Key Takeaways
- 1.AI adoption in M&A diligence has been blocked by data room privacy, not capability.
- 2.Claude Managed Agents now support self-hosted sandboxes, keeping deal files inside the firm's own infrastructure.
- 3.MCP tunnels connect agents to internal CRMs, financial models, and deal systems through an encrypted private channel with no public endpoints.
- 4.Anthropic runs the orchestration; everything the agent touches stays inside the firm's perimeter.
- 5.Firms that build familiarity with perimeter-native AI now will have a setup advantage as the infrastructure matures.
The real reason AI has been slow in M&A diligence
Claude got upgrades this week that address the data room privacy concerns that have slowed AI adoption in M&A.
The reason AI has not moved faster in deal diligence is not capability. It is data room privacy. You cannot run sensitive financial, legal, or behavioral data through a public API without triggering compliance and confidentiality concerns.
These newly released features remove that blocker at the infrastructure level.
An advisor or PE firm can now deploy an AI agent that reads internal CRMs, financial models, and deal documents, without a single sensitive file crossing a public network.
Where these features actually live
To be clear, this update is not inside the normal chat with Claude or the Claude desktop app. It lives inside a separate web app called Claude Console, specifically built to create agents, which is why these are called Claude Managed Agents. The naming is messy. Functionally, Claude Console has the same chat interface you are used to, but instead of giving you an answer, it produces your own agents. Agents here are autonomous workers that figure out how to complete a task.
The two new features for Claude Managed Agents
The names sound abstract, but here is what they actually mean for deal work.
1. Self-hosted sandboxes
The AI agent now executes inside your own infrastructure. Files, data, and code stay where they already are. None of it leaves your environment. Your existing security controls, audit logs, and network policies apply by default, so the agent operates inside the same perimeter your deal team already trusts.
2. MCP tunnels (Model Context Protocol)
Your AI agent connects to internal databases, private APIs, and deal systems through an encrypted private channel. There are no public endpoints and no firewall exceptions to negotiate. It runs over one outbound connection from a lightweight gateway you control.
Anthropic keeps the agent orchestration on their end. Everything the agent actually touches stays inside yours.
Why this matters for deal teams
For the first time, a serious agent platform is moving toward perimeter-native deployment. The perimeter is the one you already set. That changes the conversation from "can we use AI on this deal?" to "which workflows do we want to point it at first?"
Both features are still early and in public beta. The direction is clear: enterprise AI for diligence is moving toward execution inside the firm's own environment, with the model provider handling orchestration only.
The firms that start building familiarity with this infrastructure now will have a significant setup advantage when it matures. Internal data access patterns, gateway configuration, audit posture, and agent design choices all compound over time.
Reference materials
- Claude Console: https://lnkd.in/eNSnn4GX
- Claude Managed Agents overview: https://lnkd.in/etjcsai6
This piece is part of the AI in M&A series running through the Under the Hood of M&A newsletter.
Frequently Asked Questions
- Why has AI been slow to take hold in M&A due diligence?
- The blocker has not been model capability. It has been data room privacy. Sensitive financial, legal, and behavioral deal data cannot be sent to public APIs without triggering confidentiality and compliance concerns, which has held back serious AI deployment inside diligence workflows.
- What are Claude Managed Agents?
- Claude Managed Agents are autonomous agents built inside Claude Console, a separate Anthropic web app dedicated to creating and operating agents. They have the familiar chat interface but are designed to execute tasks as autonomous workers rather than answer one-off prompts.
- What is a self-hosted sandbox?
- A self-hosted sandbox means the agent executes inside the firm's own infrastructure. Files, data, and code never leave that environment, and the firm's existing security controls, audit logs, and network policies apply by default.
- What is an MCP tunnel?
- An MCP tunnel is an encrypted private channel based on the Model Context Protocol. It lets a Claude agent connect to internal databases, private APIs, and deal systems through a single outbound connection from a lightweight gateway the firm controls, with no public endpoints or firewall exceptions.
- Why does this matter for advisors and PE firms?
- It removes the privacy blocker at the infrastructure level. An advisor or PE firm can now deploy an AI agent that reads internal CRMs, financial models, and deal documents without any sensitive file crossing a public network, which makes serious AI use in diligence operationally viable.
