AI for Work
Posts
👩‍🎓 Harvard Study: AI Doesn't Improve All Work Equally

👩‍🎓 Harvard Study: AI Doesn't Improve All Work Equally

A Harvard experiment with 758 consultants reveals the "jagged frontier" - why AI makes you 40% better at some tasks while making you 19% more likely to fail at others. The surprising part? Working out which tasks fall where.

Adam Hobson
22 May

In partnership with

AI isn’t uniformly good or bad for knowledge work; it’s brilliant on some tasks, terrible on others, and your career depends on learning exactly where that line—Harvard calls it the “jagged technological frontier”—runs in your day-to-day work

Paper here.

What the Harvard/BCG experiment showed

Some tasks AI excels at (deemed inside the frontier), others - not so much (outside the frontier).

Their results:

Tasks inside the frontier	Tasks outside the frontier
758 BCG consultants using ChatGPT finished 12% more tasks, 25% faster, and delivered 40% higher-quality work on 18 creative, writing and ideation tasks.	On a single analytical case that required weighing messy evidence, the same consultants were 19 percentage points more likely to be wrong when they followed ChatGPT.
Biggest winners: consultants who’d scored in the bottom half on a pre-test—43% quality lift vs 17% for the top half.	They also finished faster—so they were confidently wrong. (Speed ≠ accuracy).

The takeaway: AI can level the playing field where it excels, but it creates hidden sinkholes elsewhere (at times with supreme confidence). Treating all tasks the same - as in “AI can do everything” is a trap.

The "Jagged Frontier" explained

Harvard calls this the "jagged frontier" because AI capabilities don't follow a smooth, predictable line. Instead, they create an irregular boundary where tasks of similar difficulty may fall on completely opposite sides.

The jagged reality: A task that seems perfect for AI (like data analysis) might trip it up completely, while a task that seems uniquely human (like creative writing) might see massive AI benefits. The frontier zigs and zags in ways that defy human intuition.

And this jagged line runs through your work. It demonstrates that you can't rely on assumptions about what AI can and can’t do. You have to discover where the frontier runs and divide tasks between yourself and tools you can trust.

Ok, so we need to divide the work?

Absolutely. The report describes three ways we currently divvy up tasks between ourselves and AI tools:

Style	How it works	Typical result
Autopilot (fire-and-forget)	Prompt once, copy-paste the answer.	Fastest—but prone to the biggest factual errors. Great for low-stakes drafts, risky everywhere else.
Centaur (as in half person/half machine)	Human and AI take turns. You hand clear sub-tasks (e.g. idea lists, first-pass outlines) to ChatGPT, then switch hats and apply human reasoning for analysis and final decisions.	Reduces risk on ambiguous or high-stakes work; easy to adopt.
Cyborg	Human and AI work simultaneously. You co-edit, critique and reprompt line-by-line, blending strengths in real time.	Highest ceiling on speed and creativity, but demands constant vigilance and good prompting habits.

The study found most catastrophic errors came from the Autopilot crowd. Top performers clustered in the Centaur or Cyborg camps because they managed the boundary between AI intuition and human judgement.

But I’ll be the first to tell you I use the autopilot style A LOT of the time - it’s not that it’s bad it’s that it’s not suitable every task on which I work.

So how do we find our own frontier?

The short answer is one of the key traits to future-proofing your career in the AI overhaul - human judgement.

Decide on the obvious ones
- Shallow work—drafting emails, summarising research, generating options—belongs on the AI-first list.
- High-stakes synthesis—conflicting data, strategic trade-offs, nuanced client feedback—goes on the human-first list.
Test it out
Draft one proposal paragraph yourself, let ChatGPT draft the next, then compare clarity, factual soundness and re-work time. Extremes reveal themselves quickly.
Listen to your gut - the money maker
- Keep thinking “let me double-check that”? The task sits outside the frontier.
- Feel yourself saying “that’s exactly what I needed”? It’s inside.

You know what’s right, you know what’s good. Think of AI outputs as generated by a human - open to criticism, not taken for gospel.

Practical tips for dividing the work

Task	How to test	Success signal
Batch of routine client emails	Draft half with ChatGPT, half manually.	AI emails need only tone tweaks—no fact corrections.
Data-heavy recommendation	Ask ChatGPT for counter-arguments, not the answer.	It surfaces blind-spots you missed.
Quarterly strategy memo	Let ChatGPT structure sections; you add evidence and nuance.	Memo feels clearer; re-writing time shrinks.
Customer sentiment report	Feed raw survey comments to ChatGPT for a summary; write one manually from a different data slice.	AI summary captures the same key themes you would flag—no obvious misreads.
Competitive-landscape slide	Ask ChatGPT to generate a SWOT table for two rivals; you build the same slide for two others.	AI SWOT is factually sound and requires little re-fact-checking.
Meeting agenda & minutes	Have ChatGPT draft tomorrow’s agenda and, after the meeting, the minutes; do both manually for a second meeting.	Participants say the AI version is clear and the minutes reflect actual discussion.
Budget forecast spreadsheet notes	Paste last year’s numbers, ask ChatGPT for a narrative explaining variances; write another narrative yourself.	Finance lead signs off on AI narrative with no corrections.
Product FAQ draft	Prompt ChatGPT to write first-pass FAQs for a new feature; you draft a parallel set.	Support team prefers the AI draft or merges <10 % edits.

These strategies give you direct human feedback on wheter AI helped you or not. Note the minutes saved or the feeling of “easiness” along with your confidence level after each trial. Within a fortnight your personal frontier will draw itself.

Career Upshot

Most colleagues will splash AI across every job and oscillate between brilliance and blunders. The scarce skill is discernment—knowing when to hit Generate and when to think. Master that, and you’ll deliver work that is faster and more reliable than either pure-human or AI-only rivals.

IN PARTNERSHIP WITH 1440

Receive Honest News Today

Join over 4 million Americans who start their day with 1440 – your daily digest for unbiased, fact-centric news. From politics to sports, we cover it all by analyzing over 100 sources. Our concise, 5-minute read lands in your inbox each morning at no cost. Experience news without the noise; let 1440 help you make up your own mind. Sign up now and invite your friends and family to be part of the informed.

Before you go!

I'd love to know what you thought of this week's email. I'm always trying to improve to bring you the best AI newsletter possible.