AI and the Changing Strategy Landscape

Recent years have been a whirlwind for artificial intelligence. Researchers have watched frontier models grow more capable and efficient while compute costs plummet and adoption skyrockets. In 2024, several new benchmark suites—MMMU, GPQA and SWE-bench—were introduced to push the limits of advanced models; within a year, scores on these benchmarks jumped by 18.8–67.3 percentage points. Meanwhile, evidence from Stanford’s AI Index and industry reports shows that inference costs for a GPT-3.5–level system fell from $20 per million tokens in late 2022 to $0.07 by late 2024, a 280-fold reduction. Model efficiency improved as well: Microsoft’s tiny Phi-3-mini matched the MMLU score of a 540-billion-parameter model with just 3.8 billion parameters.

This acceleration is driving AI from research labs into everyday life. By 2024, 78 % of organisations reported using AI, up from 55 % the year before. Autonomous vehicles are providing hundreds of thousands of rides each week, and hundreds of AI-enabled medical devices are now FDA-approved (hai.stanford.edu). AI’s trajectory, however, is not just about technological progress; it has strategic, ethical and organisational implications. To lead in this era, executives must understand AI’s performance benchmarks, appreciate why generative AI is considered a general-purpose technology (GPT), address the ethical challenges head-on, and adopt the three-phase 3Rs framework—replace, reimagine, recombine—for redesigning processes, products and business models.

1. Understanding AI’s Performance Benchmarks

1.1 Breakthroughs in capability

The past two years have witnessed remarkable advances in model performance:

• New benchmarks, rapid gains. MMMU (multimodal reasoning), GPQA (graduate-level question answering) and SWE-bench (software engineering) were created to stress-test AI systems. Within a year, AI scores on these demanding tests increased by 18.8 percentage points on MMMU, 48.9 points on GPQA and 67.3 points on SWE-bench. These gains indicate that architectures and training methods continue to unlock substantial headroom for improvement.

• Converging frontier. The AI Index reports that the performance gap between the top-ranked model and the tenth-ranked model narrowed from 11.9 % to 5.4 % in one year. The top two models are now separated by just 0.7 %, suggesting a crowded frontier where state-of-the-art performance is no longer the sole province of a single lab.

• Human vs. machine. AI agents now outperform human developers in certain constrained programming tasks. In experiments, generative AI decreased the time to complete a professional writing task by 40 % and raised output quality. Software developers using GitHub Copilot implemented an HTTP server 127 % faster (uwaterloo.ca). Yet when given ample time (e.g., 32 hours), humans still outperform AI by a two-to-one margin, underscoring that deep reasoning remains a human strength.

1.2 Scaling, efficiency and democratization

Two inter-related trends underlie these performance leaps—compute scaling and efficiency gains:

• Super-exponential compute growth. Analyses of frontier models reveal that training compute has grown by roughly 4–5× per year since 2010. For language models specifically, compute grew as fast as 9× per year between mid-2017 and mid-2024 before slowing to ~5× per year as models caught up with the frontier.

• Rise of small models. The success of small language models (SLMs) such as Phi-3-mini signals that designers can achieve high performance with far fewer parameters. On the MMLU benchmark, reaching a 60 % score required a 540-billion-parameter model in 2022, but by 2024 a 3.8-billion-parameter model accomplished the same. This 142-fold reduction lowers the cost and energy footprint of AI, enabling deployment on edge devices.

• Inference cost collapse. The cost of querying AI models has plummeted. For GPT-3.5-level performance, inference costs dropped from $20 to just $0.07 per million tokens in 18 months —a >280-fold reduction. Stanford’s AI Index notes that depending on the task, inference prices have decreased by 9–900 × per year. These declines, combined with 30 % annual hardware cost reductions and 40 % yearly improvements in energy efficiency, are democratizing access to AI.

