How to Classify Tasks in a Workflow for AI
Start with actual work, such as generating an invoice. Identify the records, decisions and handoffs, then propose where software, AI and people contribute. This guide shows how to turn that task map into a practical enterprise project and checks.
The 5-Step Methodology
Document the workflow. Break roles into tasks. Classify. Redesign. Prepare.
The five steps below connect the work to a proposed design and practical checks. Start with the workflow, document role contributions, assess task options, redesign the handoffs and prepare people to build and run the agreed work.
Document the Workflow
A workflow is a series of tasks performed in order to achieve an outcome. Before classifying tasks, identify and document the workflow. This gives every task a home and makes the classification actionable.
Example: Invoice Generation Workflow
- 01Pull customer data from CRM
- 02Calculate amounts owed
- 03Generate invoice documents
- 04Route for approval
- 05Send invoices to customers
- 06Log payments received
6 tasks in this workflow. Each will be classified independently.
Audit: Decompose every role into tasks
Start with a workflow and observe the work people actually do. Break each role’s contribution into tasks with an input, an action and an output. Compare job descriptions with records, workflow logs and interviews; the right level of detail depends on the decisions you need to make.
Example: Financial Analyst
- 01Compile monthly revenue reports from 3 ERP systems
- 02Build variance analysis models in Excel
- 03Present quarterly forecasts to department heads
- 04Review vendor invoices for budget compliance
- 05Draft board-ready financial summaries
Five illustrative tasks, not a complete task count for every financial analyst. Check the list against the selected team’s actual work.
Reference inputs can include occupational data, job postings, documented processes, task libraries and owner interviews. Generated decomposition is a proposal to check against actual work. The published posting corpus was approximately 5M task records at the 18 June 2026 provenance review; it is not 5M independently validated automation decisions.
Classify: Assign each task to a category
Use three planning categories: a candidate for automation, work supported by AI, or work retained by people. Assess each in its workflow, including missing information, exceptions and consequences of error. A label records a design judgment to test; it does not establish that an agent can perform the task.
Automate
A candidate for software execution within agreed rules, permissions and checks. An accountable owner still decides operating use.
Signal: Repetitive, rule-based, data-heavy, low ambiguity
"Compile monthly revenue reports from 3 ERP systems"
Classify against the selected workflow
Augment
A person uses AI to prepare, investigate or draft part of the task and reviews the result against relevant evidence.
Signal: Requires context, some judgment, or stakeholder interaction
"Build variance analysis models (AI drafts, human validates assumptions)"
Classify against the selected workflow
Human-Only
Work assigned to a person because its judgment, physical requirements or accountability should remain human in the proposed design.
Signal: High ambiguity, interpersonal, novel situations, physical tasks
"Present quarterly forecasts to department heads"
Classify against the selected workflow
Redesign: Rebuild the workflow from the task up
Use the task map to design who does what, which tools they use and how the result is checked. For an invoice workflow, software might prepare a draft while a billing owner resolves missing order evidence and approves issue. Define exceptions and handoffs before building.
Illustrative starting point
Illustrative starting point: a billing analyst copies approved order data into an invoice draft and investigates missing references.
Proposed change
Proposed change: software prepares the draft and identifies missing fields; the analyst checks evidence and resolves exceptions. The billing owner approves issue under the agreed rules.
Prepare: Build and practise the changed work
Prepare people for their actual responsibilities in the redesigned work. Some will configure rules or build integrations; others will investigate exceptions, test outputs or approve decisions. Practise in an environment that represents the relevant enterprise systems and records. Agree what it omits and how readiness will be assessed.
Work with AI
Supervision, validation, quality control, exception handling
Build with AI
Configuration, prompt engineering, workflow design, agent management
Four mistakes to avoid when classifying work
And how to avoid them.
Classifying at the job level instead of the task level
A data scientist may prepare data, compare models and explain a decision to a business owner. These tasks have different evidence, authority and review needs. A single label for the job hides those differences.
Decompose to tasks first. Classify each task independently.
Relying on one source of evidence
Interviews reveal exceptions and undocumented work, but recollection alone can miss steps. Postings describe expected responsibilities, not necessarily the live process. Compare both with approved records and observed cases.
Use source records and owner interviews to define the work, then test the proposed AI contribution on representative cases.
Treating classification as a one-time exercise
A policy change, new tool or different input can invalidate an earlier classification. Review the assumptions when the work changes, and retain the prior version so the reason for a new decision is visible.
Agree a review cadence and event triggers with the workflow owner. Separate currently tested capability from a proposed future design.
Stopping at the classification report
A spreadsheet showing 'these 12 tasks can be automated' changes nothing if the workflow, the tools, and the training don't change.
Continue from task labels to workflow design, practical building and evidence of what people and agents can do.
Illustrative task mixes for discussion
The percentages below are worked examples, not measured industry automation rates.
Use these example mixes to discuss how a classification changes when work and constraints differ. They do not establish what any industry can automate. For a project, prepare a task inventory and record the evidence behind each category.
| Industry | Automate | Augment | Human-Only |
|---|---|---|---|
| Healthcare | 22% | 55% | 23% |
| Financial Services | 31% | 52% | 17% |
| Manufacturing | 19% | 48% | 33% |
| Technology | 35% | 50% | 15% |
| Retail | 28% | 49% | 23% |
| Consulting | 26% | 56% | 18% |
These illustrative mixes are separate from the published posting corpus and the book’s six-industry framework. Explore the full dataset →