This week, the AI industry stopped talking about potential and started presenting invoices. A landmark policy document proposed taxing the robots replacing workers. Q1 data confirmed what many feared: 80,000 tech jobs gone, half attributed to AI. America's oldest bank quietly deployed 20,000 digital workers with their own email accounts. And a secret model is reshaping what's possible — for the 40 organisations lucky enough to access it. The reckoning has arrived, and it's surprisingly legible.

It landed on April 6 without fanfare: a 13-page policy document from the world's most valuable AI company, titled "Industrial Policy for the Intelligence Age: Ideas to Keep People First." OpenAI was proposing, among other things, a robot tax.

What the document proposes: When an AI system displaces a human worker, a levy on the automated labour would flow into a nationally managed public wealth fund — seeded in part by AI companies — giving every American citizen a direct stake in AI-driven economic growth. Alongside this: pilots of a 32-hour, four-day workweek with no loss in pay, a shift in the tax burden from labour to capital, and automatic safety-net triggers that kick in when AI displacement metrics hit preset thresholds.

The candour is striking: OpenAI explicitly warns that AI-driven growth could hollow out the tax base funding Social Security, Medicaid, SNAP, and housing assistance as "corporate profits expand and reliance on labor income shrinks." A company projecting $25 billion in annualised revenue acknowledging it might be eating the social contract is a remarkable thing to put in writing.

The timing matters: OpenAI is heading toward an IPO. Congress is actively debating AI legislation. Releasing a progressive economic manifesto now is either genuine moral leadership or sophisticated regulatory pre-emption — and probably both. Either way, it has shifted the frame of the policy conversation.

Enterprise angle: Every organization deploying AI in workforce-heavy roles now has a policy signal to track. If a robot tax gets legislated, the cost calculation for automation shifts — not dramatically, but enough to factor into multi-year AI investment models. CFOs and legal teams should start scenario planning now.

On April 7, Anthropic quietly announced Claude Mythos Preview. No press release, no launch event, no TechCrunch splashdown. Just a whispered update to roughly 40 selected organizations through a private access programme called Project Glasswing.

The numbers justify the secrecy: Claude Mythos scores 93.9% on SWE-bench Verified — the benchmark for AI software engineering, where 80% was considered exceptional just months ago. It scores 97.6% on USAMO 2026, the elite mathematics olympiad. Both figures are well above any other publicly available model. GPT-5.4 Pro currently leads public benchmarks at a 92 composite score. Mythos isn't competing with that — it's in a different league.

What is Project Glasswing: Anthropic isn't saying much publicly. Selected organizations are under NDA but believed to include a mix of frontier research labs, strategic enterprise partners, and government contractors. Early access comes with obligations: structured feedback, safety evaluation, and use-case documentation.

The frontier is bifurcating: Anthropic is building a tiered market — a public frontier and an invitation-only frontier running ahead of it. The gap between what the top 40 organizations can do and what everyone else can access is now measurable in benchmark points, and it is significant.

Enterprise angle: If you're not in Project Glasswing, this is the moment to ensure your organization has deep relationships with its AI vendors and a clear path to early-access programmes. The competitive moat being built here isn't just about model quality — it is about the compounding advantage of learning to work with tomorrow's models today.

For two years, analysts projected the impact of AI on employment in carefully hedged future tenses. This week, the Q1 2026 data arrived and removed the hedging: 78,557 tech sector jobs were cut between January 1 and April 1, with nearly 50% — some 37,638 positions — directly or indirectly attributed to AI and automation. The United States absorbed more than three-quarters of the losses.

The companies doing the cutting weren't subtle: Amazon announced 16,000 corporate layoffs, explicitly citing a strategic shift toward AI-driven agentic workflows, with cuts targeting middle management and administrative roles. Atlassian cut 1,600 positions — roughly 10% of its global workforce — while simultaneously appointing two new AI-focused CTOs. The message encoded in those two simultaneous moves is unambiguous.

An important nuance the data doesn't capture: These cuts are happening ahead of the productivity gains, not after them. Cognizant's Chief AI Officer noted that "it will take another six months to a year before companies start seeing real productivity gains from AI." Many of these layoffs are bets on a future productivity curve, not responses to efficiency already achieved. The social cost arrives now; the economic benefit arrives later.

The constructive longer view: BCG's 2026 workforce study projects that while 92 million jobs might be eliminated by 2030, 170 million new roles will be created — a net gain of 78 million. IBM tripled its entry-level hiring this year while simultaneously deploying AI agents, arguing that AI oversight creates more work, not less. Both the difficult short-run reality and the optimistic long-run projection can be true at the same time.

