The AI dividend: Why enterprises must embrace radical application development

We might be witnessing one of the shortest-lived competitive advantages in enterprise technology history and one of the most quietly squandered.

For a moment, access to frontier AI models appeared to be a source of differentiation, but that window is already closing. The same foundation models, coding agents and increasingly interchangeable toolchains are now available to every serious enterprise. GitHub Copilot, Claude Code, Cursor, Codex, IBM Bob™ and a rapidly expanding ecosystem of agentic systems are delivering measurable productivity gains in software engineering and adjacent domains. Benchmarks continue to improve and enterprise adoption continues to broaden.

Yet the enterprise outcomes are not materializing.

The uncomfortable reality is that capability is no longer scarce. Execution discipline is. Task-level productivity improvements are becoming easier to measure, but they are not translating into proportional business outcomes. Engineering teams are moving faster. Dashboards are green and AI investment continues to grow.

Yet chief financial officers (CFOs) are asking an increasingly difficult question. If everything is faster, why is nothing fundamentally cheaper, quicker or more differentiated at the portfolio level? Why is AI often increasing the cost base before delivering measurable value?

This reality does not suggest AI has failed. It suggests that most organizations are optimizing execution within operating models that were not designed for how AI now produces work.

In this context, an emerging structural problem is becoming visible: the AI dividend tax.

The AI dividend tax describes a structural condition in which AI-driven productivity gains are generated but they systematically fail to convert into enterprise value. It is not a technology limitation. It is a value translation problem embedded within the organization.

Value is created at the edge—within teams, integrated development environments (IDEs) and agent workflows—but it leaks as it moves through the enterprise. Developers ship faster. Business analysts generate more user stories and richer requirements. Test engineers automate validation. Delivery throughput improves locally.

Yet somewhere between those gains and business outcomes, the signal degrades. Time-to-market barely shifts. Product differentiation remains flat. Customer impact improves unevenly if at all. Enterprises are accelerating work without accelerating outcomes.

Boards approve AI investment expecting compounding returns. Instead, they often receive fragmented productivity: isolated pockets of efficiency that fail to reshape enterprise performance. Organizations believe that they are scaling AI when, in reality, they are scaling local optimization within unchanged systems.

This organizational reality is the tax.

Every major technology shift exposes the same failure mode. Capability is not the problem; translation is. Cloud did not create value simply by virtualizing infrastructure. It created value when organizations redesigned operating models to exploit elasticity rather than replicate on-premises constraints. Agile did not deliver value through ceremonies. It delivered value when organizations stopped treating software delivery as a sequencing problem and started treating it as a flow problem.

AI is following the same pattern, but at far greater speed and with levels of investment that amplify the consequences of poor organizational design. Many organizations are investing heavily in models, agents and tools while underinvesting in the organizational capabilities required to convert those technologies into enterprise outcomes.

Four capabilities determine whether organizations capture the AI dividend:

•    AI platforms, assets and tools: The enterprise AI foundation, orchestration and reusable assets that provide scalable delivery.
•    Engineering plane: The methods, governance and controls that make AI-native delivery repeatable and safe.
•    Operating model: How workflows across humans and AI agents, including commercial models and delivery structures.
•    Skills and talent: The capabilities required to design, orchestrate and continuously improve AI-native systems.

These capabilities are interdependent. When they are aligned, AI amplifies the organization. When they are not, it amplifies fragmentation.

Most organizations begin their AI journey by investing in models, coding assistants and agentic tooling. These technologies are rapidly becoming enterprise infrastructure rather than sources of competitive differentiation. The differentiator is how they are orchestrated across delivery.

This stage is where harness engineering becomes important.

AI systems do not produce enterprise outcomes in isolation. They require orchestration layers that determine how they are coordinated, governed and applied across real work.

Two layers are becoming structurally important:

•    Inner harnesses optimize execution within an individual task or development session through coding agents and specialized assistants.
•    Outer harnesses coordinate workflows across delivery, connecting AI with enterprise context, governance, lifecycle management and organizational knowledge.

Inner harnesses make individuals more productive. Outer harnesses determine whether that productivity translates into enterprise performance.

A compelling illustration comes from the 2026 Faros AI engineering report, which analyzed two years of engineering telemetry from more than 22,000 developers across 4,000 teams. While AI delivered substantial gains in individual productivity, downstream engineering outcomes deteriorated: pull request sizes increased, median review times rose, more pull requests were merged without review, bugs per developer increased and production incidents per pull request increased.

The report highlights a phenomenon Faros describes as acceleration whiplash—the point at which individual productivity outpaces the organizational systems needed to absorb it.

Most organizations are investing heavily in inner harnesses while underinvesting in the outer harnesses that translate productivity into enterprise performance. They are improving engines while leaving the rest of the vehicle unchanged.

The engineering plane provides the methods, controls and repeatability that allow AI-native delivery to scale safely. Governance can no longer sit alongside delivery as a compliance activity; it must become part of the execution system itself.

