Agents Take Action:
AI Executes, You Decide

Google's most consequential enterprise update this week was Workspace Intelligence reaching full rollout across eligible accounts. The system gives Gemini real-time context drawn automatically from a user's Gmail, Google Chat, Calendar, and Drive, meaning AI-assisted tasks no longer require manual context-setting on each query. Admins can control which data sources are enabled at the domain, organisational unit, or group level, and the update is on by default for users with AI Expanded Access and AI Ultra Access licences. Separately, Google AI Pro and Ultra subscribers received increased usage limits in Google AI Studio, and Gemini in Chrome added a skills feature that lets eligible Workspace users save reusable prompts and execute them with a single click anywhere in the browser. The week also saw a pre-conference buildup toward Google I/O on May 19, with Google signalling that agentic AI, new Gemini model updates, and Android 17 will all feature prominently.

For enterprise leaders, the practical implication of Workspace Intelligence is a change in how Gemini performs relative to your expectations. The same prompts that previously returned generic outputs will now return responses grounded in your organisation's actual files, meetings, and communications. That shift has value, but it also introduces a new category of question for IT and compliance teams: what does Gemini now have access to, and are the default data source settings appropriate for your organisation's governance requirements? Google has provided admin controls, but the burden of configuring them falls on the organisation, not on Google. The May 19 I/O keynote is also worth scheduling time to watch; Google has consistently used I/O to announce changes that ripple through enterprise deployments within weeks.

OpenAI replaced GPT-5.3 Instant with GPT-5.5 Instant as the default model for all ChatGPT users on May 5. The new model reduces hallucination rates in sensitive domains including law, medicine, and finance, and introduces improved personalisation by drawing on past conversations, connected files, and linked Gmail accounts for Plus and Pro users. A companion update introduced memory source visibility, allowing users to see exactly which past interactions shaped a given response and to correct or delete any source they consider outdated. Alongside the model change, OpenAI also released GPT-5.2 Codex on the API for paid ChatGPT users and launched Advanced Account Security, an opt-in setting designed for users requiring stronger protection, including passkey-only sign-in and shortened active sessions. GPT-5.5 Thinking is now available to eligible paid plans.

The more strategically significant development for business leaders is OpenAI's decision to expand advertising in ChatGPT. The company launched a self-serve Ads Manager in beta and introduced cost-per-click bidding, while maintaining that ads are kept separate from answers and that conversation data is not shared with advertisers. For enterprise and EDU customers, a separate and significant update arrived: ChatGPT for Excel and Google Sheets is now globally available with a free preview running through June 2, after which usage follows existing credits. Enterprise leaders should note this free window and use it to evaluate whether the spreadsheet sidebar reduces meaningful friction in their analytical workflows before committing to continued usage.

Microsoft made three updates this week that collectively advance Copilot from a desktop-bound assistant to an ambient work layer. Copilot Cowork, the task-delegation product available to Frontier programme users, is now available on iOS and Android. The mobile release maintains full continuity with desktop sessions: a task started from a phone carries forward to a laptop without re-context, and tasks run in the cloud whether or not any device is open. Alongside mobile access, Cowork gained a plug-in skills system that allows organisations to package their own workflows, data connectors, and approval steps into reusable templates. Also of strategic note for enterprise leaders: Microsoft added Anthropic's Claude Opus 4.7 as an available model in Copilot Cowork, giving Frontier users the option to route complex reasoning tasks to Anthropic's most capable current model directly within the Microsoft 365 environment.

The other material update this week was the confirmation that Microsoft 365 Copilot Chat is now fully integrated across Teams chats, channels, calling, and meetings on Windows, Mac, and web, with mobile support arriving shortly. This is relevant for C-Suite leaders because it means Copilot is now present in every primary communication surface in the Microsoft 365 suite, not just as a sidebar but as a native participant. For enterprise leaders evaluating Copilot's return on licence investment, this ubiquity changes the adoption equation: employees are now one click away from AI assistance in the moments they are actually doing their most collaborative work. The Microsoft 365 E7 Frontier Suite, which bundles Copilot, Agent 365, and Microsoft Entra, became transactable in CSP channels at the start of May, giving procurement teams a consolidated licensing path for the full AI stack.

