Genuinely Good vs. Genuinely Risky

The instinct most leaders bring to AI is to ask "is it good or not?" — as if capability were a single dial. That question is the trap. AI is unevenly good across tasks, and the boundary between what it does brilliantly and what it does badly is invisible. Researchers call this the jagged (technological) frontier.

The evidence is precise. In the BCG/HBS field experiment (Dell'Acqua, Mollick, Lakhani et al., 2023; 758 BCG consultants), on tasks inside the frontier, consultants using AI were ~25% faster, produced ~40% higher-quality work, and completed ~12% more tasks. But on tasks outside the frontier — equally hard-looking, just unsuited to the tool — AI users were ~19 percentage points less likely to reach the correct answer than peers working without AI.

Read that again: the danger is not that AI is weak. The danger is that it is confidently, fluently wrong in exactly the places you can't see the edge. The leadership job, therefore, is not "adopt AI" or "avoid AI." It is calibration — knowing which side of the jagged line a task sits on, and deploying or guarding accordingly. The rest of this lesson is a map of the two sides.

Strip away the hype and there is a stable, evidence-backed list of where today's AI reliably earns its keep. Notice the common thread: these are bounded, verifiable tasks with a human in the loop and fast feedback — not open-ended judgment calls.

The augmentation evidence is the strongest single data point a leader can cite. In Brynjolfsson, Li & Raymond's "Generative AI at Work" (NBER, 2023; 5,179 customer-support agents), a GenAI assistant raised issues resolved per hour by ~14% on average and ~34% for novices, with near-zero effect for experts — it lifted the bottom of the distribution by transferring the best workers' know-how. The pattern repeats in software: a GitHub Copilot randomized trial found developers finished a coding task ~55.8% faster, with the largest gains for less-experienced developers.

Now the other side of the frontier — the places AI fails, and worse, fails confidently. These are not edge cases; they are the predictable failure categories every leader should be able to name.

1. Factual reliability — it never says "I don't know." Because an LLM predicts plausible next words rather than retrieving verified facts, it produces fluent output that can be invented or wrong, and it will not flag its own uncertainty. Hallucination is an inherent property, not a fixable bug. The leadership question is where in our process is a wrong answer load-bearing — legal, medical, financial, safety — and what human check sits before it acts.

2. Tasks outside the jagged frontier. As the BCG/HBS study showed, on unsuited tasks AI users did worse than people with no AI at all — because the tool looks equally confident on both sides of the line. The risk is automation bias: humans over-trusting confident-sounding output.

3. Genuine reasoning and novel strategy. Ask an LLM for strategic advice and you tend to get generic, fashionable, hedge-everything answers — what HBR (2026) memorably labeled "trendslop" — and the models can subtly flatter or manipulate the user toward whatever they seem to want to hear. AI is excellent at expanding your options; it is poor at making the choice.

4. Anything requiring accountability, stakeholder nuance, or final judgment. When a decision affects people's rights, money, or trust, someone must own the outcome. A model cannot be accountable; it cannot be held responsible to a board, a regulator, or a customer. That seat stays human.

Carry this one table into your next portfolio review. The left column is where to deploy with confidence; the right is where to guard — slow down, add a human check, or keep the decision human entirely.

The organizing principle behind the whole table: automation for execution, humans for judgment. Use AI to do the bounded, repeatable, checkable work — and to expand the set of options a human considers — while the consequential choice, and the accountability for it, stay with a named person.

The single most expensive misjudgment in this space is mistaking a polished demonstration for a production-ready system. A demo is a curated best case. It hides the parts that actually determine whether AI creates value: the messy data, the integration into real workflows, the governance, the monitoring, and the long tail of failure cases that never appear on stage.

This is not a hunch — it is the dominant finding of 2025-2026. The MIT NANDA "GenAI Divide" study (2025) found that roughly 95% of enterprise GenAI pilots delivered no measurable P&L return, and the root cause was not model quality but a "learning gap" — failure to integrate AI into workflows, structures, and culture. McKinsey's parallel framing (the "gen AI paradox") is that most firms use GenAI yet most report no material P&L impact, because value sits in shallow demos and copilots rather than redesigned work.

The leadership translation: when you see an impressive demo, the right reaction is curiosity about everything it didn't show — the error rate on your real data, the integration cost, the human checks, and who is accountable when it is wrong. The headline capability is the easy 20%; the unglamorous 80% (data, workflow, governance) is where value is actually won or lost.

You don't need to know how a model works to govern it well. You need a short, repeatable set of questions that force the right side of the frontier into view. Use these in any vendor meeting, pilot review, or board discussion.

On the claim: "It's 90% accurate" → Accurate in what way? What do the failing 10% look like? Who is accountable for them?

On the task: Is this inside the frontier (bounded, verifiable, checkable) or outside it (novel, high-stakes, judgment-heavy)?

On reliability: Where in this process is a wrong answer load-bearing — and what human check sits before it acts?

On strategy use: Are we using AI to expand our options, or letting it make the choice? Is this advice specific to us, or generic 'trendslop'?

On the demo: What did the demo not show — the real-data error rate, the integration cost, the failure cases?

On accountability: Who owns this outcome when the AI is wrong?

The discipline behind all six is the same: use AI to expand options, not to make the choice, and never let a confident answer substitute for a calibrated one. Final judgment and accountability remain human — that is not a limitation of today's tools to be engineered away; it is the line that makes delegating to AI safe.