Beginners can't see AI mistakes. Experts don't believe they'll make them.
The Jagged Technological Frontier: AI's capability boundary isn't a wall — it's a row of jagged teeth. One step forward, you're safe. One step sideways, you fall.
On this page
Your company is training employees on how to write prompts. But nobody's training them on when not to open ChatGPT.
That's the missing half of most AI strategies.
758 consultants, one uncomfortable conclusion
A 2025 paper in Organization Science offers some of the most rigorous experimental data on AI's impact on knowledge workers to date.
A research team from Harvard Business School, MIT, and several other top institutions partnered with BCG to design 18 tasks simulating real consulting work. 758 BCG consultants were randomly assigned to three groups: no AI, GPT-4, and GPT-4 with prompt engineering training.
Results split in two directions.
Inside-the-frontier tasks (within AI's capability: writing reports, analyzing data, generating marketing copy):
"subjects using AI outperformed those not using AI, completing 12.2% more tasks and completing them 25.1% more quickly on average while also delivering solutions of significantly improved quality." — Dell'Acqua et al., Organization Science, 2025
12.2% more tasks completed. 25.1% faster. Higher quality.
Outside-the-frontier tasks (a complex management decision requiring integration of ambiguous information):
"subjects using AI were 19% less likely to produce correct solutions compared with those without AI, pointing to potential limitations of AI supporting knowledge workers." — Dell'Acqua et al., Organization Science, 2025
AI users were 19% less likely to get it right.
AI made them worse. And they didn't know it.
This introduces the paper's core concept: the Jagged Technological Frontier.
"a jagged technology frontier... where AI assistance improves performance for some tasks but worsens it for others, even within the same knowledge workflow and with a seemingly similar level of difficulty."
The "frontier" is AI's current capability boundary. "Jagged" means it's not a smooth curve — it's uneven. Two tasks that feel similar in difficulty to a human might sit on opposite sides of the line.
AI's capability boundary isn't a wall. It's a row of jagged teeth. One step forward and you're still safe. One step sideways and you fall.
Most companies have two AI gears: full throttle or full stop. That's not a strategy. That's a stance.
Wrong unit of analysis
"Our department is going all-in on AI" is too coarse a decision.
The right unit of analysis for AI strategy is the task, not the department. Saying "marketing uses AI" carries about as much information as "marketing uses electricity."
One consultant, on one project, might handle ten tasks: six inside the frontier, four outside. Going all-in or all-out will get you burned on some of them.
Some examples:
✅ Typically inside the frontier (AI adds value reliably):
- Writing a product description (with reference material, output is verifiable)
- Summarizing a 50-page report into 3 pages (extraction, cross-referenceable)
- Drafting a competitive analysis (structured task, human can verify)
- Classifying and tagging support tickets (clear rules, results are checkable)
- Drafting an email (clear goal and context)
⚠️ Typically outside the frontier (AI will hurt you):
- Deciding if a candidate fits the company culture (requires reading people, highly contextual)
- Prioritizing product features for the quarter (integrates market intuition, team capacity, strategy)
- Negotiating a key client contract (requires understanding relationship dynamics and history)
- Spotting a team member who's quietly checked out (requires sustained observation and nuance)
- Deciding whether to enter a new market (integrates large volumes of ambiguous, unstructured signals)
Three questions to run on any task:
- Can I verify the output? If there's a ground truth or cross-reference → probably inside the frontier
- Does it require integrating ambiguous context to make a judgment? Yes → might be outside
- What's the cost of being wrong? High stakes, low tolerance for error → proceed with caution
This isn't a one-time classification. It's a task map your organization needs to maintain continuously.
Half-trained skills
One detail from the experiment worth noting: the third group — consultants who received prompt engineering training — performed about the same as the untrained AI group on outside-the-frontier tasks.
Training didn't help them recognize when not to use AI.
Prompt engineering teaches you how to write effective instructions for AI. It's a real and useful skill — but it solves "how to use AI," not "when to put it down."
Prompting is an operational skill. Knowing when to close the tab is a cognitive skill. You can learn the first one in a day. The second requires domain expertise.