Anthropic's week was its busiest in months. Claude Opus 4.7 launched across all Claude products and major cloud platforms including Amazon Bedrock, Google Vertex AI, and Microsoft Foundry, at unchanged pricing of $5 per million input tokens and $25 per million output tokens. The new model brings substantially improved vision, better handling of long-running agentic tasks, and a 13% improvement on Anthropic's internal 93-task coding benchmark over Opus 4.6. Following a new compute agreement with SpaceX granting access to the Colossus 1 data centre and over 220,000 NVIDIA GPUs, Anthropic doubled Claude Code's five-hour rate limits for Pro, Max, Team, and Enterprise plans, and removed peak-hours restrictions entirely for Pro and Max users. The same compute capacity also raised API limits for Opus-class models. Anthropic's Code with Claude event on May 6 confirmed that API volume on the platform has grown 17 times year-on-year.

The most immediately actionable update for non-technical business leaders was the release of ten pre-built financial services agent templates. These cover pitchbook construction, KYC file screening, and month-end close processes, and ship as plug-ins for Claude Cowork and Claude Code, as well as cookbooks for Claude Managed Agents. Claude also launched Microsoft 365 add-ins for Excel, PowerPoint, and Word, with Outlook support arriving shortly, allowing context to carry automatically between applications without re-prompting. Separately, Claude Managed Agents gained three significant capabilities: multi-agent orchestration, a Dreaming memory system that reviews past agent sessions to identify patterns and improve future performance, and support for outcomes-based grading that allows organisations to define what a successful result looks like and measure agent performance against it.

Perplexity released an entirely new native Mac application on May 7, replacing its previous software with a redesigned interface built specifically for macOS. The new app introduces Personal Computer, a cloud-processed automation layer that operates across local files, native Mac applications, and the web. Processing shifts to Perplexity's servers rather than the device, which reduces strain on Mac hardware while enabling more complex automation. A universal Command Bar, triggered by pressing both Command keys simultaneously, gives users immediate access to AI-powered actions from anywhere in the system without switching applications. The original Mac app remains temporarily available but Perplexity confirmed it will be retired.

Separately, Perplexity's deep research capability now outputs directly to editable presentation slides, completing a research-to-deliverable workflow inside a single tool that previously required moving between research, synthesis, and presentation software. The Model Council feature, which allows users to route queries to a committee of AI models simultaneously, was extended with memory personalisation so that each model in the council can access relevant personal context to improve response quality. For enterprise and C-Suite leaders, the strategic relevance of Perplexity is less about any individual feature and more about its trajectory: the platform is building toward a research-to-action workflow that competes not only with ChatGPT and Gemini but with productivity suites that bundle research, writing, and presentation into a single environment.

Meta's Muse Spark model, released on April 8 following a year-long pause in Meta's public model releases, continued to receive post-launch review and adoption this week. The model represents a ground-up rebuild of Meta's AI stack, with native multimodal capabilities and multi-agent orchestration support. Early reviews have been largely positive relative to the poor reception given to Llama 4, though commentators note that the model sits behind the frontier set by Anthropic and OpenAI on several dimensions. More significant for business leaders is the strategic context: Meta has formally closed its open-source Llama programme. Yann LeCun, Meta's former chief AI scientist and a long-standing advocate for open models, publicly acknowledged that Llama 4 benchmarks had been overstated before his departure, and Meta's shift to a proprietary model family under the Muse brand signals a commercial pivot toward protecting its AI investments rather than distributing them freely.

For businesses that have built internal tools, workflows, or deployments on the assumption that Meta would continue releasing open Llama weights, this is a significant change in the dependency landscape. Manus, the autonomous agent platform acquired by Meta in 2025, continues to operate as both an independent service and an integration target for Meta's broader AI suite. Manus agents can complete complex, multi-step tasks including analytics, research, and web-based execution, and the platform has seen continued enterprise adoption particularly in use cases where unattended, long-running automation has commercial value. For leaders evaluating autonomous agents, Manus represents the most mature example in Meta's ecosystem of what AI execution rather than AI assistance looks like at scale.

xAI launched Connectors for Grok on May 6, enabling direct two-way integration with Google Workspace, Microsoft SharePoint, OneDrive, Notion, GitHub, and Linear. Users can now read and draft emails, update documents, analyse spreadsheets, manage calendar events, search codebases, and track project issues without leaving the Grok interface. A Bring Your Own MCP option extends Connectors to proprietary internal systems for enterprise users. The launch is available on Grok Web, iOS, and Android simultaneously. For leaders who are active daily users of X, this update is immediately visible: Grok, which is integrated into X's interface, can now access and act on your connected work tools from the same surface where you consume news, monitor competitors, and communicate publicly. The reach of the assistant has expanded from conversation to execution.

