The 70-20-10 model is a rule of thumb for how people learn at work: roughly 70 parts in a hundred from doing the job, 20 from other people through feedback and coaching, and 10 from formal courses. It is not a measured law, but a reminder that a course alone rarely makes someone competent.
Why does it matter for teams working with agents?
Because the skills that matter most with agents are judgement skills, and judgement is learned mostly on real cases. A reviewer can learn in a course what held, refused and withheld mean, and what provenance is. They learn to spot a held payment backed only by an emailed invoice by deciding real held payments, getting some wrong, and being corrected. A training plan that is all course and no practice produces people who pass a test and freeze at the queue.
The model also warns against the opposite mistake: throwing people at a live queue with no structure and calling it learning by doing. Experience teaches only when someone checks it and feeds back.
How does it apply in practice?
For a reviewer learning to decide an agent's held actions, the three parts might look like this.
| Part | What it is | Example |
|---|---|---|
| 70: experience | Real decisions on real work | Deciding held actions during the agent's shadow period, under supervision |
| 20: other people | Feedback, coaching, calibration | Weekly calibration sessions; a team lead reviewing decisions |
| 10: formal learning | Courses, lessons, tests | Lessons on how a governed agent decides, depth and the record |
The order matters. A short formal grounding comes first, so the experience makes sense. Then supervised practice and feedback carry most of the load. The broader question of whether training must come before deployment is in do employees need training before a company deploys AI agents.
An example
Take a hypothetical insurer training claims handlers to review an agent's settlement drafts. The formal part is a few hours of lessons on depth, the four verdicts and provenance. The social part is a weekly calibration and a senior handler checking every decision for the first fortnight. The experience part is the handlers deciding real held claims throughout. By the end, each handler has decided a few hundred claims with feedback, which is where their competence actually comes from.
How is it different from the Kirkpatrick model?
The 70-20-10 model is about how people learn. The Kirkpatrick model is about how to evaluate training, in four levels: reaction, learning, behaviour and results. One shapes the design of a programme; the other shapes how you judge whether it worked. They fit together: if most learning happens on the job, then most of the evidence that training worked will also be found on the job, in the record of real decisions.
A playbook is different again: it is a written guide to how a process is run, which supports the experiential part by giving people a reference while they work.
What are its limits?
The numbers are illustrative, not research findings, and the split varies by role. A builder designing agents may need more formal learning than a reviewer. In regulated work, the formal part may also carry obligations, such as AI literacy under the EU AI Act (Article 4), that a business cannot meet by experience alone. The guide to training teams for AI agents covers how to balance them.
Where the Academy fits
Pryme Intelligence Academy's Operator track mirrors the model. The formal part is about 14 hours of lessons; the experiential part is a two-week practicum of real review decisions; the social part comes from the Pryme assessor who signs it off and from the calibration a team runs around it.
Questions people ask
Does the 70-20-10 model mean courses do not matter?
No. The formal part is small but it comes first and gives people the vocabulary to learn from experience. Without it, practice teaches bad habits as easily as good ones.
Can the experience part happen before go-live?
Yes. A shadow period or draft-only mode gives reviewers real cases with full supervision before the agent acts on its own.