Invoice preparation
Draft an invoice from approved order and delivery records.
Compare amounts, tax treatment and references with the source records. Route missing or conflicting evidence to the billing owner before issue.
Answers
What percent of tasks can AI automate?
There is no single percentage for every enterprise. 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.
From The Agentic Enterprise (2026), co-authored by Giridhar LV, Kashi KS, and Rajan. Available on Amazon Kindle.
A planning reference for examining work
Use the categories to examine candidate responsibilities, then test the proposed changes.
The book’s 30/40/30 framing
30 / 40 / 30
Automate / augment / human. Use these categories to discuss a proposed division of work. They do not measure a reduction in workload or headcount.
The study discussed in The Agentic Enterprise (2026) covers 2,143,500 tasks, six industries, 60 occupational roles and 11 companies. Its 30/40/30 framing should not be combined with the wider job-posting catalogue or the separate O*NET occupational analysis. This page does not provide the task-level records needed to reproduce that study.
Read the published study context and classification framework. The book citation above identifies the publication. For a project decision, examine your own records, tools and controls rather than adopting its proportions as a target.
The decision changes with the work
Illustrative design examples. These are proposed responsibilities and checks, not deployed customer results.
Bank customer onboarding
Assemble identity-check results and draft a risk memo.
An authorized reviewer resolves screening matches, missing evidence and the onboarding decision. A drafted memo does not authorize an account action.
Website quality assurance
Propose or execute tests for an agreed user journey.
Compare behavior with executable assertions, known defects and expected permissions. Engineers review missed defects and changes to the tests before accepting a release.
Maintenance planning
Summarize asset history and propose a work priority.
Qualified owners check inspection evidence, uncertainty, permits and safety controls before authorizing field work.
Within each example, different steps may need different treatment. An agent might prepare a recommendation while a person decides. A rules-based service might execute an approved action. Missing evidence or an unusual case may require a stop and review. Record these distinctions before choosing an automation target.
Common questions
What the figures can and cannot tell you.
No. It is a planning reference used in The Agentic Enterprise (2026), not a target for every industry, role or enterprise. Classifications can change with the available tools, records, rules and consequences of error. An uncertain statement may also need clarification before assigning it a category.
No. A job description tells you about stated responsibilities. Classifying it proposes how AI might contribute; it does not show whether a system is deployed, what it can reliably execute or how much time it saves. Those questions require workflow evidence and observed results.
They may count different things: occupational descriptions, job postings, workflow steps or conversations with an AI assistant. They may also use different definitions of automation and augmentation. Compare the population, dates, categories and unit of analysis before comparing percentages. AI conversations alone cannot measure work for which people did not use the assistant.
No. Software may execute an agreed step while a person or team remains accountable for the workflow. Identify who can approve an action, what the system must stop or escalate, and how the result will be checked. A classification alone does not establish safe autonomous operation.
A task percentage does not establish a staffing decision. Work may be removed, redistributed or expanded, and new review or operating responsibilities may follow. Those changes depend on the organization’s choices and demand for the work. Task analysis cannot guarantee a particular job outcome.
Document the work, volumes, exceptions and current baseline with its owners. Identify candidate steps for software execution or AI assistance, then test representative cases using approved records and tools. Review correctness, human effort and failures before deciding what to change. Discuss enterprise-specific analysis and preparation with Nuvepro.