible to any human.

Exponential View →

Microsoft to grow AI-serving capacity from 2GW to 13GW by 2032

Bloomberg reports Microsoft plans to expand its AI-serving capacity from about 2GW today to nearly 13GW by 2032. The move is part of a broader datacenter fleet expansion from 12GW to 38GW, cited in Exponential View's newsletter.

Exponential View (citing Bloomberg) →

Anthropic's compute deals reach $517bn, per The Information

The Information compiled Anthropic's public compute agreements and reporting since October last year, totaling roughly $517 billion for 14.8GW of capacity. Google and AWS together account for 11GW of that total. The figure exceeds the $180 billion in server-rental spending through 2029 Anthropic previously told investors to expect. Other deals include agreements with Nscale, Fluidstack, CoreWeave, Akamai, Microsoft and AMD.

DataCenterDynamics, citing The Information →

AI Player of the Week

HeyGen

HeyGen turns a script, slide deck, or audio file into a video presented by an AI avatar modeled on a real person's voice, gestures, and expressions, then translates it into 175+ languages with matching lip-sync. Every custom avatar requires on-camera identity verification, and Coursera reports a 40% increase in video watch time and a 25% boost in course completion using it for training content. Marketing and L&D teams producing recurring video content at volume should shortlist it; teams needing one-off, high-stakes video where a human presenter's judgment matters should not.

HeyGen →

Tip of the Week

Ask the model to separate known, likely, assumed and missing

Prompted Psyche recommends asking a model to label which parts of its answer follow directly from the given information, which are a likely interpretation, which are assumptions, and which are missing entirely. It also warns against false precision in these labels. The point is to surface weak claims that fluent, confident-sounding prose would otherwise hide.

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