2. Why AI Is Becoming a General-Purpose Technology

Generative AI is increasingly recognized as a general-purpose technology—akin to electricity or the Internet—because it exhibits three defining characteristics: pervasiveness, continuous improvement, and innovation spawning. A June 2025 OECD analysis argues that generative AI qualifies as a GPT: its applications extend beyond software to sectors like life sciences and finance, it continues to improve rapidly, and it catalyses new innovations. The report cautions that productivity gains may take time to materialise but that long-term benefits depend on widespread diffusion and innovation spillovers

The implications are profound. With compute doubling every few months and models getting smarter and cheaper, AI is penetrating every industry—from self-driving fleets and AI-assisted healthcare  to AI-powered legal and financial services. High performance plus low marginal cost is the hallmark of a GPT: once developed, the technology is infinitely reproducible, spawns complementary innovations, and serves as a platform for new industries.

3. Ethical Issues: Building Trustworthy AI

Rapid diffusion of AI amplifies ethical challenges. Institutions and developers must address fairness, reliability, privacy, transparency, accountability and inclusiveness. Microsoft’s responsible-AI commitment frames these as core values, and an EDUCAUSE working group translates them into actionable principles for educational settings:

• Beneficence and Justice. AI should promote the well-being of all stakeholders and ensure fair treatment across groups

• Respect for Autonomy. Individuals must have the right to make informed decisions about AI interactions

• Transparency and Explainability. Systems should provide clear explanations of how decisions are made, enabling scrutiny and appeal.

• Accountability and Responsibility. Institutions and developers are accountable for the impacts of their AI systems

• Privacy and Data Protection. Personal data must be safeguarded against unauthorized access

• Nondiscrimination and Fairness. Bias mitigation is essential to prevent discriminatory outcomes

• Assessment of Risks and Benefits. Organisations should weigh benefits against risks and conduct ongoing evaluations

These principles are interdependent: transparency enables accountability, and fairness is intertwined with justice. Implementing them requires cultural change, cross-disciplinary collaboration, and regulatory oversight. As AI becomes a GPT, ethical frameworks must scale accordingly—governments are already stepping up with new regulations

4. The 3Rs Framework: Replace, Reimagine and Recombine

To navigate AI’s strategic implications, economist Joël Blit proposes the 3R framework, which describes how general-purpose technologies diffuse through the economy. The three phases—Replace, Reimagine and Recombine—offer a roadmap for organisations and policymakers.

4.1 Replace: Doing Existing Things Better

The Replace phase involves substituting old technologies or manual work with AI to boost efficiency. Blit notes that AI will displace older technologies and people doing tasks manually to increase efficiency. The process is conceptually straightforward:

1. Break down workflows into tasks

2. Identify tasks AI does well, particularly prediction and content generation

3. Calculate the return on investment (ROI) for automating each task

4. Prioritise replacement based on ROI

Blit emphasises that replacement can mean automation or augmentation: AI may empower workers rather than fully displace them. Experiments show that AI assistance improves productivity and quality, especially for less-skilled individuals. Executives should provide guardrails and encourage experimentation, because bottom-up adoption often precedes top-down mandates uwaterloo.ca. In practice, the Replace phase is already underway as AI handles customer service, marketing, coding, and legal research.

4.2 Reimagine: Creating New Business Models

The Reimagine phase is more challenging. It requires envisioning entirely new processes, structures and business models that leverage AI’s capabilities rather than simply enhancing existing workflows. Lessons from the internet age illustrate the power of reimagining: Amazon didn’t just put its existing bookstore online; it built a digital platform offering lower costs, broader selection and new services like recommendations and reviews. By contrast, incumbents that merely digitised existing processes stayed in the Replace phase and fell behind.

Blit proposes five lenses to spark reimagination:

1. Identify constraints imposed by previous technologies and ask how AI removes them. When electric motors freed factories from a single power shaft, the assembly line emerged; AI may free factories from human-centric layouts.

