Microsoft moved two of its Dynamics 365 agents to general availability on Monday. Sales Agent handles the paperwork side of a deal cycle: pre-call account summaries, post-call CRM capture in natural language, automatic logging of objections and commitments, next-step drafting. Service Agent handles the customer support side: case summarization, next-best-action prompts, resolution email drafts, record updates without the human ever leaving the case. Both are grounded in live CRM data via Model Context Protocol, which is the connective tissue Anthropic proposed a year ago and which has since quietly become the dominant standard for how enterprise agents get context they can trust. Sandvik Coromant and Northern Trust are the two named launch customers. Deva Rajamohan, Microsoft’s Dynamics 365 CX VP, put out a line about agentic AI resetting the equation by bringing intelligence into the flow of work rather than adding another tab. That is exactly the shape of pitch Salesforce is also making about Agentforce, so what customers are actually going to compare here is whether Copilot’s grip on the Microsoft 365 surface area matters more than Salesforce’s grip on the CRM record itself.
The next morning, Forbes published a hands-on comparison of Microsoft Copilot Cowork, Amazon Quick, and Claude Cowork against a small set of realistic enterprise scenarios. Product launch logistics. Executive travel coordination. Compliance report drafting. Microsoft edged out the field on tasks that required deep workplace context, which is what you would expect from the vendor that already owns the mailbox, the calendar, the file store, and now the CRM. Amazon Quick was stronger on action chains that reached across AWS and into third-party systems like Salesforce and Jira, which is the fingerprint of a company that thinks about agents as workflow automation. Claude Cowork was the most cautious of the three, refusing to take high-risk actions without human approval, which is either a feature or a bug depending on whether you have ever had an autonomous system book a wrong flight for your CEO.
All three shared the same failure. They cannot hold onto anything across sessions. Every time you come back, you re-explain your project, your preferences, your customer, your team, your context. The Forbes framing was that these are Gen One assistants that can process enormous amounts of data in the moment but treat each interaction as a clean slate unless someone explicitly retrains them. That is a real limit. A human executive assistant that forgot the CEO’s calendar preferences every Monday would be fired by Tuesday. Enterprise agents are being asked to pass, at best, on Monday morning.
The Gartner numbers everyone is quoting this week say 40 percent of enterprise applications will contain task-specific agents by year-end, and 40 percent of current agent projects will be scrapped by 2027 due to bad data, unclear ROI, or the memory problem above. Those numbers are almost certainly both directionally true, which is what makes them useful. Microsoft and Salesforce and everybody else can ship the GA banner on Monday. The category will still be sorting the ones that stick from the ones that get quietly turned off around Q3 of next year. The pilots that survive that cull will be the ones that solved the persistence problem. Nobody has yet.