ChatGPT Can Find the Work It Should Be Doing for You

Summary

The presenter argues that businesses should not start by asking AI, “What should I automate?” That broad question can lead to overengineered solutions, such as custom code or apps, that take time to build and maintain. Instead, start with evidence of the work people already do, then use existing AI features, such as skills or projects, where possible.

The recommended approach is to map work from role → activity → task → steps, including each task’s inputs, sources, decisions, and expected output. For example, a CFO’s reporting activity might include writing weekly reports, with specific steps and inputs.

Recommended Workflow

• Find recurring work from existing data. Ask AI to analyze sources such as calendars, sent email, Teams chats, or exported ChatGPT and Claude conversations. The presenter suggests reviewing the previous 90 days and gives an example of analyzing 10,000 emails.

• Require evidence, not just suggestions. For each recurring task, ask AI to provide:

  • A one-line task description
  • How often it appeared and on which dates
  • Specific supporting evidence, such as an email or calendar event If evidence is insufficient, AI should say so.

• Limit and rank the results. Request no more than 10 tasks, ranked by frequency. Have AI identify and describe work—not recommend what to automate.

• Choose a simple starting task. Assess each candidate based on:

  • Number of process steps
  • Number of information sources
  • Number of value judgments required Start with the task that has the fewest of these factors, then take on more complex tasks later.

• Document the process through an AI interview. Provide examples of good work, explain the task, and have AI ask one question at a time, using dictation if helpful. Limit the interview to 15 questions. The output should be a one-page specification covering steps, sources, decisions, and quality criteria.

• Test collaboratively before formalizing. Give the specification to AI in a new chat and complete the task together at least once. Once the result is satisfactory, ask AI to turn the conversation and its nuances into a reusable skill or project.
• Improve through review and feedback. Check AI’s decisions, correct mistakes, and refine the skill over repeated use. The goal is to reduce errors over time and eventually delegate the task with less supervision.

Metrics and Figures Mentioned

• 90 days: Suggested period for analyzing calendars and sent emails.
• 10 tasks: Maximum number to return in the initial ranked list.
• Top three: An option if prioritizing by frequency alone; the presenter otherwise recommends evaluating the three complexity factors.
• 1–2 tasks: Suggested number to focus on for AI assistance.
• 15 questions: Maximum length for the process interview.
• One page: Target length for the resulting specification.
• 10,000 emails: Example dataset size for analysis.
• “95%” of desired AI use cases: The presenter’s estimate that most can be handled with features already in the AI product rather than custom applications. This is presented as a general claim, not a measured KPI.

Practical Takeaway

Treat AI adoption as process discovery and gradual delegation: identify repeatable work from real records, document how it is done, start with a low-complexity task, test it with AI, and improve the resulting skill through feedback. Avoid custom software unless simpler built-in capabilities do not meet the need.

Presenter/source: Dylan Davis, as identified by the video channel. He says he runs an AI consulting company.