Here's the number that should change how your organization thinks about AI: more than 90% of Claude Cowork's enterprise usage isn't software development. It's knowledge work. Business operations. Content creation. The administrative load that accumulates in every department, every week, and never quite makes it onto anyone's formal job description.
Anthropic just expanded Claude Cowork to mobile and web — and that expansion matters far more than the headline suggests. This isn't a product update. It's an architecture shift in how enterprise AI works, and the implications for CIOs, CFOs, and COOs are worth paying attention to.
What Claude Cowork Actually Is
If you're not tracking Cowork closely, here's the context. Anthropic built Claude Code as an agentic AI system for software engineers — one that can reason across files, run commands, and execute multi-step development tasks. Cowork applies that same agentic model to general knowledge work.
You hand Claude Cowork a task and describe what "done" looks like. It works across your files, calendar, email, messaging apps, the web, and connected tools until the job is finished. The key difference from a standard AI chat interaction: Cowork continues working after the conversation ends. You don't need to stay engaged for progress to happen.
Until recently, that meant you needed a laptop open. The work stopped when you closed the lid.
That just changed.
Three Things That Changed This Week
Anthropic expanded Cowork to web (claude.ai), mobile (iOS and Android), and maintained full capability on desktop — which retains access to local files and browser control. Three specific things are different:
Background tasks now run without any device online. You can schedule a task for 6 AM Monday — say, building a client briefing from email threads, call transcripts, and recent news — and Claude executes it while you sleep. No laptop required. The briefing is waiting for your review over coffee.
Work follows you across devices. Start a complex task at your desk, check on it from your phone during a commute, pick up the completed output from any device. The session holds context regardless of where you access it.
Decisions reach you on mobile. When Claude reaches a decision point that requires human judgment, the question surfaces on your phone. You can redirect a draft mid-meeting and Claude continues on the corrected path. The key design principle here is deliberate: nothing ships until you've reviewed and approved it. The AI does the work; humans make the calls.
The 90% Number Is the Real Story
Anthropic's usage data is worth sitting with for a moment. The largest categories of Cowork usage aren't what most enterprise AI conversations focus on. Business operations and content creation together account for roughly half of all usage.
What does that look like in practice? Three examples from the announcement:
- Reconciling a quarter's spend and drafting the variance memo
- Turning a folder of contracts into a renewals tracker with risks flagged
- Building a client deck from call transcripts and pipeline data
These are tasks that every finance team, legal team, and sales team does repeatedly. They're not glamorous. They're not the AI transformation story that gets told at board meetings. But they represent an enormous amount of hours across every department in every enterprise — and they're now candidates for delegation to a background AI agent.
This connects to a data point IBM released alongside its own agentic platform update this week: 85% of DevSecOps professionals surveyed say AI has already shifted the bottleneck in software development from writing code to reviewing and validating it. The same pattern is emerging in knowledge work broadly. AI can draft the variance memo. The CFO still needs to sign off.
What This Means by Role
Different members of the leadership team should be asking different questions.
For the CTO and CIO: The mobile and web expansion creates new data access and security questions. Desktop Cowork has always accessed local files and browser sessions — environments your security team could model and control. Mobile sessions introduce different threat surfaces: what data is the agent accessing when it runs in the background? What happens to session context on a lost or compromised device? These aren't blocking issues, but they need to be in scope when your enterprise deployment conversation happens.
Cowork is available on paid plans — Pro, Max, Team, and Enterprise. Enterprise-tier deployment gives you the administrative controls (model-level entitlements, usage monitoring, spend alerts) that a serious rollout requires. That's the right starting point for any CIO who wants to move beyond individual productivity gains to department-level deployment.
For the CFO: Anthropic is offering doubled Cowork usage limits through August 5 as a launch promotion. That's a signal worth reading: the company wants enterprises to try larger, more complex tasks without hitting constraints during evaluation. From a budget perspective, this is also the moment to establish governance before usage patterns scale.
The $500 million AI invoice story from earlier this year — where an enterprise client ran up that bill in a single month because they hadn't implemented spending controls — is the cautionary tale every finance leader should have in mind. Cowork isn't API consumption at that scale, but the principle holds: if your organization rolls out access to background AI agents without usage visibility, costs can compound in ways that don't surface until the invoice arrives. Spend alerts and model-level entitlements exist in the Enterprise tier precisely because this is a real risk at scale.
