Samuel Arendt

Enterprise AI starts with controls around the model

An AI demonstration answers whether a model can perform a task. A corporate deployment must answer different questions: who may use it, which data enters, which tools can take action, how much it costs, and how an incident can be investigated.

Claude Cowork's enterprise updates make that difference visible. Anthropic announced private plugin marketplaces, controls over connectors, and OpenTelemetry support. A later update added group roles, spending limits, and administrative analytics. The list maps dependencies that appear when agents stop being individual experiments.

Model capability still matters. Capability without control, however, creates an operation that is difficult to sustain.

Access must reflect the organization

Companies already separate systems by function, responsibility, and sensitivity. An agent connected to those systems should respect the same boundaries. Legal, finance, and engineering can use the same AI product without sharing the same catalog of actions.

Group controls allow capabilities to be enabled by role. That reduces accidental exposure and makes it easier to explain why someone can read one source but cannot modify another. The familiar principle is least privilege: grant only the access required for the work.

This design should reach the tool level. A connector may offer both read and write operations, while the organization may need to permit only reading. Broad permission at the product level is easy to configure but weak when a single action can create a financial commitment, delete data, or communicate with customers.

Distribution is also governance

When every user installs plugins independently, the company loses a common view of the environment. Versions diverge, integrations are duplicated, and local instructions may conflict with internal policy. A private marketplace addresses part of the problem by creating a known source for approved capabilities.

Central distribution should not become central accumulation. Every plugin needs an owner, a purpose, permissions, a version, and a removal path. Automatic installation makes sense only when a group depends on that capability. Otherwise the organization replaces scattered configuration with an inflated catalog.

It also helps to separate knowledge packages from credentials. A skill can explain a process without accessing systems. A connector can provide access without deciding how the agent should act. Combining the two makes audits and revocation harder.

Observability should support a decision

OpenTelemetry provides a familiar format for sending events into observability systems. According to Anthropic, Cowork events can include tool and connector calls, files read or modified, skills used, and the approval mode applied to an action.

Collecting everything without an operational question creates another pile of logs. The organization must decide which events help control cost, investigate incidents, and identify usage patterns. It also needs rules for retention, access, and sensitive content. Agent telemetry may contain filenames, parameters, and decisions that deserve protection of their own.

An initial implementation can answer four questions: who started the session, which capability was called, whether the action required approval, and how many resources it consumed. Additional fields belong only when a concrete case justifies them.

Budget belongs to the architecture

Agents execute variable sequences. Two people can request similar outcomes and incur different costs because of the chosen route, files, and tools. Group limits and usage analysis are therefore product controls as much as finance controls.

The purpose is not to punish heavy usage. It is to distinguish productive adoption from loops, weakly designed automation, and expensive capabilities applied to simple work. A team may accept higher cost for security review and require smaller models for routine classification.

That policy becomes easier when telemetry connects spending to a workflow. Cost alone says little; cost until an accepted result guides a decision.

A minimum deployment checklist

Before expanding access, a company should be able to identify the owner of every integration, authorized groups, write operations, accessible data, and the revocation procedure. It should test employee departure, plugin removal, and investigation of an improper call.

The rollout can start with one group and a few recurring workflows. The team tracks outcomes, interventions, cost, and incidents. New capabilities enter when observed demand exists, rather than to complete an internal showcase.

"Enterprise-ready" describes operational predictability more than response quality. If the organization cannot list access, distribute configuration, observe actions, and limit spending, it still has an individual tool being used by many people.

A basic sign of maturity is answering these questions without assembling a task force. When access, cost, and responsibility appear in the normal workflow, the company can expand usage without relying on the memory of whoever installed the first integration. The same record shortens incident response when behavior departs from expectations.

If your company is weighing something similar, reach me on LinkedIn or WhatsApp.

References

#AI Agents, #Engineering Management, #Operations