Repetitive tasks at your screen, filing an expense claim, downloading a recurring report, creating a ticket with fixed fields, add up to a lot of time. OpenAI's Record & Replay and Claude's equivalent (cowork) promise that you demonstrate a task once, after which the AI carries it out itself from then on. No hype: let's look at what it really is and where it pays off today.
What is record-and-replay?
Record-and-replay means literally: record and play back. You perform a task once, by hand, while the computer watches. From that recording, the AI distills a reusable skill: a step-by-step plan it can then run on its own, each time with different input. Where a classic script breaks the moment a button moves, an AI agent reasons along and adapts. It's macro recording, but with an understanding of what you're trying to achieve.
How does OpenAI's version work?
Record & Replay lives in the ChatGPT desktop app and in Codex. You open Plugins, choose "Record a skill", approve the recording permissions, and simply perform the task on your Mac. The app observes your actions and the window content needed to learn the workflow. When you stop the recording, the system inspects the captured steps and writes a draft skill: when to use it, what input it needs, which steps follow, and how to verify the result. After that, in a new chat, you simply ask it to run that skill, with the variables for that specific case.
- Examples from the docs: filing expense claims, booking parking, creating pre-configured issues, publishing videos, and downloading recurring reports.
- Under the hood it relies on Computer Use plus browser actions and connected plugins.
- Limitations: macOS only, Computer Use has to be enabled, and the initial rollout excludes the EEA, the UK and Switzerland, relevant for European users.
- Organizations can turn it off via requirements.toml with computer_use = false.
Seeing it work
A short demonstration says more than an explanation. The video below shows recording and replaying a workflow in practice.
And the equivalent in Claude: cowork
Claude has a similar idea under the name cowork: you work together with the agent on a task, after which it captures the pattern and repeats it. Where OpenAI leans heavily on literal screen recording on the Mac, Claude tends to focus more on describing and capturing a reusable procedure, a skill the agent runs with real tools, including the browser environment and connected systems via MCP. Either way, the core idea is the same: capture a task once, then repeat it on request with new input.
The difference isn't in the magic, it's in the trust: a recorded skill is only valuable once you're willing to let it run unsupervised.
Where it pays off for small businesses
For a small business, the biggest win comes from tasks that are stable, boring and frequent. Think of retyping orders between two systems, pulling weekly reports, or creating standard tickets. Our approach stays grounded:
- Start with a task you do exactly the same way, every single time.
- Demonstrate it with realistic but non-sensitive data.
- Have a human check the first runs before you let it loose.
- Watch the regional restriction: OpenAI's version isn't fully available in Europe yet, Claude cowork is often the practical alternative there.
In short
Record-and-replay turns a one-time demonstration into a reusable skill. It's no miracle cure, but for repeatable screen tasks it's one of the most concrete time savings AI offers today. Want to know which of your processes are a good fit for this? We're happy to think it through with you, no hype, just results.
