AI can do roughly 75–80% of the work in most media tasks I have watched — with the share highest in production, assembly, and pattern-finding, and much lower in negotiation, relationships, and final judgment. That is the 75–80% principle, and the first thing to say about it is that it is a heuristic, not a benchmark.
I want to be careful with the number, because round figures attract false precision. The 75–80% is not a measured statistic from a controlled study. It is the shape I keep seeing after running these workflows myself and watching teams run them: AI takes the large, repetitive, generative middle of a task, and leaves a smaller, higher-stakes edge that stays human. The exact share moves by task, by team, and by how clean the context is. The shape has held for me; the decimal has not.
I wrote the framing into the definition of AI-native media operations because teams get into trouble at both extremes. Let me lay out where the line falls, and how to find it in your own work.
What the 75–80% principle says
The principle says that for a typical media task, a well-designed workflow with AI doing the first pass can cover most of the work, and a human covers the rest. It does not say AI does 80% of every task, and it does not say the human 20% is small in importance. Importance and volume are different axes.
Think of it as a division of labour rather than a score. The model is best at producing a lot of plausible material quickly: options, variants, summaries, first drafts, pattern calls. The human is best at deciding what matters, choosing among the plausible, and carrying the consequences. The percentage is just a reminder that the boundary exists — and that it sits further toward the model than most cautious teams assume, and further toward the human than most enthusiastic vendors admit.
Where AI's share is high
The tasks where AI covers the most ground share three properties: high volume, low consequence per unit, and a clear standard to check against.
- Creative and copy variants. Headlines, descriptions, hooks, social copy — an AI-native team produces dozens against a decided brief, then a human prunes.
- Research synthesis. Turning a pile of public sources into a structured summary is now mechanical, which frees the analyst to ask better questions.
- Reporting assembly. Pulling, joining, formatting, and narrating the weekly numbers is largely automatable once the data is clean.
- Budget and pacing maths. Reforecasting, checking spend against plan, and flagging anomalies are pattern-finding tasks a model does well.
- Localization and adaptation. Translating and adapting a campaign across markets, with a native reviewer at the end, is a volume task that scales well.
In these areas, 75–80% is a conservative estimate. Watch a good operator work with a good workflow and the model share looks higher still.
Where AI's share is low
The tasks where AI covers the least ground share the opposite properties: low volume, high consequence per unit, and no clean standard.
- Negotiation. Media buying, rate cards, partnerships, and commitments are relational and adversarial. A model can prepare you; it cannot sit across the table.
- Relationship and trust. Clients buy confidence. Confidence is built in conversations and delivered by a person whose name is on the work.
- Final judgment. Choosing the one plan that fits the client's real constraint, or the moment to stop a test, is a judgment no model can be accountable for.
- Brand and taste calls. Knowing that copy is on-brand rather than merely correct is learned human judgment, and it is where the last 20% earns its keep.
- Naming the question. A model will answer almost anything. Deciding which question is worth asking — and which metric matters — is still the senior operator's most valuable act.
In these areas the model might assist, but 75–80% would be a wild overstatement. Here the human share is most of the job.
A rough map of the split
| Task | AI share | What stays human |
|---|---|---|
| Creative and copy variants | High | Brief, brand judgment, final pick |
| Research synthesis | High | Framing the question, validating sources |
| Reporting assembly | High | Interpreting the numbers, the recommendation |
| Budget and pacing maths | High | Deciding whether to act on the signal |
| Localization and adaptation | High | Cultural nuance, native review |
| Media planning drafts | Medium | Choosing the constraint, owning the plan |
| Negotiation | Low | The whole conversation |
| Client relationships | Low | Trust, context, accountability |
| Final go/no-go judgment | Low | The decision and its consequences |
Read the table as a shape, not a scoreboard. The high-share rows are where you rebuild the workflow first, because that is where the throughput lives. The low-share rows are where you protect capacity, because that is where the value lives.
Why the number matters less than the shape
The most common misuse of the principle is to quote the number as if it were measured, and then plan around it.
A team hears "80%" and assumes it can cut four of five people. Another team hears the same number and assumes it must keep everything human to stay safe. Both are treating a heuristic as a fact. The number's job is to keep you honest about which 20% you must keep — and to stop you from either over-delegating the judgment or refusing to rework the assembly.
If you remember one thing, make it this: the volume lives in the machine's half, and the accountability lives in yours. The two halves are not interchangeable, which is exactly why a single percentage description is useful and a single percentage target is dangerous.
How to audit your own tasks
Do not start from the principle. Start from your actual work, and let the principle fall out of the audit.
- List the tasks, not the job titles. "Produce the weekly report" decomposes into pull, clean, join, format, narrate, and interpret. The first five are high-share; the last is low-share.
- For each task, ask three questions. How much volume? How much consequence per unit? Is there a clear standard to check against? High volume plus low consequence plus a clear standard means a high AI share.
- Mark the standard explicitly. A task with no review standard stays human until you define one. Volume without a standard is risk, not throughput.
- Protect the low-share edge. Name who owns the judgment, the relationship, and the final call. If that is unclear, the workflow is not finished.
The audit usually produces a surprise: a team spends most of its anxiety on a step that is easy to automate, and almost none on the judgment step where it should be investing. The Execution guide turns this audit into a rollout order for the rest of the lifecycle.
The failure modes at both ends
Over-delegating. The team ships AI output with no review, then blames the tool when a wrong claim or a competitor's name goes live. The model did what it was asked; the workflow removed the human at the wrong point. This is the 20% being stolen by the 80%.
Under-delegating. The team uses AI as a spell-checker and keeps every stage manual, then wonders why nothing compounds. This is the 80% being treated as if it were the 20%. Both errors come from the same mistake: not knowing which half you are in.
What to measure
Track the model share and the human share separately. Cycle time tells you whether the workflow is genuinely faster; quality and error rate tell you whether the review is doing its job. If cycle time falls but errors rise, you have moved work to the machine without moving the standard — which is the fastest way to lose a client's trust and the slowest way to win it back.
FAQ
Is 75–80% a measured figure?
No. It is a working heuristic from watching these workflows, not a benchmark from a study. Treat the shape as the useful part: high share in production and assembly, low share in judgment and relationships.
Can AI do 100% of a media task?
Not one where a person has to sign their name to the result. AI can produce the material and the first pass, and in narrow, low-consequence tasks it can run end to end. Anything with real consequence needs a human owner.
Does a high AI share mean fewer people?
It often means fewer people for the assembly and the same or more for judgment, relationships, and review. The functions concentrate rather than disappear, so the headcount answer depends on which half you were weighted toward.
How do I know I have the split right?
Two signals. The cycle time falls without a rise in errors, and the low-share edge — judgment, relationships, final call — has a named owner. If either is missing, keep tuning the boundary.
The short version
AI does roughly 75–80% of the work in most media tasks, with the share highest in production, assembly, and pattern-finding and lowest in negotiation, relationships, and final judgment. It is a heuristic for finding the 20% you must keep human, not a target for headcount. Audit your own tasks for volume, consequence, and a review standard; automate the high-share half, protect the low-share edge, and measure cycle time and quality together. The Execution guide is the full map of that work, and the for-team-leads track turns the audit into a rollout order.
I am still learning this field in public, and my own number drifts as the tools improve. But the shape has held: the volume is delegable, the accountability is not.
If your team has run this audit, I would like to hear what surprised you — it is usually not the task people expected.
Cheers, Chandler