The 30% rule in AI is an informal guideline for splitting work between machines and people: let AI handle roughly 70 percent of a task, the repetitive, rules-based execution, while humans keep the remaining 30 percent, the parts that need judgment, creativity, ethical oversight, and accountability. It is not a law, a regulation, or a research finding. No one formally coined it, no standards body enforces it, and different sources even state it in opposite directions. What the rule really encodes is a management instinct that has held up well: automation works best when a person deliberately keeps ownership of the decisions that would be expensive to get wrong. This guide covers what the rule says in its common forms, where the number probably came from, what the research actually supports, and how to translate it into something genuinely useful when you put AI agents to work.

TL;DR: The 30% rule says AI does about 70 percent of a task and humans keep the last 30 percent: judgment, quality control, and edge cases. It is a folk heuristic, not a standard; you will also find it stated in reverse, and as an AI-content threshold in classrooms. The evidence says the real boundary between AI work and human work is task-shaped, not percentage-shaped. The practical upgrade: stop counting percentages and place human checkpoints on the actions that are irreversible, external, or high-stakes.

Diagram of the 30% rule: AI handles repetitive execution while humans keep judgment, quality, and accountability
The split the 30% rule points at: execution to the machine, judgment to the human

The three versions of the 30% rule you will actually find

Search for the rule and you will meet three different claims wearing the same name. That is the first thing worth knowing: the phrase is a container people pour their own ratio into. The three circulating versions:

Version What it says Where you see it
Humans keep 30%AI executes about 70% of a task; people retain 30% for judgment, quality control, ethics, and edge casesThe most common workplace form, used by consultants and AI training programs
Automate only 30%Start by automating roughly 30% of a workflow; humans keep 70% until the AI proves itselfCautious-adoption guides, often aimed at regulated or risk-averse teams
Education thresholdKeep AI-generated material under about 30% of a submitted piece of workClassrooms and student guidance, usually tied to AI-detection anxiety

The first version dominates business writing about the rule. The second is the same instinct running in the opposite direction: a starting ration for automation rather than a floor for human involvement. The third lives in schools, where the number functions as a social norm about how much AI help still counts as your own work; no detection vendor or university body publishes an official 30 percent threshold, and actual academic-integrity policies vary by institution.

Notice what all three share. Each one is really a claim about ownership: how much of the outcome a person must still stand behind. That is the durable idea underneath the unstable number.

Where the 30% rule comes from (nobody owns it)

No researcher, company, or regulator coined the 30% rule. It has no paper, no author, and no first citation anyone can point to. It emerged the way most management folklore does: consultants and practitioners needed a memorable answer to "how much should we automate?", a round number filled the vacuum, and repetition did the rest.

The number itself probably stuck for two reasons. First, it echoes one of the most quoted statistics in the automation debate: McKinsey's 2023 estimate that activities absorbing up to 30 percent of hours worked in the US economy could be automated by 2030, a share generative AI pushed up from 21.5 percent to 29.5 percent in their modeling. That "30 percent of work" figure describes what could be automated across an economy, not how any single task should be divided, but headlines compressed it and the compressed version fused with the oversight heuristic. Second, 30 percent simply feels right to managers: enough human involvement to stay accountable, not so much that the automation stops paying for itself.

It is worth being blunt here because most articles about the rule are not: if you see a page presenting the 30% rule as an established principle with a known origin, you are reading marketing, not history. The honest description is that it is a useful piece of folklore.

What the research actually supports

The rule is folklore, but the instinct behind it has real evidence, and the evidence sharpens it in a way the percentage never could.

The best field data comes from a 2023 Harvard Business School and Boston Consulting Group experiment with 758 consultants, the study that introduced the phrase "jagged technological frontier." On tasks within AI's capability, consultants using GPT-4 completed 12.2 percent more tasks, finished them 25.1 percent faster, and produced results rated more than 40 percent higher in quality than the control group. On a task deliberately chosen to sit outside AI's capability, the same tool made people worse: AI users were 19 percentage points less likely to reach the correct answer, because they trusted fluent output on a problem the model could not actually solve.

Read that carefully and the 30% rule's real lesson falls out. The boundary between what AI should do and what humans should do is not a percentage of the work; it is a jagged line through the task list. Some tasks belong almost entirely to the machine. Others, superficially similar, belong almost entirely to the person, and handing them over quietly degrades quality. A fixed 70/30 split gets both halves wrong: it under-delegates the work AI does brilliantly and over-delegates the work it fails at. What actually protects you is knowing which side of the frontier each task sits on, and keeping a human on the far side. Our breakdown of real AI agent examples shows the same pattern task by task, and agentic AI explained without jargon covers why agent systems make the frontier question sharper rather than softer.

One more evidence point deserves a place in this conversation: regulation. The EU AI Act requires human oversight for high-risk AI systems, and sector rules in finance and healthcare impose review duties of their own. None of them specify a ratio. The law cares about the same thing the folklore gestures at, accountable humans at consequential moments, and it also declines to put a number on it.

