Why AI’s biggest economic shift is still ahead

Artificial intelligence (AI) hasn't yet produced the sweeping disruption many expected because most organisations are still using it to speed up workflows built for an earlier digital era, according to a new perspective paper by researchers at Microsoft Research.

The paper, titled From Augmentation to Reconstruction: Guiding the AI Disruption to the Good Place and published as an arXiv preprint, argues that the largest economic and social effects of AI will come not from better chatbots or faster task completion, but from rebuilding markets, education, news, coding and other systems around delegation, machine-to-machine interaction, continuous monitoring and auditable constraints.

AI is still mostly speeding up old workflows

The paper challenges the public debate around AI, which often focuses on whether models can pass exams, navigate websites or outperform benchmarks. The authors state that such examples are attention-grabbing but too narrow. The more important question is whether AI can help people learn new skills, make better decisions, coordinate more effectively and build systems that were not possible under human-centered workflows.

Current AI adoption remains conservative, with most organisations using the technology for summarising documents, drafting emails, generating code suggestions, answering customer questions or producing first-pass analysis. These uses can save time, but they usually operate inside existing forms, approvals, meetings, queues and decision chains. Faster drafting still waits for slow review. Faster analysis still feeds legacy decision processes. The workflow becomes quicker in parts, but the structure remains largely unchanged.

To explain this, the authors describe three stages of AI integration: augmentation, automation and reconstruction. These stages are conceptual rather than strictly chronological, and organisations may occupy more than one stage at once. But the authors argue that only the third stage produces genuine structural transformation.

1. Augmentation: AI helps humans perform existing tasks. A worker uses AI to draft an email, summarise a document, write code snippets or generate ideas. Humans still manage sequencing, judgment and accountability. This stage spreads quickly because it requires little organisational redesign. It delivers real productivity gains, but those gains are local.
2. Automation: AI takes over more routine tasks with less human attention. In education, for example, an AI system may generate lesson plans, create assessments, assist with grading and provide adaptive tutoring. Teachers and students then focus more on broader learning goals and progress. In business operations, AI may automate well-defined work segments and reduce coordination drag. The authors say this stage often feels like the first major productivity step-change. But automation still tends to run inside older architectures. The system may be doing more work automatically, but it remains organized around human-era processes. The authors describe this as a transitional stage, not the final destination. It may produce uneven productivity data because firms must first make complementary investments in data, infrastructure, workflows and organisational processes before the full benefits appear.
3. Reconstruction: According to the authors, the real disruption lies in this stage. At this point, workflows are redesigned around AI's distinctive strengths: speed, parallel processing, memory, continuous monitoring and machine-to-machine interaction. Tasks are not merely accelerated or delegated. They are re-sequenced, removed or converted into auditable constraints. Human roles shift toward higher-level judgment, goal-setting, architecture, oversight and exception handling.

The paper compares this transition to the early electrification of factories. Replacing steam engines with electric motors produced only modest gains until factories were redesigned around electricity's flexibility. The authors argue that AI may follow a similar path. The biggest gains will arrive only when organisations stop asking how AI can mimic people and start asking what can be rebuilt around AI's capabilities.

Agentic markets could change shopping, learning, news and coding

The paper uses four domains to show how reconstruction could change everyday activity: shopping, education, news and coding. In each case, the shift is not simply from humans doing tasks faster to AI doing the same tasks automatically. It is a move toward systems built around delegation, continuous monitoring and agent-to-agent coordination.

In shopping, the current digital economy still assumes that humans search, browse, compare products and navigate interfaces. AI already supports recommendations, chatbots and personalisation, but these are still layers on top of human browsing. In a more automated version, a consumer might delegate specific tasks, such as reordering groceries or finding a flight under a price limit.

The reconstructed version is more radical. Personal agents could continuously manage consumption, anticipate needs, coordinate with vendor agents, negotiate over price and terms, and help create dynamically customised goods. Markets would shift from human browsing interfaces toward machine-readable discovery and quality assurance. Firms would compete less on visibility and more on value, responsiveness and trustworthiness as expressed in terms that agents can evaluate.