What actually needs to be cultivated: the ability to evaluate your own AI usage before you start. Ask yourself:
- Can I verify what AI gives me? If not, do I have enough domain knowledge to sense when something's off?
- Am I using AI to accelerate my judgment, or to replace it? The first is augmentation. The second is outsourcing.
- If AI's output is wrong, will I catch it? If the answer is no, you're in a high-risk scenario.
An example: a marketing manager uses AI to write a competitive analysis. The output looks complete, organized, well-formatted. But a key data point about a competitor is fabricated — an AI hallucination. A manager with domain expertise would feel something was off while reading it, check, and catch the error. Someone without that background would hand it in, and the team would make decisions on bad data.
Developing this skill looks nothing like prompt training. It's closer to critical thinking.
The most dangerous thing isn't being wrong. It's not knowing you're wrong.
On outside-the-frontier tasks, AI's answers typically "look reasonable." That's the real risk.
AI's most dangerous output isn't a wrong answer. It's a wrong answer that looks right.
The consultants who got it wrong weren't guessing randomly. They received a convincing wrong answer and accepted it. The output was fluent, internally consistent, well-structured — it looked like the right answer in every way.
At the organizational level, this risk gets amplified.
False consensus
Three people on a team simultaneously use AI to analyze the same problem and arrive at highly similar conclusions — because their prompts are similar, and the training data is shared. They compare notes, see agreement, and become more confident.
False consensus is more expensive than a wrong answer. An error you check. A consensus you act on.
This is especially dangerous in strategic discussions. A few people use AI to research a new market, all get back "there's opportunity," the decision-makers sign off. But the training data lags on recent market shifts — and the lagging parts happen to be exactly what matters.
Two failure modes
Beginners can't see AI mistakes. Experts don't believe they'll make them.
New people lack domain knowledge to sense when AI output is drifting.
Experienced people face a different risk: the more familiar they are with AI, the more confident they become in their ability to judge the frontier. But AI capability is constantly evolving, and new blind spots are forming. The capability map they built two years ago has large sections that no longer apply.
The frontier moves
The boundary drawn in the GPT-4 era has shifted significantly since then.
Two years ago, "drafting a persuasive legal opinion" was an outside-the-frontier task. Today, with o1 and Claude 3.7 in the picture, law firms are using AI to generate first drafts that attorneys review. That task has moved inside the frontier.
Similar migrations are happening in code review, medical imaging, financial modeling — tasks that AI could only handle at a "reference" level two years ago now approach "trustworthy" in some contexts.
But as the frontier expands outward, new outside-the-frontier zones form too. As AI grows stronger, people assign it more complex work — some of which isn't yet at a "hand it off safely" level, even if it looks like it is.
Every major model upgrade moves the teeth. The frontier you mapped last month is already wrong.
AI risk management isn't a one-time assessment. It's ongoing work that needs to update every time a significant model ships. With each major release, it's worth asking:
- Have any tasks I previously flagged outside the frontier moved inside?
- Are there tasks I'm currently running through AI that have quietly drifted outside?
Closing: the right framing for an AI teammate
Lucius is an AI community operations tool we built. Initially we asked: how much can Lucius do? The Jagged Frontier framework made us ask a different question: what should Lucius not do?
Community management has plenty of inside-the-frontier tasks: writing welcome messages, answering FAQs, aggregating engagement data, drafting content. Lucius handles those reliably.
Outside-the-frontier tasks also exist: sensing whether friction between two members might escalate, deciding whether to remove someone for borderline behavior, maintaining relationships with key community voices, judging whether community health needs strategic adjustment. Lucius can surface information on these — but shouldn't make autonomous calls.
A tool goes where you point it. A good teammate taps you on the shoulder when you're pointing in the wrong direction.
The best AI teammate isn't the one that can do the most. It's the one that knows where its limits are.
Open your team's task list from last week. Tag each item: inside the frontier or outside. When you're done, you'll know what your AI strategy actually needs to change.
Reference
Dell'Acqua, F., McFowland III, E., Mollick, E., Lifshitz-Assaf, H., Lakhani, K., Kambhampati, S., Nowak, M., & Aral, S. (2025). Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality. Organization Science. https://doi.org/10.1287/orsc.2025.21838