For the COO: This is where the strategic question gets interesting. If AI agents can now execute multi-step knowledge work tasks in the background — across email, documents, calendars, and web sources — what does that change about how operational work gets structured?
The honest answer is: we don't know yet, at the level of detail that operational leaders need. What we do know is that the tasks most likely to be meaningfully delegated share a structure. They involve pulling information from multiple sources, applying a consistent template or framework, and producing a draft output for human review. Variance memos, renewals trackers, briefing docs — these fit the pattern. Tasks requiring judgment calls, stakeholder relationships, or context that doesn't exist in any document don't.
The COO question isn't "what can AI do?" It's "what currently takes hours in my teams that fits this structure?" Starting with one department and one class of repeating tasks is the right approach. Trying to delegate everything at once is how you end up with a fleet of agents and no accountability for output quality.
The Governance Layer That Most Deployments Skip
Cowork's design reflects a specific philosophy: the AI does the execution, humans make the decisions. That's the right default for this phase of enterprise adoption. But the governance layer around "which decisions reach the human?" matters a great deal.
Consider the client deck scenario. Claude builds the deck from call transcripts and pipeline data. A salesperson reviews it. What are the review expectations? If the reviewer assumes Claude got the numbers right and skims the content, errors in the source data propagate into client-facing material. If the reviewer is expected to verify every figure independently, the time savings evaporate.
This is an organizational process question, not an AI capability question. The enterprises that get the most value from agents like Cowork are the ones that explicitly define what "review" means for each task type — and train their teams accordingly. That takes longer than deploying the tool. It's also more important.
The parallel to cloud infrastructure is instructive here. The enterprises that avoided the horror stories of uncontrolled cloud spend didn't just set up billing alerts — they created governance processes, defined ownership of cost centers, and established review cadences. AI agent deployment needs the same treatment.
What's Next for Enterprise Knowledge Work Automation
The mobile expansion of Cowork is one data point in a broader pattern. IBM's Bob platform added multi-agent capabilities and built-in cost analytics this week. OpenAI released GPT-Live targeting real-time interaction use cases. The competition for enterprise knowledge work is intensifying across every major AI provider.
What that means for CIOs and CTOs evaluating these platforms: the differentiation between providers is increasingly about governance, integration depth, and organizational fit — not raw capability. Most of these systems can now draft the variance memo. The question is which one fits your enterprise's existing tool stack, data governance requirements, and user behavior patterns.
Anthropic's enterprise positioning is built around the controls layer: model-level entitlements, spend alerts, admin analytics, and the human-in-the-loop design philosophy baked into Cowork. That's a coherent story for risk-conscious enterprises, particularly those in regulated industries where every output carries compliance exposure.
The background scheduling capability is probably the most underappreciated feature in this launch. The ability to define when AI work happens — and have it complete without human supervision during the execution phase — is what turns an AI assistant into something closer to a delegated workflow. When a legal team can schedule contract analysis to run overnight and have a risk-flagged renewals tracker ready for review by 8 AM, the time economics of AI adoption change meaningfully.
The Honest Assessment
Claude Cowork going mobile isn't a transformative moment. It's a maturation moment. The move from desktop-only to mobile-plus-web expands where enterprise AI can operate, reduces the friction of monitoring delegated work, and enables the background scheduling that makes agents genuinely useful for operational workflows.
The 90% knowledge-work usage number is the more meaningful signal. It tells you that when enterprises give their teams access to an agentic AI system and let them decide what to use it for, they don't reach for coding tasks. They reach for the grinding, document-heavy, multi-source coordination work that makes up so much of white-collar operations.
That's not a surprise if you've spent time in enterprise environments. It's validation that the market for AI in the workplace is much larger than the developer productivity story suggests — and that the next phase of enterprise AI value isn't going to come from engineering teams. It's going to come from finance, legal, sales, and operations teams automating the invisible work that fills their weeks.
The CIOs and COOs who figure that out first will have a meaningful head start.
Follow Rajesh on LinkedIn or X/Twitter for more enterprise AI perspective.
