Zeta Labs launches AI employee Viktor in Microsoft Teams

Viktor's core claim is that it does not draft, suggest, or summarize — it executes. Where most workplace AI generates text, Viktor maintains a persistent workspace, connects to a company's actual tools, and performs actions end-to-end ^[1][2]. A request like 'create a project brief from last week's meeting notes and share it with the marketing team' is meant to result in a finished, shared brief rather than a draft handed back for review ^[2]. Each instance runs on its own persistent cloud computer where it writes and runs code to produce PDFs, dashboards, and web apps. The strategic bet, endorsed by lead investor Accel, is that owning the execution layer for an entire team — not assisting one individual — is where the next tens of billions of dollars of AI adoption sit ^[3]. Crucially, Viktor goes to where work is already discussed: it lives natively inside Slack and now Teams channels, converting in-channel conversation into completed tasks rather than asking teams to adopt yet another app or dashboard ^[1].

By launching inside Teams, Viktor competes with Microsoft Copilot on Microsoft's home platform. The two products embody opposing philosophies of workplace AI: Copilot drafts content and offers suggestions under human oversight, while Viktor positions itself as an autonomous agent that takes delegated workflows and finishes them ^[2]. On Teams, Viktor integrates with the Microsoft 365 stack — SharePoint, Outlook, Planner, To Do, Power Automate — alongside third-party SaaS such as Salesforce, Jira, and Workday, and ships SOC 2 Type 1 certification ^[1][2]. Because an autonomous agent taking real actions carries far more risk than a text generator, Viktor wraps execution in enterprise governance: admins define permitted actions, restrict sensitive channels, require human approval for high-risk tasks, and log every interaction, with customer data excluded from public model training ^[2]. The framing is explicitly a 'hire' rather than a tool, aimed at operations teams, agencies, and growing businesses that want automation without adding headcount ^[1].

The promotional layer around Viktor is loud and bullish — hands-on creator accounts describe adding it to Teams, asking it to pull email numbers and draft campaigns, and watching it execute in-channel, and the founder frames it as a teammate handling roughly ten-minute tasks autonomously. But much of that buzz appears to be part of a paid creator program, and the most substantive critical discussion sits elsewhere. Early community feedback centers less on capability than on usage economics: the recurring complaint is credit burn, where connecting apps, re-reading environments, and running Viktor on a schedule consume credits quickly. The sharpest version of the critique argues this cost is structural, not a tuning issue — tools in this tier tend to re-ingest context from scratch on each run rather than diffing against cached state, because a true checkpoint or caching layer is technically hard and often skipped to ship fast. With pricing at a free tier carrying $100 in credits and a $50-per-month team plan for 20,000 credits ^[4], that consumption model is the load-bearing question for buyers weighing the autonomous-employee pitch against monthly cost.

Beneath the product framing sits a genuinely difficult engineering problem that explains why a channel-native agent is non-trivial. The founder describes the chat surface as a deceptively complex input: threads, direct messages, edits, deletions, and emoji reactions all carry meaning an agent must interpret correctly. Harder still is memory isolation when a single Viktor serves hundreds of users across many channels and DMs — context from one conversation must never leak into another. The team also reports that swapping the underlying model for a cheaper one produces user-noticeable regressions, which constrains how aggressively the economics can be optimized. Independent hands-on comparisons frame Viktor as the autonomous end of the spectrum — working for you rather than with you — and credit it with built-in memory that learns a business over time, the same persistent-memory capability that makes the multi-user isolation challenge so central to the product.