AI's Jagged Frontier: Why It Is Like Past Technologies, and Why It Is Not

AI is similar to past general-purpose technologies in the way that matters most: its value depends on complementary innovation. The technology by itself is not the system. Its payoff depends on data quality, workflow redesign, training, evaluation, integration, governance, and organizational learning.

Research on general-purpose-technology adoption from the NBER emphasizes that productivity depends not only on the technology itself but on firm-specific benefits and the co-invention costs around it. That is doubly true for AI, where the models are general but the value is intensely contextual.

A language model can draft a contract clause, but whether that draft is useful depends on the company's legal standards, jurisdiction, risk posture, negotiation context, and review process. The model is one part of a larger operating system, and the other parts have to be built.

The first real difference is what AI touches. The IMF estimates that almost 40 percent of global employment is exposed to AI, rising to about 60 percent of jobs in advanced economies, and notes that, unlike earlier automation and information technology, AI has a distinctive ability to affect high-skilled jobs.

Exposure does not mean disappearance. It means tasks within those jobs may be automated, augmented, restructured, or repriced. But the breadth, and the reach into professional work, is the point. Earlier automation hit factory, clerical, and routine work first. AI reaches writing, analysis, legal reasoning, coding, design, and management communication.

That makes AI organizationally sensitive in a way prior waves were not. It touches work tied to status, expertise, and career progression. The adoption challenge is not only technical; it is cultural and political. People will ask whether AI devalues their expertise, whether juniors will still learn, and whether gains become support or layoffs. Those questions deserve real answers.

The second difference is the interface. The internet required connection, browsers, and broadband. Enterprise software required training, configuration, and procurement. AI can be used by typing a request in natural language. That makes adoption faster and more chaotic at the same time.

Employees can use AI before the organization has policies, security controls, evaluation methods, or approved workflows. This is why shadow AI is not a marginal issue but a predictable feature of the transition. The thing that makes AI easy to adopt is the same thing that makes it hard to govern.

The low barrier also changes what training means. Traditional software training teaches users where to click. AI literacy has to teach people how to frame a task, supply context, test outputs, verify sources, protect data, recognize failure modes, and decide when not to use the tool at all. The interface is simple; competent use is not.

The third difference is the strangest and the most consequential. AI has a jagged frontier. It performs some tasks impressively and fails at nearby tasks that look almost identical. A Harvard Business School field experiment with Boston Consulting Group found that consultants using GPT-4 completed more tasks, worked faster, and produced higher-quality work on tasks inside the frontier.

On a complex managerial task deliberately chosen to sit outside the frontier, the same AI users were 19 percent less likely to produce correct solutions than consultants working without AI. The tool did not just fail to help; it actively pulled capable people toward worse answers, because it was confident in territory where it was unreliable.

That is the central management problem with AI. The risk is not that it is good or bad. The risk is that it is inconsistently capable, and that the boundary between capable and not is invisible from the output. It will summarize a standard policy beautifully and misread an ambiguous exception just as fluently.

Spell out the failure modes, because they are not exotic. AI may write fluent code that compiles and contains a security flaw. It may generate a plausible market analysis while inventing facts. It may produce a customer response that sounds empathetic and violates policy. The same tool is useful, misleading, or dangerous depending on the task.

There is also a difference in how AI relates to authority. Search engines point you to sources. Spreadsheets show their formulas if you inspect them. AI often produces synthesized answers without transparent reasoning or reliable sourcing, and it does so in a conversational, authoritative, complete-sounding voice. That is a persuasion risk: people believe it because it sounds competent.

So output has to be treated as a draft, a hypothesis, or a recommendation until it has been verified through a defined process. Not because AI is untrustworthy in general, but because its trustworthiness varies task by task in ways the confident tone hides.

The practical response to a jagged frontier is classification. Organizations need to define where AI can draft, where it can recommend, where it can decide, where it must be checked, and where it should not be used at all. Without classification, governance defaults to one of two failures.

A blanket use-AI-everywhere policy ignores risk and eventually produces an incident. A blanket do-not-use-AI policy ignores value and will be quietly bypassed, which is worse, because now the usage is both unmanaged and hidden. Neither extreme survives contact with how people actually work.

The correct approach is differentiated control: high freedom where tasks are high-volume and low-risk, tight review where outputs are regulated, financial, legal, safety-critical, or customer-facing, and clear prohibition where the risk plainly outweighs the value. That map is the foundation of everything in the playbook that follows.