Answers

Automate, augment, or human?

Use automate, augment and human-only as planning choices for a defined task. The choice depends on the tools, data, required review and who may act. Treat it as a proposal: test representative cases and agree the boundaries with the workflow owner before deployment.

By Giridhar LV·Founder & CEO, Nuvepro. Author of The Agentic Enterprise.·June 19, 2026·6 min read

From The Agentic Enterprise (2026), co-authored by Giridhar LV, Kashi KS, and Rajan. Available on Amazon Kindle.

The three buckets

Propose a category for a clearly defined task, then review it against the actual tools, rules and evidence.

Automate

AI executes the defined task within approved boundaries. Verify outputs, exceptions and downstream effects before assigning it.

Candidate task examples

Reconcile a balance sheet
Draft a first-version contract
Update a forecast model with last week's numbers
Summarize a long document into a brief

Augment

AI drafts or analyzes; a person retains the identified decisions or approvals. Specify the handoff and the evidence that person needs.

Candidate task examples

Present scenarios to a CFO
Negotiate a contract with a long-standing customer
Choose between two architectural approaches
Update a forecast based on a business change

Human-only

Keep the responsibility with a person where policy, trust or the nature of the interaction requires it. Revisit the decision when the work or requirements change.

Candidate task examples

Hold space for a patient in crisis
Tell a team that layoffs are coming
Own a high-stakes judgment call with real consequences
Sign off on a decision the board wants a human accountable for

Why three buckets, not two

A two-bucket split tells you how AI and humans share work. It does not tell you what must stay human.

Anthropic's January 2026 Economic Index reported augmentation in 52% and automation in 45% of sampled Claude.ai conversations from November 2025. Those are interaction patterns within that sample, not the proportion of all enterprise work that AI can perform.

Enterprise planning also needs to identify work that remains with people. A sample of AI conversations cannot measure tasks where AI was never used. Decisions about responsibility require the workflow context, including approvals and operating rules.

The categories are a planning framework. Their application can change with the task boundary, model, tools and business rules. Use the classification to propose trials; evaluate the observed work before assigning responsibility.

Common questions

How to use the classification in practice.

Automate proposes that AI executes the defined task within approved permissions and checks. Augment proposes that AI contributes while a person retains specified decisions or approvals. Define the task boundary carefully: producing a draft is different from approving or issuing it.

Review the inputs, required judgment, decision authority and cost of error with the workflow owner. Propose a category, then test it using the actual tools, data and exceptions. Structured inputs can make a task a useful automation candidate, but they do not prove the agent can perform it correctly.

The 2023 Harvard/BCG Jagged Frontier experiment found that AI effects varied by the consultant task tested. Classification can help identify what to try and who should review it. It does not locate a model’s capability frontier automatically; representative trials and independent checks are still needed.

Yes. A different model, tool, dataset or operating rule can change the proposed category. Re-test the affected cases before changing responsibility. A job-posting classification is not a measurement of deployed automation, and there is no fixed current-generation ceiling that applies to every enterprise.

It makes retained human responsibilities visible, including decisions, relationships and approvals. Its size depends on the work and requirements. People also remain accountable for reviewing and operating work in the other categories.

Name a trigger, input, observable output and responsible actor. If parts of the activity need different permissions or review, split them. Preparing a daily cash-flow report is a task to examine; Excel proficiency is a skill, not a directly executable task.

What percent of tasks can AI automate?/Augmentation vs automation/The 30/40/30 pattern

See the three buckets for a real role.

Explore public role and workflow patterns in the data. Discuss the actual responsibilities, approved tools and preparation scope with your organization; detailed analysis and preparation plans belong in the tenant workspace.