The 30/40/30 pattern: a starting point for examining work
Before deciding what to automate, examine what people actually do. The useful question is which part of a workflow could change, what would improve, and who would check the result.
From The Agentic Enterprise (2026), co-authored by Giridhar LV, Kashi KS, and Rajan. Available on Amazon Kindle.
Updated 18 October 2026 to clarify the study scope and distinguish task classifications from deployment evidence.
What the pattern describes
The Agentic Enterprise (2026) uses a 30/40/30 framing for automation candidates, human and AI collaboration, and human-led work, alongside its study of 2,143,500 tasks across six industries, 60 roles and 11 companies. Use the framing as a planning reference, not a measured deployment rate or a target for your workflow.
Use the categories to ask better questions. A proposed automation needs accessible inputs, permitted actions and a result that can be checked. Augmented work needs a clear review or decision responsibility. Human-led work may still use supporting tools while its owner retains the decision.
The public task library and occupational reference data have their own sources and dates. A model-generated classification of a job description is a research hypothesis. It is not an observation that a company has automated that task, a prediction of future capability, or permission to change staffing.
Read the study context and Task Intelligence frameworkUsage evidence answers a different question
Anthropic’s first Economic Index report, published 10 February 2025, analyzed a sample of Claude conversations mapped to occupational tasks. It describes observed product use. The report also notes that it cannot tell how people used the resulting output afterward and does not treat its sample as representative of all AI use.
That is useful evidence about what people ask AI to help them do. It does not supply the missing activities, approvals or exceptions in a particular enterprise. Do not combine a usage percentage with a task-classification percentage as though they measured the same thing.
Read Anthropic’s original report and limitationsWork through the actual decision
These illustrative examples show why a label is only the beginning. They are proposed designs, not measured customer results.
Prepare an invoice
Extract approved order details and calculate a draft.
Reconcile quantities, prices and tax rules. Route missing or conflicting records to the billing owner before issue.
Review a contract
Find clauses that differ from the approved playbook.
Check the cited text and missed clauses. A legal reviewer decides the consequence and proposed response.
Test a website
Have an agent exercise agreed journeys and report failures.
Reproduce the failures, check for missed seeded defects and review false positives. Release approval stays with the application owner.
Turn the proposal into a project
Pick work with a clear owner and a result worth improving. Document the current inputs, decisions, handoffs and exceptions. Record a baseline that includes quality and review effort as well as elapsed time.
Then scope a simulation around the relevant enterprise records, tools and rules. Your people can build or configure the solution and test it on representative cases. Existing agents can enter directly for evaluation. Review human capability, agent performance and the combined workflow separately; the operating owner decides what evidence is sufficient for use.
The result may involve automation, assistance or a decision to keep part of the work with people. The project should establish which approach works under the agreed conditions, rather than force a predetermined percentage.
Explore practical guides and published workflow examples