In education, augmentation means students use AI to explain concepts, complete assignments or generate practice questions, while teachers use it to draft lesson plans or tests. Automation means AI systems produce lesson plans, create assessments, assist with grading and tutor students in real time. Reconstruction would reorganise learning around continuous personalised loops. AI would track mastery, generate tailored instruction, assess progress and feed results back into curriculum design. The focus would shift from static assignments and tests toward dynamic skill acquisition.

In the news domain, AI now summarises articles, aggregates headlines and highlights key points. A more automated system could deliver personalised daily briefings across sources and formats. Reconstruction would turn news consumption into an interactive, continuously updated dialogue. Personal agents would monitor sources, evaluate credibility, cross-check claims and synthesize evolving narratives according to a user's goals. The unit of consumption would move from individual articles to adaptive understanding.

Coding is one of the domains where elements of reconstruction are already beginning to appear. In augmentation, AI suggests code snippets, explains functions and helps with debugging. In automation, it generates code, tests, refactors and documents well-defined tasks. In reconstruction, software development is built around AI agents that can plan, generate and iterate on code. Humans move toward problem formulation, architectural guidance, editing and curation, while agent-to-agent coordination increasingly handles execution.

The authors argue that this kind of reconstruction changes what it means to shop, learn, consume news or code. The activity itself is redefined. That is why the most important AI effects may not show up immediately in familiar productivity measures. The paper links this delay to the productivity J-curve associated with general-purpose technologies, where early adoption requires large intangible investments before broad gains become visible.

The future depends on trust, infrastructure and incentives

The paper holds that AI reconstruction has not yet arrived at scale because the binding constraints are institutional, not only technical. Agentic systems require trust and accountability, machine-legible data, interoperable interfaces, redesigned workflows and incentives that reward system-level transformation rather than local optimisation.

The study identifies the following barriers:

• Trust: Organisations will not delegate meaningful authority to agents unless they can inspect behaviour, evaluate outcomes, assign responsibility and respond to failures. Agent systems need instrumentation, monitoring, evaluation, incident response and liability frameworks. Without these systems, firms will keep AI in advisory or assistive roles.
• Data and interfaces: Reliable agentic execution requires clean, interoperable data and explicit machine-readable policies. Many organisations have not made their internal operations legible enough for machines to act autonomously. Even fewer have built systems that can safely interoperate with external agents. Weak interfaces and brittle data pipelines raise risks and limit autonomy.
• Human-centered workflows: Most enterprise systems assume that humans read screens, fill out forms, interpret exceptions and mediate interactions. Agent-first systems require APIs, protocols and machine-readable rules that allow agents to transact, coordinate and escalate without constant human intervention. This is not a small software layer. It is a structural redesign of work.
• Incentives: Firms can gain short-term value by using AI to improve existing business models, but the largest benefits may require rebuilding markets and organisations in ways that threaten incumbent advantages. Large firms have the resources to invest in agentic infrastructure but may resist changes that disrupt their current business models. Smaller firms may have stronger incentives to transform but less capacity to build the necessary systems.

The authors warn that if organisations stop at automation, they risk locking in systems optimized around legacy workflows. This could limit wider gains and intensify worker uncertainty. In the short term, firms may favour capital deepening and partial automation over broad investment in new skills, even though the long-term opportunity may depend on redesigned roles and new forms of work.

The agentic future is not predetermined. AI may produce open, competitive systems that broaden welfare gains. It may also reinforce walled gardens, concentrate surplus and preserve friction in more automated form. Agents embedded inside proprietary ecosystems could lock users into closed markets, protect existing platforms and delay the broader benefits of agent-to-agent coordination.

The authors call for leaders to do more than anticipate where AI is headed. They argue that organisations should actively steer AI toward a better outcome by building open, auditable and transformative designs before lock-in becomes irreversible. That means treating advanced agentic coordination as an organisational capability developed over time, not as a feature to buy or a demo to admire.

Organisations should invest in machine-readable data and interfaces, redesign workflows to be AI-native, and define delegation boundaries and accountability structures for agent-driven decisions. These investments may not immediately look like productivity gains, but the paper argues they are necessary for the system-level transformation AI makes possible.