TL;DR

In Co-Intelligence (2024), Wharton professor Ethan Mollick offers four rules for working with AI: invite it to the table, stay the human in the loop, treat it like a person (and tell it which person), and assume today's model is the worst you will use. His field experiment with 758 Boston Consulting Group consultants found that GPT-4 made people faster and better on tasks inside AI's "jagged frontier", and worse on a task just outside it. For marketing teams, the lesson is to map that frontier task by task rather than adopt or reject AI wholesale.

Who Ethan Mollick is, and what he argues

Ethan Mollick is a professor at the Wharton School and the author of the One Useful Thing newsletter. His book Co-Intelligence: Living and Working with AI was published by Portfolio in April 2024. Its central idea is that large language models are best treated as a co-worker, collaborator or coach rather than as either a magic oracle or a glorified autocomplete.

The book is built around four rules. As summarised in reviews and in Mollick's own interviews (for example on WHYY in July 2025), they are:

  1. Always invite AI to the table. Try it on almost everything you do, within legal and ethical limits, so you learn first-hand where it helps and where it does not.
  2. Be the human in the loop. Models produce plausible text, including convincing but wrong facts and citations, so human judgement and checking stay in the process.
  3. Treat AI like a person, but tell it what kind of person it is. Giving the model a specific role and context tends to produce more relevant output, while remembering it has no feelings or sense of self.
  4. Assume this is the worst AI you will ever use. Capabilities keep moving, so conclusions about what AI "can't do" need regular review.

The jagged frontier: what the BCG experiment actually found

The book's most quoted concept comes from a field experiment Mollick co-authored with Dell'Acqua, McFowland and colleagues. It first appeared as Harvard Business School Working Paper 24-013 in September 2023, and a peer-reviewed version was published in Organization Science in March 2026.

The design was a preregistered experiment with 758 BCG consultants, about 7% of the firm's individual-contributor consulting workforce. Participants were randomly assigned to work without AI, with GPT-4, or with GPT-4 plus a short prompt-engineering overview.

  • Inside the frontier (a set of 18 realistic consulting tasks), consultants using AI completed 12.2% more tasks, worked 25.1% faster and, according to the working paper, produced more than 40% higher-quality results than the control group.
  • The levelling effect: consultants who scored below average on a baseline task improved by 43%; those above average improved by 17%.
  • Outside the frontier (a single complex business problem chosen to sit outside AI's capabilities), AI users were 19 percentage points less likely to reach the correct answer than those without AI.
Jagged frontier

the uneven, invisible boundary between tasks where AI performs well and tasks of seemingly similar difficulty where it fails.

The authors also observed two successful working styles. "Centaurs" divided work between themselves and the AI, delegating some subtasks and keeping others. "Cyborgs", in the paper's words, were "completely integrating their task flow with the AI", switching back and forth at the subtask level.

Why this matters for marketing teams

Marketing work sits on both sides of the frontier at once. Drafting variations of ad copy, brainstorming campaign angles, summarising customer interviews or restructuring a landing page are plausible candidates for the inside of the frontier, though each team should test that rather than assume it. Interpreting messy performance data, judging what a specific customer segment will believe, or deciding whether a claim is legally safe are tasks where a confident but wrong answer is costly, and where the study suggests people may over-trust the model.

The levelling effect matters too. If AI lifts weaker performers most on frontier-friendly tasks, it can raise the floor of a small team quickly. But it also means junior staff may produce polished work without having built the judgement to spot when the AI has wandered outside its competence. In our work with European SMEs we see this most often in multilingual campaigns, where fluent output can hide a factual or cultural error.

A playbook: applying the four rules to a marketing team

  1. Run a two-week "invite AI to the table" sprint. Every team member tries AI on their real tasks and logs each attempt: task, prompt, result, time saved or lost, and whether they had to correct facts.
  2. Draw your own frontier map. Sort the logged tasks into three groups: AI does well with light review; AI helps but needs expert checking; AI should not lead. Expect the map to differ by language, channel and product line.
  3. Assign a named human in the loop for each output type. Someone owns factual accuracy, brand voice and legal claims for every asset that ships. "The AI wrote it" is never an acceptable answer to a complaint.
  4. Write role prompts, not one-line requests. Brief the model as you would a freelancer: who it is (for example, a B2B copy editor for Dutch logistics buyers), the audience, the goal, constraints and examples of good work.
  5. Choose centaur or cyborg deliberately. For high-stakes work (pricing pages, regulated claims), use a centaur split with clear hand-offs. For exploratory work (ideation, first drafts), a cyborg back-and-forth is often faster.
  6. Review the map every quarter. Under rule four, a task that failed last quarter may now work; a task that worked may need new safeguards. Keep the log and revisit it.

Limits and common mistakes

The experiment is strong evidence, but it is one study, in one firm, using GPT-4 as it existed in 2023, on tasks designed by the researchers. The frontier has moved since then, which is exactly Mollick's point: you cannot rely on someone else's map. The quality ratings also measure output in a controlled setting, not commercial outcomes such as conversion or revenue.

The most common mistakes we see are treating the frontier as fixed, generalising from one good or bad experience, and confusing fluency with accuracy. The outside-the-frontier result is a warning that experienced professionals can be led astray precisely because the answer looks competent. Anthropomorphising the model can also go too far: rule three is a prompting technique, not a claim that the system understands your customers.

"Always invite AI to the table" — Ethan Mollick, Co-Intelligence

Key takeaways
  • Mollick's four rules: invite AI in, stay the human in the loop, give it a role, and assume it will keep improving.
  • In the BCG experiment, AI users inside the frontier did 12.2% more tasks, 25.1% faster, at markedly higher quality.
  • On a task outside the frontier, AI users were 19 percentage points less likely to be right.
  • Marketing teams should map their own frontier task by task and assign clear human ownership of every output.

FAQ

What is the "jagged frontier" in simple terms?
It is the idea that AI is very good at some tasks and poor at others that look similarly hard, and that the boundary between them is not obvious in advance. You only find it by testing on your own work.
Was the BCG study peer-reviewed?
Yes. It circulated as Harvard Business School Working Paper 24-013 from September 2023 and was published in Organization Science in March 2026. The headline figures for tasks completed, speed and the outside-frontier result appear in both versions.
Should junior marketers use AI more or less than seniors?
The study suggests weaker performers gain the most on suitable tasks, so juniors can benefit significantly. Pair that with clear review by someone experienced, because spotting when AI is outside its competence requires judgement juniors may not yet have.

Sources

Map your team's AI frontier

If you want help working out which of your marketing tasks AI should lead, assist or stay away from, JRK can run that mapping with your team.

Book a free growth session →

Want this working for your brand?

Book a free 30-minute growth session. We'll look at your funnel and tell you what we'd do first — honestly.

Book a growth session →