The same week saw xAI confirm that Grok 4.3, its current flagship model, is now fully available via the API. The model offers a one-million-token context window, three adjustable reasoning intensity levels, and pricing of $1.25 per million input tokens and $2.50 per million output tokens. Eight legacy models from the Grok 4 generation will be retired on May 15, with a tight nine-day migration window for any business using those models in production. For enterprise technology leaders, the Grok 4.3 pricing is worth benchmarking directly: at $1.25 per million input tokens, it positions frontier-class reasoning at a cost level that makes higher-volume use cases commercially viable. The combination of competitive pricing, connector integrations, and the X platform distribution network makes Grok a more credible enterprise consideration than it was ninety days ago.

Mistral launched Mistral Medium 3.5, a 128-billion-parameter dense model released under a Modified MIT licence, in the days leading into this week. The model consolidates what were previously three separate tools, Magistral for reasoning, Devstral for coding, and the previous Medium for general use, into a single system with configurable reasoning effort per request. The same model that handles a quick chat reply can run a complex long-horizon agentic coding task when reasoning effort is set to high, without any application-level switching. Medium 3.5 scores 77.6% on SWE-Bench Verified, ahead of comparable alternatives, and is now the default model in both Le Chat and the Vibe coding CLI. For enterprise buyers, the practical benefit is procurement simplicity: one model, one deployment target, and one pricing relationship replaces a three-tool stack.

Remote agents in Vibe, also launched this week, shift coding sessions from local machines to isolated cloud sandboxes. A developer or operations team member initiates a task, and the agent works through it asynchronously in the cloud, opening a pull request on GitHub when complete and sending a notification to the user rather than requiring them to monitor every step. This is Mistral's answer to the unattended agentic coding model that Anthropic and OpenAI are also building toward, and it is arriving from an open-weight European provider for organisations with data residency or sovereignty considerations. For the C-Suite audience, the strategic frame here is not the coding capability itself but what it signals: if AI agents can complete software tasks unattended in a compliant, auditable cloud environment, the same pattern will extend to other knowledge-work domains within the next two to three release cycles.

DeepSeek's V4 Pro preview, which entered limited access in late April, is attracting serious enterprise attention this week following evaluations showing benchmark performance on agentic tasks that sits alongside GPT-5.5 and Claude Opus 4.7. The model is open-weight, offers a one-million-token context window, and is available at approximately $1.10 per million output tokens. For organisations running high-volume, lower-sensitivity inference workloads, the economics are materially different from those of comparable closed-model alternatives. V4 Pro achieves this through a mixture-of-experts architecture that activates only a portion of parameters per task, maintaining frontier-level output quality while significantly reducing the compute cost of each inference call.

The strategic implication for enterprise AI budgets is straightforward: DeepSeek V4 Pro makes a strong case for multi-model AI architectures where the choice of model per task is determined by sensitivity requirements and cost thresholds rather than by default vendor relationships. Workloads involving large-scale document processing, structured data extraction, code generation at high volume, or research synthesis that does not involve confidential data are all plausible candidates for V4 Pro routing. The consideration for risk-conscious organisations remains data sovereignty and the provenance of a Chinese-origin model, which should be factored explicitly into any procurement evaluation rather than treated as a disqualifier without analysis.

Apple is generating pre-WWDC 2026 attention this week after reporting indicated that iOS 27 will introduce a fundamentally redesigned Siri with persistent conversational memory, on-screen awareness, and the ability to take actions across apps. The redesign is powered by Google Gemini under a multi-year agreement reportedly valued at approximately $1 billion annually, with Apple's next-generation Foundation Models running on Gemini's infrastructure. The new Siri is expected to be previewed at WWDC 2026 and to feature a standalone chat-style interface distinct from the current voice-command model. Apple also confirmed it is developing a model-swapping architecture that would eventually allow third-party AI models, including Claude, to power specific Siri features, giving users a degree of control over which AI handles which request.

For enterprise leaders, this development is relevant across two dimensions. First, hundreds of millions of iPhone and Mac users, including most of your employees, will gain a materially more capable AI assistant on devices they already own. If the Siri overhaul delivers on its preview claims, it changes the baseline AI capability available to any knowledge worker without additional subscription cost. Second, the Apple-Google agreement on AI infrastructure represents a pattern worth noting: even Apple, which has historically prioritised on-device processing and data privacy, has concluded that cloud-based AI partnerships are necessary to remain competitive. Leaders who have deferred AI adoption partly on privacy grounds should recognise that the largest privacy-conscious technology company has reached a different conclusion about the trade-off.