Blog · Review Essay · Last reviewed June 25, 2026

The Handover and the Artificial Agents Already in Charge

David Runciman's The Handover: How We Gave Control of Our Lives to Corporations, States and AIs is useful because it refuses to treat artificial intelligence as humanity's first encounter with nonhuman agency. Modern people have already built long-lived artificial actors, granted them legal powers, routed everyday life through them, and then acted surprised when those actors developed interests of their own. They are called states and corporations.

Here, artificial agency does not mean consciousness, personhood, divinity, or AGI. It means an entity can act across time through law, records, assets, tools, credentials, staff, procedures, and delegated authority. That makes Runciman's book less a prophecy about machine rule than a concrete warning about handover points: the places where human judgment is converted into institutional process and then becomes hard to call back.

The useful governance test is a handover ledger: what authority was transferred, to which artificial actor, through which records and tools, with what stop controls, and with what evidence left for people affected by the action. Without that ledger, delegation becomes a story told after the fact.

The Book

The Handover was published by Profile Books in hardback on September 7, 2023, with a 2024 paperback. Profile's publisher page lists the hardback ISBN as 9781788163675, the paperback ISBN as 9781788163682, and the ebook ISBN as 9781782836216. Its central premise is blunt: AI seems new because it is made of computation, but the political problem of living with artificial agency is older than software.

Runciman is well positioned for that claim. The University of Cambridge's Department of Politics and International Studies lists him as Honorary Professor of Politics, former head of the department from October 2014 to October 2018, a former host of Talking Politics, host of Past Present Future, and a scholar of modern political thought, democracy, the state, representation, technology, and democratic politics. The book grows out of that background: it reads AI through Hobbes, corporate personality, democratic institutions, war, markets, delegation, and the long history of collective actors that outlive the individuals inside them.

The Profile preview lays out the book's structure. The introduction is called "States, Corporations, Robots"; later chapters move through superagents, groupthink, life and death, tribes, churches, empires, transformation, ownership, work, and the question of who works for whom. The Edinburgh Futures Institute described the book as a distillation of more than three centuries of thinking about artificial agency. That is the right frame. The book is not only about AI risk. It is about the political habit of creating entities that can act through us and then over us.

Current Context

As of June 25, 2026, Runciman's argument reads less like a metaphor and more like a governance map. Agentic AI is now discussed through identity, authorization, tool permissions, observability, record-keeping, human oversight, and organizational risk management. The central problem is no longer only whether a model can produce a correct answer. It is whether an institution can safely delegate action into software while preserving responsibility for what happens next.

NIST's AI Agent Standards Initiative frames agent identity, authentication, interoperability, security evaluation, and trusted interaction as standards work. NIST's National Cybersecurity Center of Excellence separately describes software and AI agent identity and authorization as a project about identifying, managing, and authorizing actions taken by software agents, including AI agents. That vocabulary matters because a handover without a named actor is not accountable delegation; it is borrowed authority.

The April 2026 joint guidance Careful Adoption of Agentic AI Services, issued by ASD's ACSC, CISA, NSA, the Canadian Centre for Cyber Security, NCSC-NZ, and NCSC-UK, treats agentic AI as an IT and cybersecurity risk surface. Its emphasis on privilege, configuration, behavior, structural complexity, accountability, monitoring, and human oversight turns Runciman's political concern into an operational one: artificial agents become dangerous when they can act broadly while the institution cannot reconstruct or constrain the action.

Regulation points in the same direction for high-risk uses. EU AI Act Article 12 requires record-keeping capabilities for high-risk systems; Article 14 requires human oversight measures; and Article 26 sets deployer obligations for high-risk systems. ISO/IEC 42001:2023 treats AI as a management-system issue for organizations using or providing AI-based products or services. OECD AI Principles, updated in 2024, continue to place human rights, democratic values, transparency, accountability, robustness, and human agency in the trustworthiness frame. None of these sources is a general law of AI agents, but together they make the handover inspectable.

The First Handover

Runciman's most productive move is to call the state and the corporation artificial agents without making the comparison cute. These entities are not metaphors in daily life. They own property, sign contracts, sue and get sued, wage war, borrow money, hire workers, collect taxes, buy software, set policy, hold records, and make decisions no single human could make alone. They have names, addresses, rights, obligations, budgets, archives, strategies, and survival instincts.

That first handover made modern life possible. States and corporations coordinate infrastructure, science, finance, health systems, trade, education, war, logistics, and public administration at a scale no household, village, or charismatic leader could sustain. But the same scale creates an old governance problem: a machine made out of humans can become hard for humans to redirect.

The handover is not a single surrender. It is a chain of small translations: judgment into policy, policy into form, form into database, database into workflow, workflow into contract, contract into vendor system, and vendor system into automated action. By the time a person objects, the decision may already be distributed across offices, code, logs, procurement terms, and institutional habits.

