In my previous article, I argued that agentic assistants have changed knowledge work for the better but remain demanding tools that are best suited for power users. The largest software vendors have addressed this reality with what I have been calling the Claude-ification of the enterprise: a wave of look-alike desktop agents for broader use — modeled on Claude Cowork — each built with the hope of becoming the primary interface for knowledge workers.
After tracking and testing this category through the spring, my read is that two vendors, Microsoft and Amazon, have credible first entries in this segment, yet this first generation still shows real gaps. It’s important to understand how the enterprise tools differ from a power-user product like Claude Cowork, and where they fall short.
(Note: Both Amazon and Microsoft, along with many of their competitors, are advisory clients of my firm, Moor Insights & Strategy.)
Claude Cowork Remains The Power-User Benchmark
Start with the benchmark. Claude Cowork put the agentic architecture behind Claude Code into a desktop app, gave it access to a designated local folder, and let users extend it with “skills” and custom MCP connectors. That local-file access and deep configurability give Claude Cowork its superpowers but also define its ceiling. Getting professional results rewards a well-tended personal context store, also known as a “second brain” (mine lives in Obsidian), and a willingness to maintain the app’s skills and connectors. Most employees will never do that. The enterprise contenders are betting they can deliver most of that value with far less of the setup, and with the governance a CIO needs.
What Makes An Agentic Assistant Enterprise-Capable?
Before judging any single product, it helps to be explicit about what “enterprise-capable” means. From the deployments I have reviewed this year, four capabilities separate a serious enterprise assistant from a souped-up chatbot.
- Centralized Provisioning — Enterprises need a way to distribute agents, skills, flows and other agentic artifacts to the right users, and to version-control them as they change. That can be built directly into the product or delivered through a backend agentic control plane. My view on this has not changed. At scale, this layer will be mandatory, and the assistant you see on the desktop is only the tip of it.
- Memory and Personal Context Stores — These are two distinct capabilities, and the difference between them matters. Memory is managed by the agent: It learns from your sessions and the artifacts you generate, getting more useful the more you work with it. A personal context store is more like a database the user controls, a place to deposit anything contextual, from external sources of information to files and notes. This is the enterprise answer to the power user’s hand-built second brain. A serious assistant needs both, and either can be embedded in the product or supplied by external services.
- Flexibility — Model choice is still narrow today, but enterprises will increasingly bring their own and industry-specific models, and those need to be supported. Just as important, it should be easy to build connectors, via MCP or other APIs, for bespoke applications or for any app a default connector does not yet cover.
- Data Protection — This is the broad one: security, data sovereignty, language guardrails and worker-privacy protections. It is also a moving target that will evolve alongside regulation and real-world use. For most enterprises, this is the capability that determines whether an assistant ever gets past the pilot.
With this yardstick of capabilities in hand, the two most consequential enterprise challengers, Microsoft and Amazon, come into sharper focus.
The Enterprise Challengers
Microsoft Copilot Cowork
Microsoft has the most direct response. Announced in March with Anthropic’s help and made generally available in June, Copilot Cowork brings agentic, cross-app automation into Microsoft 365. Measured against the four criteria above, it is strongest on flexibility and data protection. On flexibility, it is the one tool here that offers real model choice. Depending on entitlement, you can route between Claude and OpenAI models, with a Copilot Cowork-specific model on the way. Microsoft claims 30% to 40% consumption savings versus Claude when using the Microsoft 365 connector; this is plausible given that Microsoft owns more of the stack, though I would want to see it proven in real-world customer numbers before treating it as fact.
In terms of data protection, Copilot Cowork delivers a provisioned, contained experience with the security and governance controls enterprises expect — a far better answer to the shadow-AI problem than handing power users an open tool. And rather than building provisioning into the app, Microsoft handles it through a backend platform, Agent 365 and Work IQ.
The trade-offs for Copilot Cowork are just as clear. It is cloud-only, with no local file system and an unclear story for personal context, and the pricing is layered, requiring a Microsoft 365 Copilot license plus consumption charges on top. For most enterprises, that is a reasonable bargain; for the local-AI crowd, it is a deal-breaker.
Amazon Quick
Amazon’s entry is Quick, which AWS shipped in late 2025 and continues to extend. It is conceptually close to Copilot Cowork, though the difference in philosophy is telling. Where Microsoft leans on a backend platform, Amazon builds these capabilities into the product. Provisioning is centralized inside Quick itself.
Quick also treats memory and the personal context store as two separate constructs, which is the right call. A personal knowledge graph serves as the agent-managed memory that learns your role, priorities and relationships across sessions, while connected knowledge bases give users a curated store for their own files and external sources. Together they spare users from building their own second brains. Data protection was the other major highlight of the launch, spanning security, sovereignty and privacy controls.
One caveat: I have spent hands-on time with Quick but not yet with Copilot Cowork, so I will hold off on comparative usability judgments until I have. On flexibility, my hope is that over time Quick reaches parity with Amazon’s more open AI offerings like Bedrock, with broader model choice and evaluations chief among them. The early enterprise commitments are significant, with AWS-cited deployments planned across more than 100,000 seats at some customers.
The Gaps That Define Gen One Of Enterprise AI Assistants
For all the momentum, this is a first generation of products, and it shows. Five gaps stand out.
- Agent Mobility — As with human assistants, the longer you work with an agent the more valuable, and the harder to replace, it becomes. A big reason to externalize capabilities like the second brain is that you can carry them across different agents and assistants. Just as MCP made tools portable, vendors should build mechanisms to make a user’s own memory, context stores and artifacts mobile, so they move with you and upgrade easily as new agents and devices arrive. Without that, you are locked into a single product.
- Pricing That Resists Planning — Between per-user licenses, consumption fees and variable task costs, it is genuinely hard to forecast spend. Vendors that own the full stack promise savings, but until customers see it in their own financial analyses, treat those claims as unproven.
- Governance That Is Real, Not Cosmetic — As I argued in my recent work on ERP, “human in the loop” is too often a notification that can easily be ignored rather than an effective control point. When an approver gets an alert without the context to properly evaluate it, the tendency is to click yes on reflex. Making interventions meaningful is an infrastructure problem, not a UI checkbox, and most of these products are not there yet.
- Cross-Vendor Interoperability — Standards like A2A look good on paper, but production hand-offs between agents from different vendors remain largely unproven. In a multi-vendor enterprise, that gap matters.
- Sprawl and Evaluation — The same no-code ease that drives adoption invites the old citizen-developer pathologies of sprawl, security drift and unmaintained technical debt. The encouraging sign is that AI can now help police what AI builders are doing through agent simulation, anomaly detection and better observability. Whether this generation finally makes citizen development governable is the open question I am watching most closely.
Starting Points, Not Finish Lines, For Enterprise Assistants
None of this is a knock on the technology. It is a map of where we are. The Claude-ification trend is, I believe, a durable shift in how people will engage AI. But for the traditional software players to make it stick, they need to do two things: (1) build management features for the 90% of employees who are not power users, and (2) use these assistants to showcase the best of their AI, not just e-mail summaries and chat. Microsoft and AWS have both established credible starting points. Starting points are not finish lines.
Moor Insights & Strategy provides or has provided paid services to technology companies, like all tech industry research and analyst firms. These services include research, analysis, advising, consulting, benchmarking, acquisition matchmaking and video and speaking sponsorships. Of the companies mentioned in this article, Moor Insights & Strategy currently has (or has had) a paid business relationship with AWS and Microsoft.
