Is automating a repetitive desktop task worth it?
For small digital businesses that automate work inside older Windows programs for clients. Compare a fast scripted path with a slower AI agent using your own timings, see the hours saved per month, and generate safety checks for the automation. Inspired by the System 1 agents paper (see Sources); compiling past runs into scripts is our extension, not the paper's method.
EXAMPLE ASSUMPTIONS · time both paths on your own workflows and replace these
EXAMPLE RESULTS · illustrative, computed from the assumptions on the left. Not a benchmark.
How the example math works: blended time = share × fast + (1 − share) × slow. Hours saved = tasks per day × (slow − blended) × working days ÷ 3,600. The worker count (physical cores ÷ 2) is a starting point for load testing, not a measured capacity: measure time-to-first-audio on the client's own hardware.
What this leaves out: the time to build and maintain the fast path, screens that change and break it, the cost of an error that slips through, and licensing of the legacy software. Time both paths on the client's real workflows before quoting hours saved.
Sources & what we checked
- System 1 agents (arXiv 2611.21053, Tzanis and Klontzas): local agents that call tools without generating text. A deterministic compiler turns each tool's schema and the request into multiple-choice questions. On 1,952 requests to 157 biomedical tools, the best System 1 agent chose the right tool (or correctly declined) for 97.0% of requests, vs 95.2% for the best of 18 local LLM-agent settings. A request took 1.2–10.1 s with System 1 vs 2.6–26.1 s for LLM agents with reasoning, on one 48 GB workstation GPU; LLM agents with reasoning off took 1.1–9.2 s, about the same. The paper's main win is reliability (no System 1 call contained invented content, vs 5.5–19.7% of LLM calls); the speed gain is modest.
- Launch HN: Cyberdesk (YC S25, 2025-08-14): describes itself as "a deterministic computer use agent for automating Windows desktop applications".
- Low-latency CPU voice assistant (GitHub README): reports 1.25 s time-to-first-audio streaming into KittenTTS (kitten-tts-mini-0.8), vs about 1.8–2.5 s for sentence-based setups. It cites 32-core ONNX parallelization tuned for high-core Windows machines, and its setup asks for an OpenRouter API key for the LLM, so it is not a fully offline stack. We have not tested it on budget mini-PCs.
- Task volumes, fast-path share, per-task seconds, working days and core counts on this page are example assumptions.