Role-to-Agent Automation Decomposition
Break a role into tasks, score automation potential, then prototype one workflow
- Difficulty
- Advanced
- Time to result
- ~weeks to results
- Steps
- 7
- Confidence
- 94%
Treat a job as a portfolio of recurring tasks instead of asking whether the entire role can be automated. First, use domain knowledge and an AI model to enumerate responsibilities and split them into automatable and manual categories. For each automation candidate, assess how confidently AI can perform it, what gaps remain today, and when stronger models may close those gaps. Select one bounded workflow with useful inputs and a measurable output. Turn that candidate into a project brief specifying data, steps, controls, and success criteria, then give the brief to an appropriate implementation model or coding system. Evaluate how far the prototype gets before extending the approach. This produces concrete evidence about augmentation and automation without relying on sweeping percentage claims.
Origin
Kieran Flanagan outlined this task-decomposition and prototyping process while preparing to test how much of individual marketing roles AI agents could automate on Marketing Against The Grain.
Core principles
- 01Roles are collections of tasks rather than indivisible jobs.
- 02Automation potential differs substantially between tasks.
- 03Manual and automatable work should be separated before building.
- 04Confidence and time horizon matter alongside technical possibility.
- 05Each candidate needs a concrete project brief before implementation.
- 06Prototype one workflow before attempting to automate an entire role.
How to run it
- 1
Select one role
Choose a role with a recognizable set of recurring responsibilities. Narrow the scope to one team and operating context.
Pro tip Start with a role whose outputs are already reviewed or measured.
Watch out The same job title can contain different work in different organizations.
- 2
Decompose the work
List responsibilities, then break each responsibility into discrete tasks with clear inputs and outputs. Validate the list with someone who performs the role.
Pro tip Include coordination and judgment tasks, not just visible production work.
Watch out An incomplete task inventory will exaggerate the percentage that can be automated.
- 3
Classify each task
Separate tasks AI can plausibly automate from those requiring human judgment, accountability, relationships, or physical action. Allow hybrid classifications where AI drafts and a person decides.
Pro tip Evaluate task stages separately because research may be automatable while approval is not.
Watch out Avoid binary thinking about an entire role.
- 4
Score confidence and timing
Estimate how well AI can perform each candidate now and the likely time frame for dependable automation. Record current gaps and required safeguards.
Pro tip Use confidence scores tied to testable quality criteria.
Watch out Capability forecasts are hypotheses, not guarantees.
- 5
Choose a bounded workflow
Prioritize a task with accessible data, repeatable steps, measurable output, and limited downside. Brand monitoring is one example discussed in the episode.
Pro tip Select a workflow where human review already exists.
Watch out Do not begin with a high-risk customer-facing decision.
- 6
Write the project brief
Define the workflow, inputs, processing stages, outputs, integrations, review points, and acceptance tests. Make the brief specific enough for an implementation system to act on.
Pro tip Ask the model to generate the first brief, then correct it with domain expertise.
Watch out A generated brief can omit operational constraints that were not supplied.
- 7
Prototype and evaluate
Give the brief to a capable coding or agent-building model and construct one end-to-end workflow. Test output quality, reliability, and human time saved.
Pro tip Measure the prototype against the existing manual process.
Watch out A single successful run is not evidence of dependable automation.
In the wild
Kieran asked Claude to break brand marketing into subtasks, including guidelines, content, sentiment monitoring, personalization, partnerships, and performance analysis. The model separated automatable and manual tasks. He proposed adding confidence and time-frame scores, selecting brand monitoring, generating a project brief, and passing it to Gemini Advanced for implementation.
→ A broad question about automating marketers became a specific workflow that could be prototyped and tested.
A revenue team decomposes sales operations into data cleanup, routing, forecasting, process design, exception handling, and stakeholder coordination. It selects duplicate-record resolution as a low-risk candidate, writes a project brief, and builds an agent that proposes merges for human approval.
→ The team measures time saved and error rates before considering more consequential automation.
Common mistakes
Automating the job title
Asking whether a whole role is automatable hides the major differences between its production, judgment, coordination, and accountability tasks.
Ignoring confidence and time frame
A task that may eventually be automatable is not necessarily safe or useful to automate with today's models.
Starting with the whole role
Attempting an end-to-end role replacement creates too many failure points to diagnose and measure effectively.
Is it for you?
Best for
Leaders and operators evaluating how AI agents could change a specific marketing or business role.
Not ideal for
Organizations seeking an instant workforce-replacement verdict without documenting the actual work people perform.
From the transcript
“I asked it to break down individual marketing roles”
“it breaks into automatable tasks and manual tasks”
“under each automatable automatable task have it create a project brief of how would automate that and then plug that into Gemini advaned”
From the episode
How This GPT-4 Prompt Is Breaking a $257 Billion Industry