As AI adoption expands, many organizations accumulate fragmented policies across tools, teams and environments. Governance exists, but not consistently enough to become operationally meaningful. Teams either slow down to satisfy controls or bypass them entirely. Neither approach scales.

The shift now required is significant. Governance must become embedded in how work is executed rather than applied after work is completed. In AI-native delivery, governance becomes execution infrastructure rather than oversight.

Most enterprise delivery models were designed for a world where software progressed sequentially through requirements, design, development, testing and deployment. AI fundamentally changes those assumptions by enabling software to be generated, evaluated and iterated in near real time.

Organizations inserted AI across the delivery lifecycle without redesigning the flow between those stages. The result is predictable: local activities accelerate while overall delivery shows minimal improvement.

It is entirely possible to increase developer productivity without materially improving delivery timelines. This dynamic is not a paradox. It is the consequence of accelerating work inside a system whose primary constraint is no longer engineering capability, but the operating model itself.
These operating model changes extend beyond delivery workflows into commercial structures.

Most enterprise delivery economics still assume a linear relationship between effort and output. AI breaks that assumption. Context engineering transforms software delivery into a flywheel, allowing knowledge, context and automation to compound rather than reset with every delivery cycle.

When delivery systems generate output for the same effort, but commercial models remain anchored in time, utilization and staffing, much of that value is left uncaptured. Some organizations improve delivery without reducing costs. Others reduce costs without changing how they price or measure value. In both cases, the economic benefits of AI remain largely unrealized.

The underlying issue is structural. Enterprise delivery relies on full-time equivalents (FTEs), utilization and rate cards. Efficiency is not rewarded; it is often disincentivized. Service providers maximize effort because revenue depends on it, while procurement mechanisms reinforce that equilibrium.

Capturing the AI dividend requires a different commercial model. Traditional rate cards are becoming an increasingly poor unit of commercial control and risk. Future partnerships will need to shift toward forward deployed units (FDUs), small, cross-functional human and AI delivery teams organized around business outcomes rather than functional hand-offs. They will also require shared accountability and outcome-based commercial models that align incentives around enterprise value rather than hours consumed.

Technology alone will not determine which organizations capture the AI dividend. The limiting factor is no longer access to AI. It is the organizational capability to apply it effectively.

As AI takes on more of the mechanics of software production, the value of engineering shifts from writing code to designing systems that humans and agents build together.

Developers devote less time producing code and more time to defining intent, managing context, validating outputs and orchestrating specialized agents.

Business analysts increasingly curate deeper business context for AI-enabled delivery. Architects design systems that optimize collaboration between people and intelligent agents. They stop treating AI as another development tool.

This shift changes the capabilities organizations must build.

Technical depth remains essential, but it is no longer sufficient. Organizations require skills in context engineering, AI evaluation, orchestration, governance, systems thinking and product discovery. Leaders must also learn to manage hybrid teams where humans and AI agents work together as integrated delivery units.

The organizations that build these capabilities fastest will create advantages that competitors cannot purchase through access to better models.

Taken individually, each of these capabilities delivers incremental improvement. Together, they form a fundamentally different enterprise delivery system. We call this radical application development (RAD).

RAD is not an incremental improvement to software delivery. It is a structural redesign of how enterprises build software in an AI-native environment.

RAD brings together four integrated capabilities:

•    AI platforms, assets and tools that form the foundation for enterprise AI delivery.
•    An engineering plane that embeds governance, consistency and repeatability into execution.
•    AI-native operating models that enable agentic teams and new ways of organizing delivery.
•    Skills and talent that enable humans and AI agents to operate as a single high-performing delivery system.

Individually, these capabilities are optimizations. Together, they create a delivery system capable of translating AI productivity into enterprise value. Most organizations will adopt elements of this model. Few will redesign their enterprise around it.

That distinction will define the next generation of competitive advantage.

The next phase of AI transformation will not depend on access to models. That advantage has already become table stakes. The differentiator is execution.

Some organizations are beginning to generate compounding gains: shorter delivery cycles, tighter feedback loops and stronger alignment between engineering output and business outcomes. In these organizations, AI is not simply making individual contributors more productive. It is reshaping how the enterprise operates.

Others will continue to experience isolated productivity gains without corresponding improvements in time-to-market, customer outcomes or competitive differentiation. The difference is not the sophistication of their AI tools. It is whether they have redesigned the system around them.

At its core, this transformation is as much about people as technology. Roles are evolving and new capabilities are emerging. Teams are becoming smaller and more agentic, and the nature of work is shifting from execution to orchestration. The AI dividend requires more than deploying AI. It requires building entirely new organizational capabilities.

AI does not automatically create enterprise value. It creates the conditions for it. Whether that value is realized depends on the systems, operating models, commercial structures and human capabilities that surround it.

That shift is the promise of radical application development. It shifts the focus from optimizing individual tasks to redesigning end-to-end delivery systems where productivity compounds across the enterprise. It prevents productivity from remaining trapped within isolated teams.
The organizations that unlock the AI dividend will not be the ones with the best models. They will be the ones that build the best systems around them.