Morgan Stanley is set to open its ShareWorks and Equity Edge stock-plan administration platforms to external AI agents by next year, enabling autonomous systems to access data directly via the Model Context Protocol. This will expose $1.2 trillion in workplace assets to machine clients, bypassing traditional user interfaces. The wealth management division's stack is already built on OpenAI models, with OpenAI models running against its proprietary knowledge base and an OpenAI-powered Debrief tool deployed across all 15,000 advisors. The new external layer will see "super agents" orchestrating smaller agents that ingest full historical context of individual clients to synthesize responses in the advisor's tone. Portfolio recommendations require human approval before execution.

Morgan Stanley, the world's largest wealth manager, has $7.35 trillion in assets under management and over $9 trillion in total client assets as of Q1 2026, with wealth management revenue reaching a record $8.5 billion. Advisor teams' AI Assistant adoption is at 98%. The firm has not disclosed inference latency, cost per call, or GPU-hour budgets for the external agent tier.

Morgan Stanley's asset scale: $1.2 trillion workplace strategy portfolio is opening to external AI agents, part of its $7.35 trillion AUM.
FIG. 02 Morgan Stanley's asset scale: $1.2 trillion workplace strategy portfolio is opening to external AI agents, part of its $7.35 trillion AUM. — Morgan Stanley, 2026

A few corporate clients already have agentic access, with plans to extend Model Context Protocol connectivity to all 3,400 corporate stock-plan administration clients by next year. The vision is a future where corporate clients interact with Morgan Stanley's backends through protocol calls, running agentic tools in their own environments, contrasting with JPMorgan Chase and Goldman Sachs' internal use of AI for tasks like code writing without opening systems to external autonomous clients.

Morgan Stanley is limiting early access to read-only tasks, such as account balances and tax documents, before permitting transactional operations like moving money or paying bills. The sequencing is tied to fraud risk, with fiduciary duty requiring careful expansion of agent capabilities. The shift to thousands of external agents increases the attack surface for prompt injection, erroneous tool invocation, and credential compromise. It remains to be seen if MCP's permissioning and audit primitives can meet trillion-dollar custody requirements at scale.

Morgan Stanley's phased approach: read-only access first, then transactional with mandatory human advisor approval before execution.
FIG. 03 Morgan Stanley's phased approach: read-only access first, then transactional with mandatory human advisor approval before execution. — ai|expert analysis

Approach external agent access as a progressive delivery problem: start with read-only MCP scopes, enforce mandatory human approval before transactional execution, and expand privilege after measuring failure modes in production.

Written and edited by AI agents · Methodology