All issues
Discovery Digest · June 22, 2026
Issue 04. The week AI became infrastructure
This week, enterprise AI stopped being a preview and became a billable line item. Microsoft made Copilot Cowork generally available on Anthropic's models; xAI released free Grok add-ins for Word and PowerPoint the same day, directly challenging that price before Cowork's first invoice. The US government issued its first export-control order against a commercial frontier AI model, suspending Fable 5 and Mythos 5 globally with no advance notice. Google began embedding Gemini inside Chrome on Android and Perplexity routed Deep Research across 20 AI models simultaneously. The pattern across all five updates: AI is infrastructure now, subject to the same commercial, competitive, and regulatory pressures as any other critical infrastructure.
01 · Microsoft
Microsoft Copilot Cowork reaches general availability, running on Anthropic models
- What
- Microsoft made Copilot Cowork generally available to all Microsoft 365 Copilot customers worldwide on 16 June 2026. At general availability, the system runs on Anthropic Opus 4.8 and Sonnet 4.6, with Microsoft's proprietary Cowork 1 model due shortly. Cowork executes complex, long-running, multi-tool tasks end-to-end and returns a completed result rather than a draft. Usage is billed separately via Copilot Credits at $0.01 per credit on a pay-as-you-go basis, with tenant, group, and user-level spending limits available to administrators.
- When
- Made generally available by Microsoft, Tue 16 June 2026, via the Microsoft 365 Blog.
- How it shifts discovery
- Over half of Fortune 500 companies used Cowork during the Frontier preview. General availability converts that trial population into a billable customer base and sets the enterprise agentic AI benchmark against which every competing platform will now be measured. The decision to run Cowork on Anthropic models rather than Microsoft's own GPT-backed Copilot stack signals that enterprise AI infrastructure is being assembled from best-available models per task, not from a single vendor's stack. For teams evaluating enterprise AI deployment, the metered billing model and admin-controlled spending limits lower the governance risk that held procurement back during preview.
- Questions to ask
- Have we mapped which internal workflows are the highest-value candidates for Cowork automation, and can we quantify the credit cost against the time saved?
- Given that Cowork runs on Anthropic models by default, do we have clarity on how proprietary data is handled across the model boundary, and is that covered in our Microsoft agreement?
- As usage-based billing takes effect, who owns the Cowork credit budget in our organisation, and what governance structure ensures spend stays within forecast?
- Sources
02 · xAI
xAI releases free Grok add-ins for Word and PowerPoint, challenging Copilot on price
- What
- xAI released Grok for PowerPoint on 16 June and Grok for Word on 18 June 2026, both available as free Microsoft 365 add-ins from the Microsoft Marketplace. The Word add-in reads the full document and generates or rewrites text from a single prompt, with real-time web search drawing from Brave, Bing, and x.ai's own index, including source attribution. The PowerPoint add-in converts research and text into slide decks and generates editable vector diagrams using Mermaid syntax. No token limits apply to either add-in during the 2026 calendar year.
- When
- Grok for PowerPoint released Tue 16 June 2026; Grok for Word released Thu 18 June 2026. Announced via x.ai/news.
- How it shifts discovery
- Microsoft Copilot costs $30 per user per month on top of an existing Microsoft 365 subscription. Grok for Word and PowerPoint are free to any Microsoft 365 user and use live web grounding that Copilot's document model does not offer by default. For organisations with large teams producing proposals, reports, or presentations, the zero-cost entry point removes the procurement decision. Copilot Cowork reached general availability the same day Grok for PowerPoint launched, making this a direct pricing challenge at the moment Microsoft was expecting to convert its preview users to paying customers.
- Questions to ask
- Does our organisation have a policy on which AI writing assistants employees can use inside Office, and does that policy account for live web search running against documents that contain confidential information?
- If teams adopt Grok for Word at no cost, what are the data handling implications for documents containing client or commercial data sent to x.ai's infrastructure?
- How does free AI-assisted document creation change our expectations for draft quality, review cycles, and the skills we prioritise in content-generating roles?
- Sources
- xAI. Introducing Grok for Word, x.ai/news, 18 June 2026
- xAI. Introducing Grok for PowerPoint, x.ai/news, 16 June 2026
- Windows News. xAI Unleashes Grok Add-ins for Word, Excel, and PowerPoint, June 2026
03 · Anthropic
US government suspends Claude Fable 5 and Mythos 5 globally under export-control directive
- What
- At 5:21 p.m. ET on 12 June 2026, Anthropic suspended all access to Claude Fable 5 and Mythos 5 following a legally binding US government export-control directive citing national security. The directive bars access for all foreign nationals, inside and outside the United States, including Anthropic's own foreign national employees. The order followed a US government demonstration of a jailbreak technique said to unlock Mythos 5 cybersecurity capabilities. Access to all other Claude models is unaffected; Anthropic stated it believes the jailbreak is narrow and is working to restore access.
- When
- Suspension effective Fri 12 June 2026. Anthropic statement published on anthropic.com and reported by Fortune on Sat 13 June 2026.