2. Focus on end goals, not incremental improvements. IBM Credit reimagined its credit-approval process from scratch instead of marginally accelerating each step, delivering a much better customer experience.

3. Leverage digital scalability. Digitised services can scale to millions at near-zero marginal cost. Companies like Netflix and Airbnb digitised their entire business model, enabling massive economies of scale.

4. Capture tacit knowledge. AI can encode expert knowledge, making services like education and healthcare more scalable and personalised. Reimagining these sectors around AI could transform access and quality.

5. Exploit newly cheap or feasible tasks. AI reduces the cost of analysis, content creation and recommendation to near zero; reimagined hiring, grocery and entertainment experiences will make abundant use of these capabilities.

Organisations that embrace these lenses will generate new products, services and business models. The goal of the Reimagine phase is not incremental efficiency but structural reinvention.

4.3 Recombine: Fusing Technologies for New GPTs

The final phase, Recombine, lies further on the horizon. It involves combining AI with other technologies to create entirely new systems. In Blit’s words, Recombine “harnesses the synergies between AI and other current and emerging technologies to radically expand what is possible”. Examples include cobots that integrate flexible robotics, AI and 6G connectivity to perform physical tasks with human-level dexterity and intelligence. Recombine is where AI merges with biotech, quantum computing, or new materials to spawn the next generation of GPTs.

At this stage, companies must look beyond AI alone and build ecosystems that integrate multiple technologies. The true winners will be those that orchestrate platforms and alliances across sectors.

5. Strategic Implications for Leaders

5.1 Act Now—The Cost of Waiting Is Rising

Waiting for AI to mature is no longer prudent. In the Replace phase, generative AI already delivers tangible productivity gains. Early adopters not only improve efficiency but also build organisational learning and cultural readiness for more radical changes. Leaders should start by inventorying tasks, identifying AI-compatible functions and piloting solutions with clear ROI.

5.2 Develop Ethical and Governance Frameworks

Adoption without ethics breeds risk. Adhering to principles of fairness, transparency, privacy and accountability must be baked into AI projects from the outset. Governance frameworks should include:

• Bias audits for training data and models.

• Explainability assessments to ensure decisions can be understood and contested

• Privacy-by-design techniques to protect sensitive data.

• Clear accountability structures assigning responsibility for AI decisions

5.3 Build Capacity for Reimagination

Reimagining requires creativity and risk-taking. Executives should cultivate cross-functional teams that pair domain expertise with AI literacy. Encourage moonshot thinking: ask how AI could eliminate bottlenecks rather than marginally improve them. Create safe spaces for experimentation, invest in up-skilling and foster partnerships with startups and research labs. Reimagining is where the largest value lies; organisations that merely replace will cede the future to more daring innovators.

5.4 Prepare for Recombine

Although the Recombine phase is nascent, leaders should scan adjacent technologies and build flexible architectures. For example, manufacturing firms might explore how AI can pair with next-generation robotics and connectivity to build autonomous factories. Healthcare providers could integrate AI with biotech and wearables to deliver personalised, proactive care. Investing in modular infrastructures and cross-industry collaborations will position firms to ride the wave of the next GPT.

Conclusion

AI’s rapid advancements in performance and efficiency, coupled with its emerging status as a general-purpose technology, are reshaping the strategic landscape. Benchmarks show steep improvements and narrowing performance gaps, while compute scaling and cost reductions are making AI ubiquitous. Yet with great power come profound ethical responsibilities—fairness, transparency, privacy and accountability must guide every deployment.

The 3Rs framework—Replace, Reimagine and Recombine—offers leaders a practical lens to navigate this transformation. Start with replacement to harvest efficiency, but quickly move toward reimagining entire processes and business models. Keep an eye on future recombinations of AI with other technologies; they will spawn new industries and redefine value chains.

Ultimately, those who embrace AI’s capabilities while upholding ethical principles and fostering creativity will not only ride the coming wave but shape its direction, becoming true thought leaders in the age of general-purpose intelligence.