From Chatbot to Coworker

The biggest AI companies are shipping tools that want to work for you, not just chat with you.

Something shifted last week, and it wasn’t subtle. Within the span of a few days, OpenAI launched an enterprise platform built around the concept of “AI coworkers,” Anthropic released a model explicitly designed for what it calls “vibe working,” and Google rolled out features that let its AI read your email, study your photos, and use what it learns to reshape your search results. Each announcement, taken alone, might register as an incremental product update. Taken together, they mark a distinct transition in what these tools are actually asking of their users, and what they’re offering in return.

The common thread is this: AI is crossing from conversational to operational. The chatbot era, in which one typed a question and received an answer, is giving way to something more intimate and more ambitious. The new tools don’t just respond. They remember, they act, they learn from the context of your work and your life, and they do so with increasing autonomy. For anyone paying attention to the trajectory of these products, the last week offered the clearest picture yet of where all three major AI companies think the future lies.

OpenAI Frontier: The Agent Gets an HR Department

On Wednesday, OpenAI introduced Frontier, a platform that lets enterprises build, deploy, and manage what the company insists on calling “AI coworkers.” The language is deliberate. Frontier doesn’t frame its agents as tools or assistants; it frames them as employees. They get onboarded. They receive performance reviews. They build memories over time, turning past interactions into institutional knowledge. They have identities, permissions, and access controls that mirror what a human worker would receive.

The platform connects to an organization’s existing infrastructure, pulling together data warehouses, CRM systems, ticketing tools, and internal applications into what OpenAI describes as a “semantic layer for the enterprise.” The idea is that an AI agent working within Frontier understands not just the question it’s been asked, but the organizational context in which that question matters: how information flows, where decisions happen, what outcomes the company cares about.

Fidji Simo, OpenAI’s CEO of Applications, framed the problem Frontier is solving by recalling her time as CEO of Instacart: her team spent months integrating AI tools from hundreds of vendors, only to end up reinforcing the very silos they were trying to break down. “Every CEO has asked me, where is this all going?” she told reporters. “I tell them that it’s about humans and AI collaborating on one platform.”

The early adopter list is revealing: Uber, State Farm, Intuit, Thermo Fisher, HP, Oracle. These are not experimental deployments at scrappy startups. They are Fortune 500 companies embedding AI agents into operational workflows that directly affect revenue and customer relationships. Whether Frontier delivers on its ambitious framing remains to be seen, since the platform is currently available only to a limited group of customers, but the intent is unmistakable. OpenAI is no longer selling a chatbot. It is selling a workforce.

Anthropic’s “Vibe Working” and the Opus 4.6 Thesis

On the same day OpenAI shipped Frontier, Anthropic released Claude Opus 4.6, its most capable model to date, and used the occasion to make an explicit argument about where AI productivity is heading.

“Everybody has seen this transformation happen with software engineering in the last year and a half, where vibe coding started to exist as a concept, and people could now do things with their ideas,” Scott White, Anthropic’s head of product for enterprise, told CNBC. “I think that we are now transitioning almost into vibe working.”

The term is a conscious extension of “vibe coding,” the practice (coined by Andrej Karpathy in early 2025) of describing software in natural language and letting AI handle the implementation. “Vibe working” applies the same logic to knowledge work broadly: describe the outcome you want, and let the model figure out the steps. Opus 4.6 is built to make that practical. Its 1-million-token context window (a first for the Opus family) lets it ingest entire codebases or massive document sets in a single session. Its new “Agent Teams” feature allows multiple AI agents to work in parallel on different components of a task, coordinating autonomously. And a new PowerPoint integration, available in research preview, lets users build presentations with Claude reading layouts and fonts to match corporate templates.

But the most telling metric may be the benchmark Anthropic chose to highlight. Opus 4.6 holds the top spot on the Finance Agent benchmark, which evaluates how well AI performs core financial analyst tasks. This is not a model optimized for conversation. It is a model optimized for work product.

