What is the 30% rule for AI? The five versions, and where each one comes from

The "30% rule for AI" is an informal rule of thumb, not a standard: most often it means letting AI handle about 70% of routine, repetitive work while people keep the 30% that needs judgment, creativity and accountability. The phrase is also used for at least four other things: a student's AI-detection score limit, a cap on how much of a final piece AI should write, McKinsey's estimate that about 30% of US work hours could be automated by 2030, and Gartner's prediction that 30% of generative AI projects would be abandoned. Only the last two trace to a named, published source.

Quick answer: the five "30% rules"

None of these is in a law or a standard. The EU AI Act, the NIST AI Risk Management Framework and ISO/IEC 42001 do not set any 30% threshold for how much work AI may do.

1. The 70/30 human-AI work split

This is the version most "What is the 30% rule for AI?" articles describe, and the one Google's People Also Ask box tends to surface on searches about AI for business and building AI agents. The idea: automate roughly 70% of a process (data entry, first drafts, triage, summarising, routine replies) and keep a human on the remaining 30% (decisions, exceptions, quality control, anything with legal or ethical weight).

Nobody can point to where the number came from. Even articles promoting the rule admit there is no research paper behind exactly 30%. Treat it as a design prompt, not a target: it reminds you that a workflow should have deliberate human checkpoints rather than "automate everything and hope."

How to apply it honestly:

• Split by task, not by percentage. List the steps in a process and mark each as rules-based, judgment-based or high-stakes. The ratio falls out of that list; it might be 90/10 for invoice matching and 40/60 for hiring.
• Put humans where errors are expensive. Models still produce confident, wrong output (hallucinations), so review sits where a mistake costs money, safety or trust.
• Use deterministic steps first. A reliable pattern is code for the predictable parts, a model only where language or judgment is needed, and guardrails plus a human approval step around any action that changes real systems. That is also the core of how to build an AI agent that doesn't go wrong in production.

2. The output cap: no more than 30% AI-written

A close cousin of the first rule: for any final deliverable (a report, an article, a design), AI should contribute no more than about 30%, with the rest being your own thinking and writing. It is used by some teams and creators as a guard against generic, unoriginal work and over-reliance.

It has no published origin either, and it is hard to measure: what counts as "30%" when AI suggested the outline and you wrote every sentence? A more useful version is to define what must be human (the argument, the data, the decisions, the sign-off) and disclose AI use where your organisation, client or publisher requires it.

3. The school version: keep the AI score under 30%

Among students, the 30% rule usually means "your AI-detection score must be below 30%," with the rest counted as original work. This is the version most likely to cause real harm, because:

• No detector vendor defines a 30% rule. It is an informal threshold that some instructors use and many don't. Your syllabus or your institution's academic integrity policy is the only rule that applies to you.
• Detectors make mistakes, especially at low scores. Turnitin, the most widely used, does not show a number for AI scores between 1% and 19%; it shows an asterisk instead, because results in that range are less reliable and false positives are more likely. Turnitin also says its AI score should not be the sole basis for adverse action against a student.

If you are a student: ask your instructor what is allowed, keep your drafts and notes (version history is good evidence of your own work), and cite any AI help the way your course requires. If you are a teacher: treat a detector score as a reason for a conversation, never as proof.

4. The jobs version: 30% of work hours could be automated

When news stories say "AI could automate 30% of work," they are usually citing McKinsey Global Institute's July 2023 report Generative AI and the future of work in America. It estimated that without generative AI, automation could take over activities accounting for 21.5% of hours worked in the US economy by 2030, and that with generative AI that share rises to 29.5%, rounded to "about 30%." The report presents this as a midpoint within a very wide range of scenarios.

Two things to keep straight:

• It is hours and tasks, not jobs. Automating 30% of the hours in a role usually changes the role rather than removing it.
• It is a 2023 estimate. McKinsey's later research uses different horizons and numbers, so quote the specific report and year rather than "the 30% rule."

5. The project version: 30% of generative AI projects abandoned

The fifth version comes from Gartner. On 29 July 2024 it predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, because of poor data quality, inadequate risk controls, escalating costs or unclear business value. In business settings, "the 30% rule" is sometimes shorthand for this warning: plan for a meaningful share of pilots never reaching production.

The useful lesson is in Gartner's reasons, not the number. Before starting an AI pilot, write down the business metric it should move, the data it depends on, what it will cost to run at scale, and who owns the risk. If you can't fill those in, you are likely building one of the 30%.

So which 30% rule should you follow?

None of them is binding, and you won't find a "30% rule" in any regulation or standard. If you need an operating principle for AI at work, use this instead:

1. Map the process and classify each step: deterministic, judgment, or high-stakes.
2. Automate the deterministic steps with plain code or workflow tools, and use models where language understanding adds value.
3. Keep a named human accountable for judgment and high-stakes steps, with review before anything irreversible happens. This matters even more with agentic AI, where AI agents take actions rather than just suggesting them.
4. Measure and adjust. Track error rates and time saved per step, and move the human line up or down based on evidence, not on a round number.

That gives you the benefit the 70/30 rule is reaching for (efficiency without giving up judgment) with a ratio that fits your actual work.