Product Update · Governance
Claude Cowork Goes GA
April 14, 2026 · 6 min read · Liam OBrien
Anthropic shipped six enterprise controls on 9 April. Most of the “is this safe for hedge funds?”takes are getting the governance story wrong. Here's the accurate read for investment managers.
On 9 April 2026, Anthropic moved Claude Cowork out of research preview. It's now generally available on all paid plans across macOS and Windows, with six enterprise controls aimed at organization-wide deployment.
For investment manager CIOs and CCOs evaluating Cowork, the governance story is materially stronger than it was a week ago. With the right scoping and a written policy, Cowork is deployable today for research and operational workflows.
What Shipped
Cowork brings Claude Code's agentic architecture into Claude Desktop — task delegation, local file access, sub-agent coordination, scheduled tasks, and professional outputs including Excel with working formulas and formatted documents. The GA release layers six organization-level controls on top:
Role-based access controls
Enterprise-only. Group users manually or via SCIM from your identity provider, assign custom roles, turn Cowork on for specific teams first.
Group spend limits
Per-team budgets from the admin console. Predictable costs, adjustable as usage patterns emerge.
Usage analytics
Cowork activity in the admin dashboard and Analytics API. Per-user activity, skill and connector invocations, DAU/WAU/MAU alongside Chat and Claude Code metrics.
Expanded OpenTelemetry
Emits events for tool and connector calls, file access, skills used, and whether each AI-initiated action was approved manually or automatically. Compatible with Splunk, Cribl, Elastic, Datadog.
Zoom MCP connector
AI Companion meeting summaries, action items, transcripts, and smart recordings flow into Cowork workflows.
Per-tool connector controls
Admins can restrict actions within each MCP connector — e.g., allow read but disable writes. Configured org-wide from the admin console.
Enterprise admins can also require that all Cowork activity execute inside a virtual machine sandbox on member devices — a meaningful control for shops with strict endpoint data-loss policies.
The Governance Takes Most People Are Getting Wrong
The headline most people are running with: “Cowork activity isn't in Anthropic's audit logs — so don't use it for regulated workloads.”
That framing is doing a lot of work. For investment managers, it's mostly wrong.
The core distinction: activity vs. content
The OpenTelemetry stream gives near-complete visibility into what Claude did — every prompt, every tool call, every data-source query, every file write, every approval decision, with costs and timing, streamed into the firm's existing SIEM. What OTEL doesn't stream centrally is content: Claude's reasoning commentary on screen, the full body of output documents, the raw data pulled from connected sources. That content lives where it always has — at the source, on disk, or in the user's local session history.
Captured in OTEL → your SIEM
- — Full user prompt text
- — Every MCP / tool call: server, endpoint, query parameters
- — Every file path Claude reads, modifies, or writes
- — Every human approval decision
- — Model, tokens, cost, duration per API call
- — Correlation ID linking everything from one prompt
- — User identity, session, workspace paths
Lives elsewhere (not in OTEL)
- — Raw data returned from connected data sources — in the source system's own audit log
- — Full content of output files — on disk, under existing file controls
- — Claude's on-screen reasoning and commentary — local to the user's Cowork history
A 204-2 Clarification for the CCOs
Rule 204-2(a)(7) under the Investment Advisers Act applies to written communications sent and received— not to information generated inside an application. From Skadden's September 2024 analysis of SEC recordkeeping rules applied to AI-generated content:
“AI-generated information that is not subsequently transmitted likely does not constitute a written communication that must be retained under relevant IA and BD recordkeeping rules… The generation of information in an application cannot reasonably be interpreted as a sent or received communication.”
— Skadden, Arps, Slate, Meagher & Flom LLP, September 2024
An analyst using Cowork to pull fundamentals from FactSet, synthesize, and draft a research memo is not creating a 204-2 record inside Cowork. The eventual email or memo to the PM is the record — and it's captured by the firm's existing communications archiving.
Where the audit gap does matter is narrower: AI-generated marketing or advertising content (204-2(a)(11) captures internal working papers for disseminated advertisements), firm-policy audit mandates under 206(4)-7 compliance programs, client-facing workflows where AI is the communication channel, and the SEC's proposed predictive-data-analytics rule if finalized. Real cases to scope around — not a blanket ban.
A Concrete Research Workflow
An analyst prompts Cowork: “Pull Apple's revenue, EPS, and consensus estimates for the next quarter, and write me a one-page summary.”
Claude calls the FactSet MCP for Fundamentals (historical revenue and EPS) and Estimates Consensus (forward expectations), synthesizes, and writes the output as a markdown file.
Where the record lives
- OTEL → the prompt, each FactSet query with parameters, the file-write for the markdown, size of what was pulled, who ran it, when, for how much
- FactSet's own audit log → the data pulls at the source
- On disk → the markdown file itself, as a real file under existing file controls
- Email archive → when the analyst shares the note with the PM, that email is the firm's 204-2(a)(7) record
No dead-ends for a CCO reconstructing the workflow. The record is distributed across the firm's existing source-level, file-level, and email-level controls, with OTEL providing the activity glue.
What a Defensible Rollout Looks Like
With the new controls, a defensible Cowork rollout at a typical hedge fund or asset manager is straightforward:
Scoped access via RBAC
Cowork enabled for groups whose work benefits most from agentic execution — research, IR, operations. Use-case boundaries (not blanket bans) for workflows where AI output enters regulated channels such as marketing collateral or client communications.
Spend discipline
Group spend limits prevent a single team's experimentation from surprising the CFO.
SIEM integration
OTEL events piped into the firm's existing Splunk / Cribl / Elastic stack — same surveillance layer as the rest of the tech estate.
Connector discipline
Read-only access to systems of record where writes aren't needed; write access gated by policy for specific workflows.
VM sandbox
All Cowork execution inside an isolated VM on the endpoint — limits blast radius from prompt injection and reduces data-exfiltration surface.
This is the first Cowork release where the full deployment stack — identity, permissioning, observability, spend, sandboxing — is coherent enough for an institutional rollout.
Bottom Line
Anthropic's own conservative language — “do not enable Cowork for regulated workloads” if formal audit trails are required — is a defensible posture from their side. Read alongside the actual OTEL coverage and where the recordkeeping rule actually applies, a scoped Cowork deployment with a written policy sits well inside what a reasonable CCO would approve today.
The governance story moved in a week. Most of the FUD didn't.
Sources
Making Claude Cowork ready for enterprise
Anthropic · 9 April 2026 — GA announcement and enterprise feature overview
Use Claude Cowork on Team and Enterprise plans
Anthropic Claude Help Center — admin controls, compliance limitations, VM sandbox
Cowork Monitoring Reference (OpenTelemetry)
Anthropic Claude Docs — full event schema for OTEL integration
How and When SEC Recordkeeping Rules May Apply to AI-Generated Content
Skadden, Arps, Slate, Meagher & Flom · 13 September 2024
17 C.F.R. §275.204-2 — Books and Records to be Maintained by Investment Advisers
Electronic Code of Federal Regulations
Evaluating Cowork for Your Firm?
We work with investment managers on Claude Enterprise deployments — role model design, OTEL pipeline setup, connector policy, and the written governance needed to satisfy a CCO and an auditor.
Book a 30-Minute Scoping Call