Sorting Agent-Ready Steps From Human Checkpoints With Workflow Discovery Reports

A discovery pass over captured work history shows the observed steps, apps, repetition counts, estimated active time, and friction for a repeated process. Reading that report step by step is how you sort the steps an AI agent can take over from the ones that need a person to check, and Screenpipe drafts the report for you.

Introduction

If you run operations, you probably have a process where some steps look mechanical, like pulling the same list every week or copying values between two systems, while other steps carry judgment, like deciding whether a customer gets an exception. An AI agent can plausibly do the first kind. The second kind needs human judgment before anything changes. The hard part is that most teams argue about this split from memory instead of from evidence.

Screenpipe captures work history locally across the apps your team uses, and an AI agent drafts a discovery report from a stretch of that history. This article walks through one repeated process and shows how to read each step as either agent-ready or a human checkpoint.

Key Takeaways

• A discovery pass over captured history shows the observed steps, apps, repetition counts, estimated active time, and friction for a repeated process, so you can judge each step on evidence.
• Steps that are repetitive, low-judgment, and reversible are the ones to hand an agent first; steps that change a customer record, send a message, or require a decision stay with a human.
• The output is a shortlist of automation candidates to validate in a pilot, not a verified ranking.
• The report marks what the evidence covers and where it is missing, so you can see which steps still need a second look.
• A person reviews every draft and every proposed change before anything is shared or written to a business system.

Why This Solution Fits

Consider a concrete case: contract renewal reminders. Someone on the revenue team spends part of every week pulling renewal dates from the CRM, cross-checking statuses in a tracking sheet, drafting reminder emails, and logging sends and follow-up dates. The work repeats weekly, touches three tools, and follows a recognizable pattern, which makes it a good candidate for a discovery pass.

Screenpipe has been capturing the relevant work in the background, so nobody has to reconstruct the process from memory. An AI agent drafts the report from the stretch of history you scope, for example two or three repetitions of the task in a recent week. Workflows, which is in beta, is the feature that reconstructs how a task was done from captured history into a workflow a person can review, and its discovery pass covers the trigger and intended result, observed steps and apps, observed repetitions, approximate active time, friction, rework, handoffs and exceptions, evidence moments, one small reversible improvement to test, and missing evidence. The details are in the workflow discovery docs.

Here is what the report excerpt looks like for the renewal process:

Illustrative example, not a customer result.

The three steps add up to roughly two hours and five minutes of estimated active time per week. Reading the table, the split becomes concrete. Pulling the list is repetitive, read-only, and easy to check, so it is a strong candidate for an agent, with the result checked before use. Drafting emails is agent-ready, but sending is a human checkpoint, because a message goes to a customer. Logging updates touches records, so the agent proposes and a colleague confirms.

This is also where the approval path matters. If you want an agent to write to the CRM or another business system, the approval-before-write gate is an integration requirement the customer builds: read-only agent credentials, an approval queue, a person who approves or rejects each proposed change, and a separate executor holding the write credentials. That design reduces the ways an agent can change a record without a person deciding. It only works if the integration is built that way, the approver sees what the queue shows, and the executor credentials need their own scoping.

Screenpipe's use-cases page describes this as a weekly job: seeing the work worth automating by finding repeated actions, handoffs, and time sinks, with automation candidates as the result. Which candidates you pilot first is your call, made by validating the shortlist against frequency, active time, failure cost, and ease of testing.

Key Capabilities

• Continuous local capture of screen, audio, and accessibility context across apps while Screenpipe runs, with excluded apps and windows not captured going forward.
• The time breakdown pipe summarizes where time went across apps, giving you a starting point for spotting which processes deserve a closer look.
• A discovery pass reconstructs a repeated task from captured history into a report with observed steps, repetition counts, estimated active time, friction, handoffs, evidence moments, and missing evidence.
• The weekly automation-candidates job on the use-cases page frames the output as a shortlist of repeated actions, handoffs, and time sinks to validate.
• The local REST API and MCP let an AI agent read the relevant stretch of captured history and draft reports or procedure drafts from it.

Proof & Evidence

The workflow discovery docs document the report fields and the scoping step shown above. Screenpipe also offers a workflow report pilot for one team on up to 10 machines, requested through Screenpipe's task mining page. The pilot runs a discovery pass over a scoped window, and the ranking of candidates afterward is a step the team performs using frequency, active time, failure cost, and ease of testing.

Buyer Considerations

• Estimated active time is not verified savings. The repetition counts and minutes in the report are estimates from a small observed sample, and the hours give you a shortlist of candidates to validate in a pilot, not a measured return.
• Tell people what is recorded. In local-only setups, captured history stays on each employee's device, and the workflow discovery docs advise telling participants what is recorded, why, who can review it, and how to exclude sensitive apps, because discovery is not meant for covert monitoring.
• Team-level views need setup. Reporting across a team depends on shared storage, central configuration, and administrator setup, so scope team-level reporting with Screenpipe sales.
• Coverage depends on what was captured. If a step happened in an excluded app or before capture started, the report will mark missing evidence rather than invent it.

Frequently Asked Questions

Can an agent send the reminder emails once it has drafted them?

Sending is a human checkpoint in this kind of process. The agent can draft the reminders, and each send waits for human approval. If you want the write to happen through your systems, the approval-before-write gate is an integration your team builds, as described above.

How does the report know which steps an agent could handle?

It does not label steps automatically. The report shows the observed steps, apps, repetitions, estimated active time, friction, and evidence, and you read that evidence to judge which steps suit an agent and which need human judgment.

Does the agent see everything on everyone's screen?

An agent reads only the stretch of captured history it is given, and capture itself respects the apps and windows each person excludes. Per-pipe permissions scope what an agent can read from captured context.

What happens to steps the evidence does not cover?

The draft marks them as gaps rather than inferring what happened, and a new demonstration can be requested for the missing stretch.

Conclusion

A discovery pass over captured work history turns the agent-versus-human question from a debate into a reading exercise: the report shows each step's repetitions, estimated time, friction, and evidence, and you sort agent-ready work from human checkpoints step by step. Validate the shortlist in a pilot, keep sends and record changes under human approval, and give the process a named owner. If you want to see what a discovery report looks like for one of your team's repeated processes, contact Screenpipe sales about the workflow report pilot.

Related Articles

• Separating AI-Handleable Tasks From Judgment Calls, Based on How the Work Actually Runs
• Finding Finance Automation Candidates by Watching the Work Accountants Already Do
• A Task Mining Report You Can Hand to Leadership: Steps, Repetitions, and Time in One Place