How to apply the 30% rule to AI agents (checkpoints, not percentages)

The rule was born in a world of AI assistants: tools that draft while you supervise every output. Autonomous agents break the arithmetic, because an agent runs multi-step work end to end. It might perform 95 percent of the keystrokes in a workflow you barely watch, and that can still be a healthier setup than a 70/30 assistant arrangement, provided the human holds the right 5 percent. The difference between assistants and agents is the subject of our autonomous vs assistive AI guide; here is what the 30% instinct becomes once work is delegated rather than co-piloted.

Let the agent own end to end:

Keep a human checkpoint on:

The mechanics of wiring these checkpoints in are covered in our guides to adding a human approval step to an agent and human-in-the-loop patterns. The short version: approvals belong at the moments listed above, not sprinkled evenly through the workflow. An approval on every step recreates the job you were trying to delegate; an approval on the two steps that matter keeps the accountability without the drag. As an agent proves itself, checkpoints can loosen deliberately, which is a trust decision, and our piece on agent trust models covers how teams stage it.

In practice this is how modern agent platforms are built. On Gravity, you describe the outcome you want in plain words, an expert-built agent runs the execution, and anything consequential comes back to you as an approval rather than a surprise. The free tier covers one agent, and paid plans start at 20 dollars per month with 20 dollars of usage included, so the sensible pattern is to start with one low-stakes agent, watch where you actually intervene for a few weeks, and let that observed boundary, not a borrowed percentage, set your oversight level.

When the 30% rule misleads

Used as a conversation starter, the rule is harmless and often helpful. Used as a target, it fails in four specific ways.

The unit is wrong. Percentages of "the work" are unmeasurable in practice. Nobody can say whether the human did 30 percent of a report. Checkpoints on named actions are countable, auditable, and enforceable; ratios are vibes.

Stakes beat ratios. A newsletter draft and a wire transfer might each be "10 percent of the workflow." One deserves zero human review and the other deserves two approvals. Reversibility and blast radius should set oversight, never share of effort. Our guide on agent safety and guardrails works through this stakes-first logic.

Fixed ratios rot. The right split this quarter is the wrong one next quarter, in both directions. Agents earn autonomy on tasks they handle cleanly, and new failure modes appear as scope grows. Treat the boundary as something you re-inspect on a schedule, not something you set once.

Oversight theater. The subtle failure: a team "keeps 30 percent" by having a human rubber-stamp everything the AI produces. Approval fatigue sets in within weeks, the human stops reading what they approve, and you get the accountability of full automation with the cost of full review. Fewer, sharper checkpoints that a person genuinely engages with protect you more than broad shallow ones.

The test worth keeping from all of this fits in one sentence: for every task you hand to AI, can you name the moment a human would catch it if it went wrong? If yes, your split is fine, whatever the percentage. If no, no ratio will save you.

Frequently asked questions

What is the 30% rule in AI?

The 30% rule is an informal guideline for dividing work between AI and people: let AI handle roughly 70% of a task, the repetitive, rules-based execution, while humans keep the remaining 30%, the judgment calls, quality control, ethics, and edge cases. It is a rule of thumb from business practice, not a law, regulation, or research finding, and different sources state it in slightly different ways.

Who created the 30% rule for AI?

Nobody in particular. No researcher, company, or regulator formally coined the 30% rule. It spread through consultants, managers, and AI training materials because it gives teams a simple starting ratio for human oversight. The number likely stuck partly because it echoes a famous McKinsey estimate that activities taking up to 30 percent of US work hours could be automated by 2030.

Is the 30% rule an official standard or law?

No. There is no regulation, industry standard, or peer-reviewed study that requires or validates a 70/30 split between AI and human work. Regulations like the EU AI Act require human oversight for certain high-risk systems, but none of them specify a percentage. Treat the 30% rule as a conversation starter about where humans must stay in the loop, not as a compliance target.

Is it the 30% rule or the 70% rule? Which side is the AI?

Both versions circulate, which tells you how informal the rule is. The most common form has AI executing about 70% of the work with humans keeping 30% for judgment and review. Some sources flip it for cautious adopters: automate only about 30% of a workflow at first and let people keep 70% until the AI earns trust. The useful idea in both is the same: decide deliberately which parts stay human.

Does the 30% rule apply to autonomous AI agents?

The spirit applies, the arithmetic does not. With autonomous agents you do not ration a percentage of the task; you place human checkpoints at the moments that carry risk: approving anything irreversible, anything customer-facing, and anything that moves money. An agent might do 95% of the keystrokes while the human keeps 100% of the accountability at two or three approval points.

What is the 30% rule for AI in schools?

In education the phrase is used differently: keep AI-generated material under roughly 30% of a submission, so the work remains substantially the student's own. No detection vendor or university body publishes an official 30% threshold; it is a norm students and teachers repeat, and individual institutions set their own policies. Always follow your school's actual rules rather than the folk number.

Sources