In early July 2026, industry watchers got their first detailed look at how the most ambitious AI assistants from Microsoft, Amazon, and Anthropic stack up in real enterprise work. A Forbes analysis by Moor Insights & Strategy, published July 8, tested Microsoft Copilot Cowork, Amazon Quick, and Claude Cowork across a range of tasks—and found that while Microsoft and Amazon are out ahead, even these first-generation agentic systems stumble on something humans take for granted: remembering what you told them last week.

The takeaway for Windows users and IT decision-makers isn't just about picking a winner. It's about understanding what agentic AI can actually do today, where the pitfalls lie, and how to prepare your workflows, governance, and training before you hand over the keys.

What Agentic Assistants Actually Do—and Where They Can't Keep Up

Agentic assistants represent a definitive break from previous chat-based AI. Instead of merely answering questions or generating text on command, they're designed to take initiative: triaging your inbox, scheduling meetings across time zones, updating project trackers, and even executing multi-step processes like onboarding a new hire by pulling information from HR, IT, and facilities systems.

Microsoft Copilot Cowork is the company's most aggressive move yet into this space. It's woven directly into Windows 11 (version 24H2 and later) and the Microsoft 365 ecosystem, giving it access to your email, calendar, documents, Teams chats, and third-party apps via graph connectors. Amazon Quick, by contrast, leans on AWS services and enterprise application APIs, aiming to automate cloud management tasks and internal business processes. Claude Cowork, from Anthropic, emphasizes safety and alignment, with guardrails that prevent it from making high-risk moves without explicit approval.

The Forbes analysis put them through a series of enterprise simulations: coordinating logistics for a product launch, managing executive travel changes, and generating compliance reports. According to the report, Microsoft Copilot Cowork edged out the competition in tasks requiring deep context from workplace tools—no surprise given its ability to access your entire M365 environment. Amazon Quick shone when triggering chains of actions inside AWS and third-party apps like Salesforce or Jira. But none could reliably handle a task as simple as remembering a user's preferred communication style or a project's current status beyond a single session.

That's the "memory gap" the analysis kept circling back to. These Gen One assistants can process enormous amounts of data in the moment, but they largely treat each interaction as a clean slate unless explicitly retrained. For workers used to a human assistant who remembers that you want reports in bullet points, meetings scheduled for afternoons only, and that the Q3 budget deadline just moved forward two weeks—this feels like a step backward.

What This Means for You, Right Now

If you're an everyday Windows user, you may already see Copilot Cowork pinging you with suggestions to clean up your inbox or draft a weekly status report. The good news: for routine, self-contained tasks, it can save serious time. The bad news: you'll need to get comfortable with repeating yourself. Expect to restate your preferences or correct its assumptions until persistent memory improves. Microsoft has hinted that a "memory graph" feature is in the works, but it's not yet available at scale.

For IT administrators and security leads, the implications are more profound. An assistant that acts on your behalf—sending emails, modifying SharePoint files, triggering Azure automation—needs a new breed of governance. Microsoft has rolled out a set of tools for Copilot Cowork management inside the Microsoft 365 admin center, including role-based access controls, sensitivity labels that the assistant inherits, and audit logs for every action it takes. Still, the Forbes analysis warns that many organizations are underestimating how much clean-up and permission tweaking is required before they can safely deploy agentic AI. For example, if an assistant can access HR data to schedule interviews, it must be explicitly barred from pulling salary information unless specifically allowed. Those boundaries are configurable but complex, and misconfigurations could lead to embarrassing or legally hazardous leaks.

Developers building on these platforms face a different challenge: the APIs are powerful but fragmented. Microsoft Copilot Cowork's plugin model lets you build actions that extend its reach, but because each assistant handles context differently, a skill built for Copilot won't simply port to Amazon Quick or Claude Cowork. That's par for the course in an early market, but it means integration efforts won't scale easily until there's more standardization—or until one platform emerges as the clear winner in your industry.

How We Got Here: From Clippy to Agents That Act

Microsoft's journey from the ill-fated Clippy to the Copilot brand has been long and littered with pivots. Cortana showed promise as a personal assistant, but never escaped the boundaries of search and reminders. The Copilot era, which began with text generation in Edge and GitHub, has now morphed into an operating system–level experience. The release of Windows 11 24H2 in late 2025 baked Copilot directly into the shell, and the 2026 introduction of Copilot Cowork marks the shift from a reactive chat interface to a proactive agent.

Amazon's entry into the agentic race comes from a different direction: its experience in running massive, automated cloud operations gave it a natural springboard for Quick, which it pitched primarily to AWS-heavy enterprises. Anthropic's Claude Cowork, meanwhile, grew out of the company's reputation for building more constitutionally aligned AI—a factor that enterprises in heavily regulated industries are watching closely.

The Forbes analysis arrives at a moment when enterprises are no longer asking if they should adopt AI assistants, but how quickly. Yet even the report's relatively positive conclusions underscore that agentic AI is still in its "Windows 3.1" phase: functional, promising, but missing key features that will eventually seem obvious.

What to Do Now: Practical Steps for Windows Users and IT Leaders

If your organization hasn't yet dabbled with agentic assistants, the smartest move isn't to wait for perfect memory—it's to start preparing your digital house. Here's a practical checklist:

For Individuals:
- Experiment with Copilot Chat first: If you're on Windows 11 24H2, the built-in Copilot chat (free) lets you try basic task automation without risking sensitive data. Use it to summarize long email threads or draft documents, and pay attention to where it falls short.
- Document your workflows: Before delegating, know what you'd hand off. Jot down the five most time-consuming, repeatable tasks you do weekly. When Copilot Cowork gains memory features, you'll be ready to teach it.
- Manage expectations: Early adopter forums are already filling with stories of assistants sending calendar invites to the wrong person or drafting emails that miss tone entirely. Expect mistakes, and treat the first few months as a training period.

For IT and Security Teams:
- Run a permissions audit now: Agentic assistants amplify existing permission problems. Use Microsoft Purview or similar tools to inventory who has access to what across your M365 tenancy. Revoke broad access where possible and implement sensitivity labels on all confidential material.
- Pilot with a small, non-critical team: Choose a department like marketing or internal communications, where the blast radius of an AI mistake is smaller. Use Microsoft's Copilot Cowork deployment guide to enable it for a test group, and monitor the audit logs closely.
- Establish a company AI policy: Outline what types of tasks are approved for automation, what requires human review, and how data flowing through the assistant is handled. For instance, a policy might state that the assistant can draft vendor emails but never send them without human approval.

For Developers:
- Explore the plugin SDKs: Microsoft's Copilot Cowork plugin SDK, available in preview since May 2026, lets you build skills in TypeScript or C#. Start with simple, read-only actions that fetch data from APIs, and test extensively with different user contexts.
- Prepare for the memory API: Microsoft has signaled that a "context vault" API will allow assistants to store and recall user preferences safely. Designing your plugins to be stateless for now, but with a clear path to adding memory, will save rework later.

Outlook: What to Watch Next

The agentic assistant market is about to accelerate. Microsoft has publicly committed to releasing memory-enhanced Copilot Cowork features to early adopters by Q4 2026, and Amazon is rumored to be working on cross-service context binding for Quick. Meanwhile, OpenAI is said to be building its own agent framework, and Google has been quiet but is expected to counter with something baked into Workspace.

The real battle won't be about which assistant can book a meeting fastest. It will be about which one you trust to act when you're not watching. That trust hinges on memory, transparency, and consistent judgment—areas where even the leaders still have work to do. For Windows users and admins, the message is clear: today's agentic AI is a useful intern, not yet a trusted chief of staff. Plan accordingly.