AI-Proof - Weekly AI Pulse
A concise summary of the week’s most important AI developments
Executive Summary
Two themes emerged this week.
First, AI oversight is back on the agenda in Washington. After dismantling the previous administration’s AI safety framework earlier this year, the White House is now reportedly drafting an executive order that would require pre-release government vetting of frontier AI models. Microsoft, Google DeepMind and xAI also signed pre-deployment evaluation agreements with the Commerce Department’s Center for AI Standards and Innovation (CAISI) on 5 May, joining OpenAI and Anthropic. The direction of travel is clear: the most capable models are no longer being treated as ordinary software releases. They are becoming strategic infrastructure.
Second, evidence continues to grow, showing that AI adoption is moving from experimentation into operational redesign. Coinbase confirmed a workforce reduction on 5 May and directly linked the move to AI deployment across customer service, fraud and compliance. PayPal has also announced a multi-year restructuring programme alongside the creation of an internal AI Transformation Group. The important point for business leaders is not that AI automatically means layoffs. It is that investors and management teams are now starting to connect AI deployment with measurable changes to cost base, service delivery and operating leverage.
For businesses, the practical takeaway is this: AI is no longer just a productivity tool sitting at the edge of the organisation. It is beginning to reshape workflows, team structures and expectations around speed, cost and responsiveness. The companies that benefit will not be the ones that simply buy more tools. They will be the ones that understand where AI can genuinely improve repeatable work, where human judgement still matters, and where the operating model needs to change around the technology.
What to Try This Week
What to Try This Week
Run a 30-minute workflow audit with your senior team.
Pick two or three repeatable workflows where your business already spends significant time, money or management attention. Avoid the obvious AI use cases such as basic content generation. Look instead at areas such as compliance review, contract analysis, financial reporting, account research, sales operations, regulatory monitoring, customer onboarding or management reporting.
For each workflow, answer four questions:
What work is genuinely repeatable?
What data, documents or systems does the work depend on?
Where does human judgement add real value?
What could AI realistically improve today: speed, cost, accuracy, consistency, customer experience or management visibility?
The goal is not to create a redundancy plan. It is to build a clearer view of where AI could improve the business in practical terms, and where the organisation is not yet ready because the data, process or governance is too weak.
This is the useful lesson from this week’s news. AI is starting to show up in operating models, not just product demos. Businesses should respond by mapping the workflows where the technology could create measurable value, then deciding which ones are worth testing properly.
Trump Administration Reverses Course on AI Oversight
The White House is drafting an executive order for mandatory pre-release government vetting of frontier AI models, a sharp reversal of its January position. The proposal mirrors the UK’s pre-deployment safety regime and was triggered specifically by Anthropic’s Claude Mythos cyber capabilities. No timeline has been published. Industry reaction is split between welcoming the policy clarity and concern about chilling effects on US frontier development. Sources: New York Times (4 May), Wall Street Journal, Politico, CIO.
Pentagon Signs Classified-Network AI Deals with Seven Firms; Anthropic Excluded
On 1 May, the Department of Defense confirmed agreements with SpaceX, OpenAI, Google, Microsoft, AWS, Nvidia and startup Reflection AI to deploy LLMs on classified networks for “any lawful purpose.” Anthropic was excluded after refusing the unrestricted-use clause. The fine print allows the government to require adjustments to safety filters, with no contractual veto for the vendor. Constitutional and congressional scrutiny appears likely. Sources: New York Times, Associated Press, Tom’s Hardware.
NYT: Federal Safety Net “Not Ready” for AI Job Displacement
On 5 May, the New York Times argued that economists across the political spectrum agree the US federal safety net is structurally unprepared for AI-driven displacement at the scale now measurable. The piece cited Goldman Sachs research showing net 16,000 US jobs lost per month to AI substitution, with Gen Z and entry-level workers bearing the brunt. The widening entry-level wage gap is now empirically quantified, not speculative. Source: New York Times.
AMD Q1 Earnings Reset the AI Infrastructure Trade
On 5 May, AMD reported Q1 revenue of $10.3 billion (up 38% year on year), with data centre at $5.8 billion (up 57%). Q2 guidance of $11.2 billion came in well above the $10.5 billion consensus. The stock rose 13 to 16% after hours and triggered a sector-wide rally in Micron, Intel and Arista. Morgan Stanley revised its price target up the same day. The print broadens the AI infrastructure thesis beyond Nvidia in a meaningful way. Sources: Reuters, CNBC, AMD IR, EBC.
Coinbase and PayPal Cut Workforces, Cite AI Directly
On 5 May, Coinbase confirmed approximately 700 layoffs (14% of headcount), explicitly attributing the cuts to AI-driven consolidation across customer service, fraud detection and compliance. The stock rose 3.5% on the announcement. PayPal CEO Enrique Lores announced a 20% reduction over two to three years and stood up an internal AI Transformation Group led by former Walmart technology executive Anshu Bhardwaj. Investors are now treating AI-led restructuring as a margin event. Sources: New York Times, Fortune, SFGate, Bloomberg, HeyGoTrade.
Sierra Raises $950M; Cognition in Talks at $25B
Bret Taylor’s enterprise AI customer-service startup Sierra closed a $950 million Series E on 4 May, led by Tiger Global and Google’s GV with Benchmark, Sequoia and Greenoaks participating. Valuation: $15.8 billion. Separately, on 6 May Bloomberg reported that Cognition AI (maker of Devin) is in early talks at a $25 billion valuation, 2.5 times its September 2025 mark. Agentic enterprise software is now in price-discovery mode. Sources: TechCrunch, CNBC, Bloomberg, Techzine.