This is where the book belongs near Seeing Like a State, Cloud Empires, The Black Box Society, and Autonomous Technology. All four ask what happens when systems built for coordination become systems that define reality for the people inside them. Runciman adds a sharper institutional genealogy: AI does not arrive outside these older actors. It arrives through their procurement offices, cloud contracts, militaries, courts, platforms, schools, hospitals, and corporate roadmaps.

That matters because the AI future is often narrated as a direct confrontation between humans and machines. The more plausible story is triangular. Humans meet AI through institutions. Institutions meet AI through incentives. AI systems reshape the institutions that deploy them, and those institutions then reshape the people who must live under their decisions.

The Institutional Mind

The book's language of artificial agency is especially helpful for thinking about institutional cognition. A state does not think like a person. A corporation does not want like a person. But each can gather information, process signals, allocate resources, pursue goals, defend itself, forget inconvenient facts, and act in ways that no individual participant fully intended.

This is not mysticism. It is bureaucracy, law, accounting, procedure, hierarchy, contract, memory, and interface. An agency knows through forms and databases. A corporation remembers through records, dashboards, institutional routines, and balance sheets. A government sees through censuses, licenses, benefits systems, police reports, tax data, procurement files, and regulated categories. A platform sees through telemetry, logs, profiles, user graphs, ranking metrics, and moderation queues.

AI intensifies this because models can become cognitive middleware inside already artificial agents. A benefits office uses a chatbot at the front desk. A company uses a copilot to draft policy, summarize meetings, triage customers, rank leads, or monitor workers. A police department uses transcription and summarization systems that enter official memory. A hospital uses an AI scribe that turns conversation into record. In each case, model output becomes institutional perception.

The key safety question is therefore evidentiary. Does the institution know when a record is observed fact, human judgment, model summary, tool output, vendor transformation, or generated draft? If those layers collapse into one official file, the artificial agent called an institution starts citing its own transformations as reality.

The danger is not only error. It is that the institution may treat the model-mediated record as cleaner than the lived event. Once a summary, score, risk flag, generated report, or recommendation enters the file, it can become the thing later actors cite as evidence. The institutional mind has not become conscious. It has become easier to automate.

The AI Reading

Read in 2026, The Handover is a warning against two bad stories. The first says AI is an alien force that suddenly descends on human society. The second says AI is merely a tool and therefore politically neutral. Runciman's older artificial agents make both stories look thin. AI is neither outside society nor simply inside a user's hand. It is being absorbed by powerful entities that already have legal standing, budgets, coercive reach, data access, and reasons to optimize their own survival.

The key question becomes: which artificial agent is being strengthened? A model used by a public agency can either improve service or make administrative discretion harder to contest. A model used by a corporation can either help workers or turn workplace knowledge into surveillance and automation leverage. A model used by a state can either improve capacity or deepen the ability to classify, predict, and intervene. A model used by a platform can either help users navigate information or make the platform's private ordering of reality harder to see.

That question should be asked at procurement time, not only after a scandal. An AI system is not introduced into a blank social field. It enters a budget, a chain of command, a labor process, a compliance culture, a vendor relationship, and a set of people with more or less power to refuse. The same model can serve the public in one arrangement and harden institutional self-protection in another.

That question is more useful than asking whether AI will rule us in the abstract. AI rule will usually look like institutional rule with a model inside it. It will look like the benefits chatbot that becomes the front door of the state, the enterprise assistant that becomes corporate memory, the model policy that becomes compliance common sense, the answer engine that becomes public knowledge, the procurement pilot that becomes infrastructure, or the risk score that becomes the reason no one has to take responsibility.

Runciman's book also helps explain why "human in the loop" can be a weak safeguard. States and corporations are already loops made of humans. The issue is not whether a human appears somewhere in the workflow. The issue is whether any human has power to interrupt, appeal, explain, repair, refuse, or be accountable for the system's action.

Governance and Safety

The practical lesson is to govern the handover, not just the model. Before an AI agent, copilot, classifier, or summarizer enters an institution, the responsible actor should be named: the public body, employer, vendor, platform, contractor, officer, or board that is authorizing the system to act. A model cannot be accountable in the civic sense. The institution using it can be made accountable if its authority, records, approvals, procurement terms, and appeal paths are designed for contestability.

That means agent governance needs mundane controls. The system should have scoped identity, least-privilege access, tool permissions, sandbox boundaries, logging, monitoring, revocation, incident review, and a clear rule for when a person with real authority must approve, interrupt, reverse, or compensate for an action. These are not decorative compliance rituals. They are the difference between delegation and abdication.

The minimum artifact is a handover record connected to an AI system inventory, procurement file, agent identity, and observability plan. It should name the institutional owner, task boundary, delegated authority, tools, credentials, data sources, records changed, human approval gates, rollback path, appeal or complaint path, vendor obligations, incident owner, monitoring rule, and retirement criteria.

