Computer-using agents marketed as 'personal assistants'

What AI 2027 Predicted

The scenario describes mid-2025 as the era when computer-using agents are heavily marketed as “personal assistants.” The framing emphasizes consumer-friendly tasks: ordering food, managing spreadsheets, browsing on your behalf. However, the scenario also predicts that adoption would be limited — the agents are “impressive in cherry-picked examples, but in practice unreliable,” and most people would not find them worth the hassle for everyday tasks.

How We Track This

We monitor:

• Launches of computer-use agents from major labs (OpenAI Operator, Anthropic Computer Use, Google Project Mariner)
• Marketing language and positioning of these products
• Consumer adoption metrics and usage data
• User reviews and reliability reports
• Computer-use benchmark scores (CUB, WebArena, OSWorld)

Current Evidence

This prediction has been confirmed. Multiple labs launched computer-using agents in 2025, marketed with exactly the personal-assistant framing AI 2027 described:

OpenAI Operator (January 2025): OpenAI launched Operator as a computer-using agent available to ChatGPT Pro subscribers. Marketing emphasized tasks like booking restaurants, ordering groceries, and filling out forms — classic personal assistant positioning.

Anthropic Computer Use: Claude’s computer use capability was released in beta in October 2024 and expanded through 2025, enabling the model to control a computer desktop to complete tasks.

Google Project Mariner: Google launched its computer-use agent through the Gemini ecosystem, with similar personal-assistant-style marketing.

Adoption remains limited: Consistent with AI 2027’s prediction, adoption has been tepid. Computer-use benchmark (CUB) scores remain low — the best agents achieved around 10% on comprehensive real-world task suites as of early 2026. User reports consistently describe agents as impressive in demos but frustrating in practice. The AI Futures Project graded the “agents impressive but unreliable” prediction as “broadly accurate.”

The marketing-reality gap: Agents are marketed for tasks like “order me a burrito on DoorDash” but frequently fail on multi-step real-world tasks involving authentication, dynamic web interfaces, or error recovery. This gap is consistent with the scenario’s depiction of the period, though it’s worth noting that a marketing-reality gap in new technology is a common pattern, not a uniquely prescient prediction.

Sources:

• Agentic Computer Use: Ultimate Deep Guide 2026 — o-mega.ai
• 2025-2026 AI Computer-Use Benchmarks & Top AI Agents Guide — o-mega.ai
• Grading AI 2027’s 2025 Predictions — AI Futures Project

Counterevidence & Limitations

• Coding agents (Claude Code, Cursor) have seen much stronger adoption than consumer personal-assistant agents — the scenario somewhat understates the developer-tool path
• Some agents may be gaining traction in narrow enterprise use cases that don’t match the consumer framing
• The “personal assistant” marketing may be shifting toward “coding assistant” or “work agent” as companies realize consumer adoption is limited
• It’s hard to measure adoption precisely — lab revenue numbers conflate agent and chatbot usage

What Would Change Our Assessment

• Already confirmed. Would strengthen with quantitative adoption data showing low consumer agent usage relative to marketing spend
• Partial revision: If computer-use agents achieve high reliability (>50% CUB scores) and mainstream adoption, the “low adoption” part of the prediction would weaken, though the marketing framing would remain confirmed

Update History

Date	Update
2025-05	OpenAI Codex (May 15) launched with “works for you asynchronously” framing. Claude Code marketed around enterprise productivity. Google I/O (May 20) positions Gemini and Astra as personal and work assistants. The “personal assistant” framing is becoming universal.
2025-10	OpenAI launches Operator; Anthropic ships Claude Computer Use; Google releases Project Mariner. All positioned as personal assistants for computer tasks.
2026-03	Marketing framing matches prediction precisely. Adoption remains low — CUB benchmark scores ~10% — but the product category is established as predicted.