Anthropic is shifting its enterprise AI strategy toward what it calls “multiplayer AI,” arguing that the biggest problem in business adoption is not model smarts but the fact that workers still use AI in isolation. Scott White, Anthropic’s head of product for enterprise, outlined the approach in an interview, framing it as a move from single-user chatbots to agents that read organizational context and insert themselves into teamwork without being prompted. The strategy’s centerpiece is an update to Claude Tag, the company’s Slack-based agent, which now reads entire conversations instead of evaluating messages one at a time. Anthropic says this change makes Claude about 30% better at judging when to jump into a channel unprompted and when to stay silent.
White described the update as a fundamental shift in what AI does for organizations. He said Claude previously felt like a personal chief of staff but now, in a company deployment, feels like the organization’s chief of staff. He mapped the industry’s evolution in three phases: AI first handled part of a single task, then whole tasks, and now it can pursue higher-level goals such as keeping a product bug-free or speeding up legal NDA reviews. That move toward goals, he argued, forces AI to become collaborative because goals inherently involve multiple people, and knowledge work lacks the coordination infrastructure that software engineering has in tools like Git.
White credited three factors for enabling this shift. The first is connectivity, citing Anthropic’s Model Context Protocol, an open standard introduced in late 2024 that rivals later adopted, which he called the USB-C for AI connectors. The second is a new threshold of model intelligence that makes proactive intervention useful rather than annoying; the model must connect data from many sources before it can spot a problem and offer to fix it. The third is form factor, meaning putting Claude where collaboration already happens, such as Slack, where it has a federated identity with permissions and channel awareness.
The Claude Tag update removes a previous classifier that evaluated each message in isolation. Now Claude reads full channel context plus its memory and instructions, then chooses among four actions: reply inline, start deeper work in a thread, route the message to an existing workstream, or say nothing. Anthropic gave an example of two engineers chasing the same bug from opposite ends; individually their messages needed no response, but together they contained a theory and supporting evidence, so Claude opened a thread with an investigation underway. The company also built in restraint, stating that an annoying agent is worse than an unhelpful one, and Claude goes dormant in channels where it repeatedly adds nothing.
Asked whether this actually makes teams faster, White pointed to his own experience. He said he used to wait a day or two for a data scientist to analyze a business question, but now Claude does that analysis, while the data scientist builds infrastructure to keep Claude accurate. The handoffs that once consumed time have been shortened dramatically, he said, leaving him more time to discuss strategy with peers. He cited site reliability engineering as an external proof point, where Claude collects error logs, links them to code changes and Slack discussions, and pulls in the right people.
The claims come as the broader market struggles to turn AI enthusiasm into financial returns. A McKinsey survey found that 88% of organizations use AI in at least one function and 62% experiment with agents, but only 39% attribute any earnings impact to AI, and just 6% qualify as high performers. Deloitte predicted that 25% of enterprises using generative AI would deploy agents in 2025, with that figure doubling to 50% by 2027. Anthropic is betting that redesigned workflows, not bolted-on chatbots, will separate winners from the rest, making a product that restructures team coordination a more durable advantage than a slightly better model.
Giving an agent standing access to Slack, documents, and connected systems raises security concerns, particularly prompt injection, where malicious instructions hide in content the AI reads. White described a layered defense, including classifiers trained to prevent such attacks. He noted that Anthropic delayed the launch of its Chrome extension, placing it on a waitlist to study prompt injection risks in a browser context, and then used novel classifier training on that data to improve defenses. The company also exposes compliance and analytics APIs for enterprise oversight.
More AI news from TechManNews.
