Agentic AI · 14 October 2026
The 30% rule in AI: where it came from and what to do with it
What is the 30% rule in AI? It is a loose planning guide, also called a heuristic, for dividing work between an AI and a person. It is not a law, a product setting, or a percentage that tells you what your own job needs.
The seven-to-three split counts 7 named preparation actions and 3 named owner decision groups in one fictional catering plan. It does not measure time, value, accuracy, cost, or a universal AI capability.
The version I hear most often says an AI can prepare roughly 70% of a job. A person keeps the 30% that needs judgment, responsibility, or a decision that reaches somebody else.
Keep the word "roughly" in that sentence. The number starts a useful conversation. Put the actual boundary in the task itself.
What is the 30% rule in AI?
a useful shorthandOne catering inquiry may need an owner after the first missing detail. Another may have clear information until the reply itself. The important line sits where a wrong decision becomes expensive.
That line belongs to the work in front of you. A percentage can remind you to keep a person involved. A percentage cannot choose the exact moment.
Should I make AI stop at 30% of every task?
No. Stop the AI where it reaches a missing fact, an outside action, or a decision you must own. A short task may need a person after one prepared line. A well-checked internal task may contain more preparation before that point.
The same number carries several different meanings
ask which jobThe phrase has two opposite workplace readings. Gravity's 70/30 work split gives AI roughly 70% of repeatable preparation. XLev's cautious 30/70 start gives AI a smaller 30% of a workflow first.
The same digits also appear in a separate 10-20-70 change model. That model divides the effort of an AI change across tools, data, and people. It does not divide one job between an AI and a person.
Three similar percentages that answer different questions
| Model | What the number is trying to discuss | What it cannot tell you |
|---|---|---|
| 70/30 work split | AI prepares routine work while a person keeps judgment and review | Whether 70/30 fits your job |
| 30/70 cautious start | Start with a smaller automated portion while a person keeps more of the workflow | Whether the first portion should be 30% for your work |
| 10-20-70 change model | Where AI-change effort sits across tools, data, and people | Who should send a reply, set a price, or approve a booking |
When somebody says "the 30% rule," ask which job they mean. Then ask what the person still owns when the AI finishes its part.
Where the work-split version came from
not a measured lawI cannot point you to one study, standard, or inventor who set this number. Current explainers describe it as an informal guideline. They also turn the split around.
One version says AI prepares most routine work while a person keeps the judgment. Another says begin by giving AI a smaller, safe portion of a workflow. A recent BDO industry report calls a 70/30 human-AI workflow principle "emerging."
That disagreement tells you what the phrase really is. It is a shorthand for sharing work. Treating the percentage as a measured promise turns a useful reminder into a bad rule.
I split one fictional catering job into preparation and decisions
a fictional preparation runI gave an AI three fictional catering enquiries and a small menu note. The source material was fictional. The instructions allowed one local preparation note and prohibited a reply, a price, an availability check, a booking, payment, and account changes.
The AI copied the dates, headcounts, food requirements, and open questions. The AI matched only the menu notes that appeared in the fictional file. The AI also made a list of what the owner still had to decide.
The owner kept availability, date holds, allergy accommodation, delivery, menu choice, price, and the final reply. Those decisions depend on information the practice file did not contain.
I counted seven preparation actions: read the inquiries, copy dates, copy headcounts, copy food requirements, copy open questions, match stated menu notes, and list missing decisions. I counted three owner decision groups: the date, the offer, and the external reply.
The date group covers availability and a date hold. The offer group covers allergy accommodation, menu choice, delivery, and price. The reply group decides whether any message leaves the business.
That gives this fictional plan a 70/30 shape. The count does not measure time, value, accuracy, or what another job should do.
7 AI preparation actions
- Read the three enquiries
- Copy the requested dates
- Copy the headcounts
- Copy food requirements
- Copy open questions
- Match stated menu notes
- List missing decisions
3 owner decision groups
- Date. Decide availability and any hold.
- Offer. Decide accommodation, menu, delivery, and price.
- Reply. Decide whether an external message leaves the business.
The check replaces the percentage
write the stop lineBefore you hand a task to an AI, write the action that stops the run. A useful stop line can say, "Prepare the internal note, then wait before a reply, quote, booking, payment, or account change."
That sentence gives a person something to inspect. A percentage alone does not.
The check also changes how you talk to a capable builder. A person who builds a real system may need time to understand your work, connect the right tools, and handle unusual cases. You still need to agree where their system stops and where you decide.
Choose one recurring pile. Pick something you already understand, such as enquiry notes, meeting follow-ups, or a weekly document review.
Put copies in one local folder. Use invented or approved practice material first. Keep private records and real accounts outside the test.
Name what the AI may prepare. Let the instruction say which files it may read and which internal note it may create.
Name what stays with you. Include every outward message, price, booking, payment, account change, and decision that needs facts outside the folder.
Compare the note with the source files. Check the date, request, missing detail, and stop line before you use the result.
The whole thing on one card
if you read nothing elseThe 30% rule in AI is a rule of thumb, not a measured law. You may hear it as AI doing roughly 70% of repeatable preparation while a person keeps judgment. You may also hear the percentages reversed.
Start
One local pileChoose work you can read beside its source files.
Prepare
Internal materialLet the AI sort, copy, and flag what it can show you.
Decide
Keep the outside actionHold replies, prices, bookings, payment, and account changes.
Use 70/30 as a prompt to divide a job. Do not use it as a measured promise.
Write the exact stop line before a run starts. Compare the result with the files.
The human-in-the-loop habit belongs in Agentic AI Explained: How AI Agents Actually Work. Begin small, then widen a boundary you can explain.
Where do you stand?
15 short situations about handing a job to an AI tool. In each one you pick what you would do next.
- ▸Whether you sized the job right
- ▸What you handed over
- ▸How you checked what came back
- ▸When you stopped
About 5 minutes. You see your score, your weakest area and the reasoning behind every answer before anybody asks you for an address.
Questions people ask next
answered in one line eachWhat is the 30% rule in AI?
The common work version is a rough split: AI prepares repeatable work and a person keeps judgment, review, and responsibility. It is a rule of thumb, also called a heuristic, not an official standard or a percentage that every job must follow.
Who created the 30% rule in AI?
There is no single study, standards body, or named inventor behind the work-split phrase that I can point you to. Current explainers describe it as an informal guideline, and they do not all use the percentage in the same direction.
Does the 30% rule mean AI can do 70% of any job?
No. A ratio cannot tell you which facts are missing, which action reaches another person, or where a mistake becomes expensive. Divide the work by the source files and the decision boundary instead.
What should a person keep when using AI at work?
Keep the decisions that need facts outside the AI's approved files, and keep outward actions with a person at the start. That includes sending, pricing, bookings, payment, and account changes until you have a checked process.