In 2023, researchers from Harvard Business School and Boston Consulting Group gave 758 consultants tasks resembling their everyday work. Those using GPT-4 completed 12% more tasks, worked 25% faster and produced work rated over 40% higher in quality. Yet on a task outside AI's capabilities, they were 19 percentage points less likely to arrive at the correct answer.
The researchers called this the jagged technological frontier. AI performed extraordinarily well on some tasks yet failed at others that seemed equally straightforward. Even among highly qualified professionals, outcomes varied depending on how people used the technology.
Now consider what happens when this technology enters organizations designed around making people work in largely the same way.
The economics of divergence
For decades, commercial computing has been predominantly deterministic. Programs follow rules, producing predictable outcomes. This enabled firms to standardize processes, reduce variability and deliver reliable services at scale.
Generative AI introduces a parallel, probabilistic paradigm. It interprets ambiguity, generates alternatives and explores possibilities beyond explicitly programmed rules.
The economics are compelling. Generating ten approaches to a problem costs relatively little. Most may prove useless, but one could deliver a breakthrough. Where the downside is contained, the upside can be ten or a hundred times greater.
Crucially, the returns depend on how individuals use AI. Two employees with comparable qualifications and identical models may achieve radically different outcomes. One finishes an assignment faster. Another questions the assignment, reframes the problem and discovers an entirely new opportunity.
This presents a contradiction. AI rewards productive divergence, while the modern corporation was built to suppress it.
The inversion
The industrial firm created value through standardization. Repeatable processes made operations scalable, employees more interchangeable and customer experiences predictable. This logic eventually shaped knowledge work, from consulting methodologies to software development.
And for good reason. Customers expect consistency. A company cannot reinvent its operations every morning.
Consider Airbnb. Every property, host and location is different. That variety is part of its appeal. Yet the business relies on dependable listings, bookings, payments and support. Diversity in the experience is made possible by consistency in the system.
The AI-enabled firm faces the same challenge, but with a different operating logic.
Maximize variance in exploration. Minimize variance in promises.
Rather than standardizing every process, the inverse firm standardizes constraints, quality thresholds and commitments while allowing employees greater freedom to discover better solutions.
Management shifts from prescribing how work gets done to ensuring that experimentation produces dependable outcomes.
The emerging inverse firm
Early examples of this approach are appearing in industries once defined by highly structured professional workflows.
In law, Harvey and Legora are building AI-native platforms that allow lawyers to approach research, contract analysis and drafting through a combination of human judgment and AI-powered exploration. Harvey enables firms to build specialized agents grounded in their own knowledge and workflows, while Legora emphasizes collaborative work between lawyers and AI. Neither removes the need for legal accuracy or professional accountability. Instead, they expand how work can be performed while retaining human oversight.
Beyond law, Glean is developing enterprise agents that operate across organizational knowledge and workflows, combining flexibility in execution with shared context, permissions and governance. These companies offer an early glimpse of how professional work can become less dependent on prescribed sequences of tasks and more dependent on outcomes, judgment and evaluation.
Mercor offers another example, this time in recruitment. The company recruits specialists across professional domains to train and evaluate AI models. But the process also improves Mercor's own ability to identify talent. As experts complete assignments for AI labs, Mercor receives performance feedback that helps it refine candidate assessment and matching. Its founders describe receiving this information within days rather than waiting months for conventional employer feedback. Recruitment, evaluation and improvement become parts of the same continuous system. The operation learns from the work it performs, improving its ability to perform the next assignment.
The firm that learns fastest
The implications extend well beyond productivity. Innovation, traditionally organized around the edges of an established business, can become a continuous property of its operating system. Firms that capture the results of experimentation, evaluate outcomes and retain successful approaches can build sovereign intelligence: proprietary knowledge, memory, workflows and feedback loops that compound independently of whichever foundation models they use. This makes compute, inference efficiency and the economics of continuous improvement increasingly important competitive resources. Startups can build around this architecture from the beginning, while large incumbents can exploit their capital, distribution and proprietary data; mid-sized companies may be particularly exposed if they carry the inertia of incumbency without comparable resources. Yet size alone guarantees nothing. The defining advantage of the inverse firm is increasingly its learning rate: its ability to explore divergent possibilities internally, deliver dependable promises externally and durably improve what it is capable of doing next.