Anthropic did not launch Claude Tag as a productivity bot or an AI search tool. It launched it as a team member — a named identity that joins channels, holds context, and acts without being prompted. That framing is not accidental. The difference between a bot you invoke and a teammate you delegate to is the difference between a tool and a colleague, and the latter creates a category of reliance that is qualitatively different from subscription software. When an organization's Slack channels start producing decisions shaped by Claude Tag's proactive suggestions, the boundary between human judgment and AI output becomes genuinely hard to locate.

The architectural choice to give each Slack channel its own isolated Claude identity looks like a privacy feature. It functions as a product strategy. By keeping channel contexts separate, Anthropic avoids the liability of a single model that knows everything about an organization — but it also ensures that the knowledge embedded in each channel has no exit path. An organization that has used Claude Tag in its engineering channel for a year has deposited into that channel's identity a compressed version of every architectural debate, every incident postmortem, every onboarding shortcut. That knowledge is not exportable in any form that retains its usefulness. The always-on AI teammate in Slack becomes the only interface through which that accumulated context is navigable — and Anthropic holds the key.

Anthropic's disclosure that 65% of its product team's code now comes from an internal version of Claude Tag is calibrated for a specific audience: enterprise procurement teams who have spent the past two years hearing AI vendors make capability claims with no production evidence. The number sidesteps the benchmark conversation entirely. It says: we run this at the team level, in the product that builds Claude, and it changed how we ship. For a buyer who has watched AI tools generate impressive demos and then underperform in deployment, a vendor that is also a reference customer at scale is a materially different proposition than one pitching from the outside.

The decision to launch Claude Tag as a research preview rather than a full product release is standard for features that need real deployment data to mature — but it also means Anthropic is explicitly asking early enterprise adopters to shape the product through their usage. The organizations that integrate Claude Tag into their workflows in the next six months are not testing a finished tool; they are generating the behavioral data that determines what the finished tool looks like. This is a common pattern in enterprise AI, but it carries an asymmetry worth naming: the enterprise learns whether the tool is useful; Anthropic learns how enterprise teams actually use persistent AI teammates, a dataset that has no substitute and that will advantage Anthropic's product motion regardless of whether any individual early adopter renews their subscription.

The organizations most at risk from Claude Tag's lock-in mechanics are not the ones that make a deliberate choice to centralize their institutional knowledge in an AI system. They are the ones that adopt Claude Tag as a convenience — faster answers, lighter async load — and discover twelve months later that their new hires learn the codebase through Claude Tag, that their incident response has been shaped by what the channel's AI identity flags, and that the people who knew the pre-Claude context have moved on. Dependency built through accumulated convenience is harder to reverse than dependency built through deliberate integration, because it has no single decision to point at and revisit. Anthropic's Claude Tag strategy lands in a market where that dynamic is already well understood — and where enterprises are adopting anyway, because the productivity case is real enough to outweigh the lock-in concern until the lock-in is already complete.

Task automation is a race Anthropic will not win alone. OpenAI, Google, and a growing field of vertical AI tools all offer async task handling, codebase access, and proactive suggestions. The feature set that distinguishes Claude Tag today will be table stakes within a product cycle. What does not commoditize at the same rate is the organizational memory that Claude Tag is positioned to absorb — the specific, non-transferable context that makes a general-purpose model useful inside a particular company. Anthropic's bet is that by the time the feature competition catches up, the institutional knowledge deposited in Claude Tag channels will make the switching question moot. The enterprises evaluating Claude Tag as a productivity tool are also, without deciding to, evaluating whether they will be in a position to switch later — and for organizations that adopt early and use the product as designed, the accumulated context will have already made that decision for them.