Anthropic Finalises Approximately $1.5B Wall Street Joint Venture
On 4 May, Anthropic confirmed a joint venture backed by Blackstone, Hellman & Friedman, Goldman Sachs and General Atlantic to embed Claude across private equity portfolio companies. The structure mirrors Palantir’s forward-deployed engineer model and positions Claude as enterprise infrastructure for a substantial share of the global buyout market. Reported total commitment is approximately $1.5 billion. This is a distribution play, not a product sale, and a significant strategic move. Sources: Reuters, Wall Street Journal, GuruFocus, Investing.com.
OpenAI Raises $4B+ for Enterprise “Deployment Company”
Within minutes of the Anthropic announcement on 4 May, OpenAI confirmed it had raised more than $4 billion from TPG, Brookfield Asset Management, Advent International, Bain Capital, Dragoneer and SoftBank for an enterprise deployment subsidiary valued at approximately $10 billion. The structure remains majority-controlled by OpenAI. The simultaneous announcements signal that enterprise distribution, not model capability, is the decisive battleground both labs now see for 2026. Sources: Times of India, Quartz, Financial Times.
Model & Platform Updates
Anthropic Lifts Rate Caps and Launches Financial Services Agents
Following the SpaceX compute deal, Anthropic announced on 6 May the removal of five-hour rate caps for Claude Pro, Max, Team and Enterprise users and lifted peak-hour limits on Claude Code for paid subscribers. On the same day, Anthropic unveiled a financial services agent suite covering KYC, compliance monitoring, trade documentation and portfolio analysis. This is an explicit play for Wall Street mandates and a direct competitive thrust against OpenAI and Bloomberg AI. Sources: Anthropic, Ars Technica, Bloomberg.
OpenAI Rolls Out GPT-5.5 Instant as ChatGPT’s New Default Model
On 5 May, OpenAI began rolling out GPT-5.5 Instant as ChatGPT’s default model for all users, replacing GPT-5.3 Instant. The update claims sharper, more concise answers, improved image and STEM performance, better web-search decisions and materially lower hallucination rates. More interesting commercially, personalization is being expanded through past chats, files and connected Gmail, with memory-source controls to show users what context shaped responses. Source: OpenAI.
Microsoft Brings Copilot Cowork to Mobile and Plugins
On 5 May, Microsoft expanded Copilot Cowork with iOS and Android access through the Microsoft 365 Copilot mobile app, plus native and custom plugins that package skills and business knowledge into reusable workflows. The update also adds Agent 365 integration for observability, security and governance, bringing Cowork closer to enterprise-grade workflow infrastructure. Important caveat: Cowork remains a Frontier preview feature, so availability and capabilities may change. Source: Microsoft.
Mistral Ships Medium 3.5 and Le Chat Agentic “Work” Mode
On 3 May, Mistral AI released a meaningful product slate: Mistral Medium 3.5 (claimed 77.6% on SWE-Bench Verified), a 128-billion-parameter flagship, async cloud-based coding sessions and an agentic Work mode in Le Chat. The product direction is squarely on production-grade agent orchestration and developer workflows, two areas where European labs have struggled to compete with US frontier players. The release reframes Mistral’s commercial positioning. Source: Mistral via AI news trackers.
xAI Releases Grok 4.3
On 1 to 2 May, xAI released Grok 4.3 with “always-on reasoning,” a one-million-token context window, lower API pricing and a Custom Voices voice cloning suite. Intelligence Index score reportedly improved by four points over Grok 4.20. The release is incremental rather than a step change, but the pricing move is the more interesting commercial signal as the cost curve continues to fall across all major frontier providers. Sources: VentureBeat, llm-stats.com.
Stripe Launches Link Wallets for AI Agents
On 2 to 3 May, Stripe introduced digital Link wallets that autonomous AI agents can use to make payments, alongside protocols built with Cloudflare enabling agents to autonomously create accounts, register domains and deploy applications. This is foundational economic infrastructure for agent-native commerce. For any business with online checkout, the relevant question is no longer whether AI agents will be customers, but whether your systems are ready for them. Source: Stripe.
Quick Hits
Greg Brockman’s stake disclosed at approximately $30B in Musk trial: OpenAI’s president took the stand on 4 May, confirming financial entanglements with Sam Altman and a stake worth approximately $30 billion. Cross-examination focused on board independence ahead of OpenAI’s anticipated IPO.
Musk admits xAI distilled OpenAI models: Under cross-examination on 4 May, Elon Musk admitted that xAI trained Grok by “distilling” OpenAI’s GPT models, a violation of OpenAI’s terms of service. Legal observers flagged it as a significant trial moment.
Cathie Wood’s ARK sells AMD post-earnings: On 5 May, ARK Investment Management sold AMD post-earnings and rotated into Shopify and GeneDx, a notable read on whether AMD’s near-term upside is now priced in.
Asia tech rally hits records on AI-led optimism: The MSCI Asia Pacific Index gained 2.3% on 4 May, the strongest single-day move since early April, with South Korea and Taiwan tech indices both up more than 4.5% on AI chip demand.
We work with leadership teams to move from experimentation to execution safely, commercially, and at speed. Talk to us.