Most organizations running AI pilots can point to some wins. A faster first draft. A summarized meeting. A report that used to take an afternoon and now takes twenty minutes. Those are real. They are also the smallest version of what AI can do, and most leaders know it. The harder question nobody has a clean answer to is: what do we actually hand over? Which workflows, which processes, which repeatable tasks should an AI be running instead of a person? That question has stalled more AI programs than any technical limitation. A new category of tooling is answering it directly, and the major platforms are all moving toward the same idea: show the AI how the work gets done, and it figures out the rest.
The Discovery Problem Has Always Been the Real Blocker
The conversation around AI ROI usually starts in the wrong place. It focuses on the model, the vendor, the subscription tier. Those are not the constraint. The constraint is organizational imagination, and that is a harder problem to solve than picking a platform.
Knowledge workers know their jobs intimately. They do not naturally think about their jobs as sequences of discrete, transferable steps. When you ask someone what they do, they describe outcomes, not workflows. They tell you they manage reporting, or they own QBR prep, or they handle onboarding. The workflow underneath those outcomes is invisible to them because they have executed it so many times it stopped feeling like a series of steps.
That gap between “I manage reporting” and “here are the seventeen steps I follow to produce the report” is exactly where AI adoption stalls. AI is capable of running the process. The process only exists in someone’s head.
What These Tools Actually Do
All three work from the same core idea: instead of asking a knowledge worker to describe their workflow in a format AI can use, observe the workflow happening and build the automation from what you see.
OpenAI Codex record and replay, launched in June 2026, is currently available on macOS. A knowledge worker performs the task the way they always have. Codex watches: what platforms they touch, what they pull, what they combine, what the output looks like. From that observation, it generates a reusable skill that can be triggered on demand or run on a schedule. Those skills are shareable, so one person’s recorded workflow becomes an asset the entire team can run.
Microsoft Skill Recorder is the equivalent for organizations in the Microsoft ecosystem. It is a desktop app that records your on-screen work session: the clicks, the app switches, the pages visited, and optionally your spoken narration describing what you are doing and why. The GitHub Copilot CLI then reconstructs that recording as a clear intent plus an ordered list of steps, and converts it into a reusable Skill or Automation that runs inside Copilot Studio, Microsoft Scout, or Copilot Cowork. The output is not a fragile screen-click replay. It generalizes from your one recorded example, so the agent understands the task well enough to handle variations.
Anthropic’s Claude Cowork added the same capability with “Record a Skill,” which shipped July 21, 2026 inside the Claude desktop app. The mechanism is identical: screen-record yourself performing the task, narrate what you are doing and why, and Claude converts that into a reusable Skill saved to your library. Available to Pro, Max, and Team subscribers.
The platform differs. The mechanism is the same across all three: do the work once in front of the AI, and the AI owns it from there.
The QBR Example Makes It Concrete
Quarterly business review preparation exists in almost every organization, and it is one of the clearest candidates for this kind of automation.
The task is pulling together reports from multiple platforms, CRM data, financial summaries, operational metrics, project status, formatting them consistently, and assembling them into a presentable package. Done manually, this takes hours. It requires logging into several systems, exporting or copying data, reformatting across different structures, and reconciling numbers that live in different places. It happens every quarter under deadline pressure. It follows the same logic every time.
Record it once. The AI learns which systems, which reports, which fields, which format. Every subsequent quarter, the skill runs the process. The person responsible for QBR prep shifts from executing the assembly to reviewing the output.
The time savings are measurable. The more important outcome is that the process becomes reliable and delegatable. It no longer depends on one person knowing all the steps. And once that skill exists, it can be shared with every team running the same process across the organization.
The Snowball Effect Is Real
The organizations that will get the most from this category of tooling are not the ones with the most sophisticated AI strategy. They are the ones that give people permission to try it on everything.
The first workflow you automate does something important beyond saving time. It shows you that automation is possible for things you assumed were too complex, too judgment-heavy, or too dependent on human context. That realization changes how people look at everything else they do. The question shifts from “can AI do this?” to “have we tried it yet?”
That shift is the snowball effect. It does not come from a top-down AI strategy. It comes from frontline AI champions in departments who are empowered to record, test, and share. The QBR example becomes fifteen more examples. The skill one person built becomes the template others build from. The organization that encourages that behavior at scale finds more ROI than it was looking for.
The teams that stay cautious, that wait for a defined use case before experimenting, stay where they are. Capability is not the limit. Appetite is.
The ROI from AI programs has been smaller than it should be not because the technology is not ready, but because organizations have not solved the discovery problem. This category of tooling does not require a new strategy. It requires a different starting point: let your people show the AI how the work gets done, and let the machine take the volume from there. The capability exists today across the platforms most organizations are already paying for.
What workflows in your organization would you capture first? And what has been the biggest barrier to getting your teams experimenting with automation at this level?
Discussion
Comments are powered by GitHub Issues via utterances. A GitHub account is required to comment.