White put it plainly: “Claude went from a model that you can sort of talk to to accomplish a very small task or get an answer, to something that you can actually hand real significant work to. Opus 4.6 is a model that makes that shift really concrete for our users.” When combined with Cowork, the agentic desktop tool Anthropic released in January, and the 11 open-source plugins it shipped for legal, finance, sales, and other professional domains, the picture that emerges is of a company systematically building the infrastructure for AI to operate as a knowledge worker, not a knowledge retriever.

Google’s Personal Intelligence: When Search Starts Reading Your Mail

Google’s contribution to this moment arrived in January, though its full implications are still unfolding. The company rolled out what it calls “Personal Intelligence” across the Gemini app and AI Mode in Search, and simultaneously overhauled Gmail with a suite of AI features powered by Gemini 3.

Personal Intelligence is the more consequential of the two. It allows Gemini to connect, with the user’s opt-in consent, to Gmail, Google Photos, YouTube history, and Search data, then use that combined context to deliver personalized responses. The practical examples Google offers are revealing. Planning a trip? AI Mode can reference your hotel booking in Gmail and your travel memories in Google Photos to suggest restaurants and activities tailored to your family’s actual preferences. Need a new coat? It can check your flight confirmation to determine you’re heading to Chicago in March, cross-reference your past purchases to identify preferred brands, and recommend something windproof.

The Gmail overhaul operates in a similar spirit. AI Overviews now summarize long email threads and answer natural-language questions about your inbox (”What was that contractor’s last quote?”). A new AI Inbox feature, currently in limited testing, filters messages by priority, surfacing VIPs and time-sensitive items while suppressing noise. Help Me Write and Suggested Replies are now free for all users, with the latter learning to match individual writing styles.

Google emphasizes that Personal Intelligence is opt-in, that it doesn’t train directly on one’s Gmail inbox or Photos library, and that users can toggle individual app connections on or off. These are meaningful guardrails. But the trajectory is clear: Google is leveraging the depth of its existing ecosystem, which includes billions of emails, photos, searches, and viewing histories, to build an AI layer that doesn’t just know the world’s information but knows the user. No other company has this particular advantage. OpenAI and Anthropic can build smarter models, but Google already has the data that makes personal context possible at scale.

Three Approaches to the Same Idea

What makes this moment worth marking is not that any one of these products is revolutionary on its own. It’s that all three arrived within days of each other, and that all three are built on the same underlying bet: that the next phase of AI is not about generating better text, code, or images, but about understanding the user’s context well enough to take meaningful action without being told exactly what to do.

OpenAI is approaching this through enterprise infrastructure, building the plumbing that lets AI agents operate across corporate systems with defined roles and governance. Anthropic is approaching it through model capability, making Claude smart enough to handle complex, multi-step professional work that previously required a human analyst or associate. Google is approaching it through personal data, using the information it already holds about billions of users to make its AI anticipatory rather than reactive.

Each approach carries distinct risks. OpenAI’s Frontier asks enterprises to trust AI agents with operational autonomy and access to sensitive systems, and the question of what happens when those agents make consequential mistakes has no established playbook. Anthropic’s “vibe working” thesis depends on outputs being reliably production-ready, and anyone who has used AI for professional work knows how far that reliability still has to go. Google’s Personal Intelligence raises the most intuitive privacy concern: the company is asking users to let AI read their email and study their photos, and the fact that this is opt-in doesn’t fully resolve the discomfort many will feel about the proposition itself.

But the competitive logic driving all three companies is identical, and it suggests that this particular genie is not going back in the bottle. The chatbot, as a product category, was always a transitional form. It was the interface that made sense when AI could talk but couldn’t do much else. Now that it can do more, the interface is changing to match. The question for the rest of 2026 is not whether AI will become more agentic, but how quickly the tools will become reliable enough for users to trust them with real stakes.

For now, the honest assessment is that these products are impressive in their ambition and uneven in their execution. Frontier is available to a handful of Fortune 500 companies. Opus 4.6 is genuinely more capable than its predecessors but still operates within the familiar limitations of large language models. Google’s Personal Intelligence acknowledges, right in its own documentation, that it “might incorrectly make connections between unrelated topics or not fully understand the context.” These are not finished products. They are bets, placed simultaneously by the three companies best positioned to make them, on a future in which AI is less like a search engine and more like a colleague.

Whether that future arrives in months or years, the direction is no longer in question.