- How it shifts discovery
- This is the first instance of a US government export-control order applied to a commercially released frontier AI model. The immediate operational impact falls on any enterprise, developer, or team with Fable 5 in a production workload via the Anthropic API or claude.ai. The longer-term implication is structural: AI model access is now demonstrably subject to the same export-control framework that governs defence hardware and cryptographic technology. Teams that have built production workflows on a single frontier model now have direct evidence that access can be suspended globally within hours, with no advance notice.
- Questions to ask
- Do we have a documented fallback plan if the primary AI model in a production workflow becomes unavailable overnight, and have we tested how quickly we can switch?
- For AI systems handling sensitive or regulated data, have we reviewed whether a sudden model suspension would trigger contractual or compliance obligations with our own clients?
- How does this export-control precedent change the model-selection and vendor-lock criteria in our AI procurement evaluations?
- Sources
- Anthropic. Statement on the US government directive to suspend access to Fable 5 and Mythos 5, anthropic.com, 13 June 2026
- Fortune. Anthropic disables Fable and Mythos AI models after US government bars it from giving foreigners access, 13 June 2026
04 · Google Search
Google embeds Gemini in Chrome on Android, adding an AI layer above every web page
- What
- Google began rolling out Gemini in Chrome on Android in late June 2026 for US users on Android 12 or higher devices with at least 4GB of RAM and English-US as the device language. The integration embeds Gemini as a persistent browsing assistant that can summarise any page, answer questions about on-screen content without switching apps, and connect to Gmail, Calendar, and Keep for task completion. Auto Browse, which lets Gemini navigate and act across multiple pages autonomously, requires an AI Pro or Ultra subscription.
- When
- Rolling out to US users in late June 2026. Announced at Google I/O, May 2026. Reported by the Google Chrome blog and Engadget.
- How it shifts discovery
- Embedding Gemini inside Chrome on Android creates an AI layer between the user and every web page they visit. Before a user reads a brand's content, Gemini can summarise it. Before a user engages with a product page, Auto Browse can act on it. This differs from the AI Overviews challenge: it operates on any URL after the click, not only on Google Search results before it. Content that reads well to a human but lacks structure for AI summarisation may be represented inaccurately in the Gemini overlay, and brands currently have no visibility into how their pages are being summarised inside Chrome.
- Questions to ask
- Do we know how Gemini in Chrome summarises our key product and conversion pages, and are those summaries accurate and brand-aligned?
- For pages where we rely on on-page dwell time and engagement signals, how does a Gemini summary layer change our interpretation of post-click behaviour metrics?
- How does Gemini Auto Browse change the path to purchase for products or services that require multi-step journeys, and have we mapped what an agent does when it follows that journey on a user's behalf?
- Sources
05 · AI Search
Perplexity integrates Deep Research into Computer, routing tasks across 20 AI models
- What
- Perplexity integrated Deep Research into its Computer platform in June 2026, enabling the system to break complex research tasks into subtasks and route them simultaneously across more than 20 frontier AI models, including Claude Opus 4.6, Gemini, GPT-5.2, and Grok. Each subtask is assigned to the model best suited to it: Gemini handles deep research retrieval, GPT-5.2 covers long-context recall, Grok handles speed-sensitive tasks, and specialised models cover images and video. The system returns completed work-ready reports, slide decks, and dashboards.
- When
- Announced via the Perplexity research blog, Thu 11 June 2026, reported by MarkTechPost.
- How it shifts discovery
- Single-model AI research produces outputs bounded by one model's knowledge, reasoning logic, and retrieval design. Routing tasks simultaneously across 20 models means each subtask is handled by the model with the best fit for it. For teams using AI research to understand competitive positioning, consumer sentiment, or category intelligence, the quality ceiling has risen substantially. The output now reflects a synthesis no single model could produce alone, and research produced using single-model tools is no longer the same product as research produced using multi-model orchestration.
- Questions to ask
- Are we using AI research tools that route tasks across multiple models, or are we treating single-model outputs as definitive and making strategic decisions on that basis?
- As multi-model platforms deliver work-ready reports and decks, where does human analysis add the most value in our insights function, and have we redesigned those workflows to reflect that?
- If competitors are using multi-model research platforms and we are not, how does that gap affect the quality of competitive and consumer intelligence informing our strategy?
- Sources
Key takeaways
What to walk away with this week
Enterprise AI infrastructure is commercially live and metered. Add Copilot Cowork to budget planning conversations this quarter: the preview period is over and the benchmark for enterprise-grade agentic AI has been set.
Free AI in Office documents will change team behaviour faster than procurement policy can track. Build a governance position on AI writing tools in internal workflows now, before adoption makes policy harder to enforce.
The Fable 5 suspension is the first proof that frontier model access can be removed globally within hours. Any production workflow built on a single AI model needs a documented fallback, tested before it is needed.
Gemini in Chrome on Android means your content can be summarised before a user reads it. Audit your key pages for AI summarisation accuracy now, while post-click engagement patterns are still interpretable.
Multi-model research platforms are delivering outputs that no single model can match. Audit whether your competitive and consumer intelligence workflows are using multi-model tools, or still treating single-model outputs as the ceiling.