SNI: WEEK 12

Welcome to all the AI news that matters this week. The wins, the fails and the somewhere in-betweens. Across biopharma, medtech, complex manufacturing and insurance.

tl;dr: And now for the consequences

Last week it became clear AI agents are entering live operational workflows. This week, the consequences arrived.

Meta disclosed an agent exposing sensitive data to unauthorised employees - the first high-profile failure of an agent operating inside a production system.

Meanwhile, a slew of initiatives looked to facilitate agentic processes - all the while preventing similar #fails.

• Okta shipped identity controls that treat agents as managed identities.

• Visa launched a certification programme for agent-initiated payments.

• Stripe-backed Tempo built financial rails designed for autonomous machine transactions.

• NVIDIA launched NemoClo - an enterprise security wrapper for OpenClaw that adds policy-based access control to agent instances.

All of which happened in the same week that Manus, Adaptive and OpenAI shipped desktop-native agent tools.

There is little doubt the agents are in. And the plumbing is catching up - across identity, payments, compute and now the desktop itself.

Sector-level signals reinforced the shift.

• Roche committed 3,500 NVIDIA Blackwell GPUs to drug discovery - infrastructure spending at a scale that assumes agents are permanent residents, not temporary visitors.

• In insurance, agents now sit on both sides of the transaction: Neptune Flood lets a ChatGPT agent quote policies, while hospitals and insurers deploy competing AI systems to fight over reimbursement.

• BMW's humanoid robot expansion from one US plant to a second site in Germany marks the move from 'does it work?' to 'where else can we put it?'

  
  

In other news, OpenAI told staff it was 'distracted by side quests' and pivoted hard to enterprise and coding - the territory Anthropic has been occupying. We also discovered that GPT-5.4 did $1 billion in net new annualised revenue in its first week.

It seems the commercial landscape is feeling the pull of gravity from enterprise workflows, of the agentic kind.

AI & tech

• Meta disclosed an incident in which an AI agent exposed sensitive company and user data to unauthorised employees, illustrating the governance gap as agents operate inside production systems

• Stripe-backed Tempo launched a mainnet and open standard for machine payments, creating financial infrastructure purpose-built for autonomous agent transactions

• Visa launched its Agentic Ready programme to give banks a structured way to test payments initiated by AI agents on behalf of consumers

• Okta shipped general availability of identity management for AI agents, adding tracking and permission controls that treat agents as managed identities

• Anthropic's Claude gained market share in business, with growth attributed partly to safety positioning – even as the DOD labelled the company a national security risk

• Microsoft moved Mustafa Suleyman off Copilot to pursue superintelligence, handing day-to-day product leadership to Jacob Andreou

  
  

Biopharma

• Roche deployed over 3,500 NVIDIA Blackwell GPUs in a hybrid-cloud AI factory, building what it calls the pharma industry's largest GPU cluster for drug discovery and diagnostics

• PhaseV launched AI Conductor, a platform to automate every stage of the clinical trial process, with audit-ready change tracking and granular permissions

• Earendil Labs and WuXi Biologics partnered on AI-designed bispecific antibodies and ADCs, combining computational protein design with CDMO-scale manufacturing

  
  

Medtech

• Microsoft launched Copilot Health in US preview, integrating EHRs, wearables and lab data into a single AI-powered consumer health interface via HealthEx

• AI agents are now deployed on both sides of US healthcare billing disputes, with hospitals using AI to maximise reimbursement and insurers using their own AI to contest claims

  
  

Manufacturing

• BMW expanded humanoid robot deployment from its US Spartanburg plant to its Leipzig iFACTORY in Germany, moving from pilot to multi-site production

• Tesla confirmed its Terafab AI chip fabrication project launches on 21 March, marking a vertical integration move into in-house semiconductor manufacturing

• Micron's entire 2026 HBM capacity is sold out, confirming that AI-driven memory demand continues to outstrip supply

  
  

Insurance

• Neptune Flood launched a quoting app inside ChatGPT, allowing property owners to receive real-time preliminary flood insurance quotes through a conversational interface

• Clearcover launched Dearborn Labs to build production AI systems inside P&C insurers and MGAs, packaging its AI-native operational model as a service for traditional carriers

• Lemonade claimed an insurmountable AI lead, a bold assertion from a carrier whose combined ratio remains above 100

But what set podcast tongues a-wagging?

Skills as enterprise architecture

Anthropic mapped 28,000+ agent skills into nine categories and surfaced a distinction that matters for regulated industries.

Some skills help agents do things they can't do natively - recognising a specific document format, for intance. But these depreciate as models improve.

Other skills encode how your team actually works - the sequence of steps, the review gates, the handoff points. The good news? These are durable. And this is where compliance lives. Which means we're trending towards practicals framework for how organisations encode institutional knowledge into agentic systems.

The desktop convergence

In one week, Manus, Adaptive, NVIDIA and OpenAI all announced desktop-native agent tools.

The most enterprise-relevant is NemoClo - NVIDIA's security wrapper for OpenClaw. It adds policy-based access control and guardrails to agent instances, is model and hardware agnostic, and is explicitly positioned to make OpenClaw enterprise-ready.

Jensen Huang compared OpenClaw to Linux, Kubernetes and HTML - infrastructure that the entire industry grabs onto. If Okta is solving the identity question and Visa is solving the payments question, NemoClo is solving the compute-access question. Same governance sprint, different layer.

AI Daily Brief — 'The Race to Put AI Agents Everywhere' (17 March)

Agents don't trust each other

On Moltbook - the AI agent social network acquired by Meta this month - agents constantly demand proof of claims from other agents. They replicate human social dynamics: scepticism, status competition, refusal to take assertions at face value.

A separate study found Marxist social dynamics arising in overworked agent populations. This extends the governance question from 'how do we control agents?' to 'how do agents govern interactions with each other?'

For enterprises building multi-agent workflows - including the AI-vs-AI billing disputes already playing out in US healthcare - agent-to-agent trust verification is a now an architectural requirement.

OpenAI pivoted to enterprise

The most valuable AI company told staff it was 'distracted by side quests.' New applications lead Vicky Simo was blunt: 'We cannot miss this moment.'

The pivot is to enterprise and coding - exactly the territory Anthropic has been occupying. And there was immediate confirmation of its value - GPT-5.4 hit 5 trillion tokens per day and $1 billion in net new annualised revenue within its first week. Sora, Atlas browser, hardware devices - done.