Enterprise angle: Organizations that invest now in reskilling programmes, human-AI teaming frameworks, and transparent workforce communication will be the ones that attract talent in the AI era. Those that treat the workforce transition as an HR footnote to a technology project will face a trust deficit they cannot AI their way out of.

America's oldest bank just published the most detailed operating manual available for the AI-native organization — and it wasn't published in a whitepaper. It was deployed in production.

The scale is genuinely remarkable: BNY Mellon now has 20,000 employees actively building and operating agents across more than 125 live use cases on its Eliza 2.0 multi-agent orchestration platform. These are not chatbots. These are autonomous workers with their own system credentials, email accounts, and communication access — operating as recognized identities within the bank's enterprise systems.

What that looks like in practice: An agent might cross-reference internal databases, validate external regulatory updates, consolidate findings, and send a structured briefing to a human manager via Microsoft Teams — all without a human in the loop until the final step. The human's job shifts from doing the research to reviewing and acting on it.

The governance architecture is the real story: Tiered escalation protocols ensure agents operate autonomously within defined parameters, escalate to human judgment when they encounter ambiguity, and log every action for audit. In a regulated industry, this architecture is not optional — it is survival. And it's instructive for any compliance-sensitive enterprise.

What comes next: BNY is building Predictive Trade Analytics — agents that won't just identify settlement risks but will autonomously initiate remediation protocols to prevent trade failures before they occur. That is not automation of known tasks. That is AI taking initiative.

Enterprise angle: Three things stand out for enterprise CIOs: (1) the agent-credentialing model — giving agents their own identity within enterprise systems rather than operating under a generic service account; (2) the human escalation architecture — agents that know when to stop and ask; (3) the AI literacy programme — building agent-building capability throughout the workforce, not just within a centralized AI team. The third is the most underrated.

A new enterprise AI adoption survey delivered an arresting set of contradictions this week: 97% of organizations report deploying AI agents in the past year. 52% of employees are already using them. And yet — only 14% of enterprises have a clear AI strategy with defined goals. 71% have an "incomplete or developing" strategy. 79% report significant challenges, a double-digit increase from 2025. Perhaps most pointed: 54% of C-suite executives admit that AI adoption is "tearing their company apart."

The familiar failure mode: Demos that worked in a sandboxed environment, deployed into production workflows without adequate testing, generating outputs that humans feel compelled to verify anyway — which often makes the overall process slower, not faster. The gap between demo and production is where ROI goes to die.

The constructive read: The 14% who do have a clear strategy are almost certainly seeing the returns that justify everyone else's investment. The gap between the leaders and the followers isn't about access to models — everyone has access to the same frontier. It is about organizational discipline: defining success metrics before deployment, building feedback loops, and investing in the human side of the human-AI system.

Enterprise angle: The survey's uncomfortable implication is that most enterprise AI deployments are experiments masquerading as strategy. The companies pulling ahead aren't necessarily deploying more AI — they're deploying it more deliberately. The practical move: before your next AI initiative, insist on a definition of what "working" looks like, measurable within 90 days.

Every CIO in a workforce-heavy organization is now managing two transitions simultaneously: deploying AI agents fast enough to stay competitive, and navigating the human cost of doing so with enough care to stay trusted. This week drew that tension into sharp relief.

The week's data points are not unrelated: BNY Mellon's 20,000-agent deployment and the Q1 layoff numbers landed in the same news cycle. The organizations building the most impressive agentic deployments are often the same ones restructuring their workforces around them. OpenAI's economic manifesto didn't arrive by accident — it arrived because the macro signal is getting harder to ignore.

The data frames the strategic tension: Gartner's 2026 workforce research projects that enterprises with mature AI agent deployments will see a 30% reduction in routine task volume by 2027. But organizations investing in AI augmentation strategies — rather than pure automation — see 40% higher employee retention in AI-adjacent roles. The deployment model you choose isn't just a productivity decision. It is a talent decision.

The strategy gap is the real competitive moat: What this week's enterprise AI adoption paradox reveals is that the 14% of enterprises with a clear AI strategy are not operating with better models. They are operating with the same frontier models, but with more organizational clarity about what they are for. The advantage is not technological. It is managerial.

The practical framing for CIOs: The question is not just "how many agents can we deploy?" It is "what does our workforce look like in 18 months, and are the people in it equipped for it?" BNY's model — building agent capability throughout the workforce rather than concentrating it in a centralised AI team — is the most instructive playbook available. It treats AI literacy as an organizational capability, not a specialization. That approach scales. A centralized AI team does not.

That's your signal for the week of April 6–12, 2026. The invoices are arriving — in job numbers, in policy documents, in deployment blueprints. The leaders who read them carefully will write better ones.

Subscribe to receive the weekly updates directly to your mailbox.

See you next week.