Market Intel — May 26, 2026

The Privacy Problem That Was Killing SMB Agent Adoption Just Got Solved

Anthropic’s MCP tunnels and self-hosted sandboxes let small businesses run Claude agents against private data — without exposing it to the public internet. Here’s what that actually means.

There’s been a consistent friction point in selling AI agents to serious small businesses: the data problem.

Accounting firms. Law offices. Healthcare practices. Real estate brokerages. These are exactly the clients who could benefit most from autonomous agents. They’re drowning in repetitive document work, client intake, scheduling, and reporting. But the moment you say “your agent connects to your database,” someone in the room asks the question: where does that data go?

Until recently, the honest answer was “Anthropic’s cloud.” For a two-person law firm handling client files, that was a dealbreaker. Now the answer is different.

What Anthropic Just Shipped

In the last three weeks, Anthropic has quietly flipped the enterprise privacy calculus for Claude Managed Agents:

• Self-hosted sandboxes (public beta): Tool execution — the part where agents actually touch your files, databases, and APIs — now runs in your own infrastructure. Anthropic handles orchestration and reasoning. Your data never leaves your environment.
• MCP tunnels (research preview): Agents can reach internal databases, private APIs, knowledge bases, and ticketing systems through an encrypted tunnel with zero inbound firewall rules. One lightweight gateway, one outbound connection. Your internal systems stay dark to the public internet.

Combined, these features mean the agent’s brain runs at Anthropic but its hands operate inside your walls. That’s a fundamentally different risk profile than what existed six weeks ago.

Why This Matters More Than It Sounds

The SMB AI adoption data tells a clear story. Adoption among businesses with 10–100 employees jumped from 47% to 68% in one year — a reversal where small businesses are now adopting AI faster than large enterprises. The average small business uses five AI tools.

But “AI tools” mostly means chatbots, copywriting, and scheduling apps. The real productivity leverage — agents that work against your actual business data, close the feedback loop, and take autonomous action — has stalled in regulated and data-sensitive industries for one reason: compliance anxiety.

That anxiety is legitimate. A dental practice with patient records. A property manager with lease agreements. A CPA firm with client tax data. These aren’t edge cases — they’re the SMB economy’s backbone. And they’ve been frozen out of the productivity gains everyone else is talking about.

MCP tunnels and self-hosted sandboxes dissolve that freeze. The agent can query your internal database, draft the report, update the CRM — and your sensitive data never touches a public endpoint.

The Competitor Move You Should Know About

OpenAI is moving hard in the other direction. Two weeks ago they announced the OpenAI Deployment Company — a B enterprise services unit built to deploy agents inside large organizations. They acquired Tomoro, an AI consulting firm, to bring 150 engineers onboard from day one.

Read that clearly: OpenAI is betting on high-touch, high-cost enterprise deployments. Their money is going upmarket.

Anthropic is making the infrastructure better for everyone else — including the service providers building on top of Claude. That gap is a window. The SMBs who can’t afford a B deployment company still need agents. Someone has to build and manage those deployments. That’s the business opportunity.

Three Industries to Target Right Now

With this infrastructure in place, here are the verticals where the privacy objection collapses and the ROI case writes itself:

1. Legal services (small firms, solo practitioners): Contract review, client intake summaries, deposition prep, billing time analysis. Private data stays on-prem. Agent runs against actual case files.
2. Healthcare-adjacent (chiropractic, dental, therapy practices): Patient intake, insurance pre-auth research, appointment scheduling intelligence. HIPAA-sensitive workflows are now viable.
3. Financial services (bookkeepers, CPAs, independent advisors): Transaction categorization, client report drafting, anomaly flagging. Connecting to live accounting software without data leaving the client’s environment.

These aren’t new use cases. They’ve been sitting in the “not yet” column because the privacy infrastructure didn’t exist. It does now.

What to Do This Week

• Request MCP tunnels access at claude.com/blog/claude-managed-agents-updates — it’s research preview, limited access, worth getting in the queue early.
• Audit your prospect list for any deal that stalled on “where does our data go.” Reopen those conversations. The answer has changed.
• Update your intake questions. Start asking about internal tooling — what databases, what CRMs, what internal APIs does the client run. That’s the connective tissue for an MCP-tunneled deployment.
• Position your offering as infrastructure, not chatbots. The market is flooded with chatbot builders. Very few people are offering private-network agent deployments for sub-M revenue businesses. That’s your differentiation.

The bottom line: The privacy gap that blocked AI agents from the most valuable SMB verticals just closed. The businesses that move first — and the service providers that know how to implement this — will own the next 18 months of market share in those industries. The window is real. It won’t stay open long.