Those controls do not solve Runciman's problem by themselves. They make the right questions concrete. Who can see the agent's actions? Whose credentials did it use? Which tools could it call? Which records did it change? Which people were affected? What evidence is preserved? What rights to explanation, review, appeal, correction, or rollback exist after the fact? If a corporation or state cannot answer those questions, it has not adopted an intelligent assistant. It has created a new opaque office inside an older artificial agent.

The Recursive Trap

The recursive pattern is simple. Institutions adopt AI to see and act more efficiently. The AI system formats what the institution can see. Workers, applicants, customers, students, patients, and citizens adapt to the formatted reality. Their adapted behavior becomes new data. The institution then treats the updated data as confirmation that the system understands the world.

This is not hypothetical. Search rankings shape websites; websites then become training material for search and answer systems. Workplace dashboards shape work; the shaped work becomes evidence for future dashboard metrics. Eligibility systems shape how people describe need; the descriptions become administrative truth. Model-generated records shape later investigations; later investigations cite those records as memory. Platform algorithms shape culture; cultural producers adapt to platform incentives; the adapted culture returns as proof of demand.

The Handover is useful here because it identifies the older container for the loop. Recursive reality does not float in the cloud. It is housed in institutions with legal personality, money, infrastructure, and coercive consequences. A bad model can be fixed. A bad institutional arrangement can keep generating new models that reproduce the same priorities.

The practical AI governance question is therefore institutional, not just technical. Who owns the system? Who defines success? Who audits the records? Who can refuse the interface? Who benefits from automation? Who carries the error? Who has standing to challenge the decision? Who can slow deployment when the artificial agent called a corporation or the artificial agent called a state says speed is necessary?

Post-deployment monitoring is the test of whether the handover remains governable. The institution should watch for expanding tool permissions, silent scope changes, worker dependence, appeals that cannot reach a responsible human, vendor lock-in, rising exception queues, unreviewed secondary use of agent traces, and records that become harder to correct because a model helped write them.

Where the Book Needs Friction

Runciman's analogy is strong, but it should not become too neat. States and corporations are artificial agents, but they are not all the same kind of agent. A constitutional state, a petrostate, a platform monopoly, a nonprofit, a bank, a school district, a church, a military contractor, and an AI lab do not delegate power in the same way. Their legal forms matter. Their histories matter. Their violence, accountability, labor systems, and public obligations differ.

The phrase "we gave control" also needs pressure. Control was not handed over equally or voluntarily. Colonized people, workers, debtors, migrants, welfare recipients, prisoners, patients, and students have often been made legible to state and corporate systems before they had meaningful power to shape those systems. AI extends a very uneven history of delegation.

The book's institutional optimism also needs limits. If states and corporations are the older artificial agents that might restrain AI, they are also the actors racing to deploy it. The state wants security, efficiency, and capacity. The corporation wants scale, rent, data, market power, and defensible margins. Treating these agents as the solution without asking who controls them risks handing the future to the very entities that made the problem administratively attractive.

That does not weaken the book. It clarifies its best use. The Handover should be read as a map of the terrain, not a guarantee that existing institutions will save us. It shows why AI governance has to repair public capacity, corporate accountability, labor rights, procurement, audit power, appeal rights, and democratic oversight at the same time it evaluates models.

What This Changes

The strongest lesson is to stop asking only what AI can do and start asking which artificial agent it empowers.

When a model enters a workflow, name the actor it is serving. Is it serving the affected person, the worker, the public, the agency, the vendor, the platform, the shareholder, the security apparatus, or the institution's desire to make a messy judgment look objective? If those interests conflict, where is the conflict recorded? If the model makes the institution faster, does it also make the institution more answerable?

Then inspect the handover point. What decision, memory, explanation, or relationship is being transferred from situated human judgment into model-mediated process? Can the affected person see the transfer? Can they contest it? Can a worker override it without punishment? Can a public body explain it without hiding behind a vendor? Can the system be removed after dependency forms?

The practical rule is simple: no artificial agent should gain consequential authority faster than the institution can name it, limit it, observe it, explain it, revoke it, and repair the record after it harms someone. That rule applies to the old agents of law and capital as much as to the new agents of software.

The Handover belongs on an AI reading shelf because it changes the scale of the question. The future is not humans versus machines. It is humans living among artificial agents old and new, some made of law and bureaucracy, some made of capital and contract, some made of models and data centers. The political task is to keep those agents from using one another to make human judgment ornamental.

Source Discipline

This review keeps three claims separate. First, the book claim: Runciman argues that states and corporations are earlier artificial agents, and the publisher's preview defines the relevant agency as the ability to act in the world. Second, the interpretation claim: that analogy is most useful when applied to institutional deployment, procurement, memory, and authority rather than to speculative machine consciousness. Third, the current-governance claim: as of June 25, 2026, AI-agent governance is increasingly discussed through standards, identity, authorization, logging, human oversight, cybersecurity guidance, and organizational risk management.

Those current-governance claims are sourced to primary standards bodies, regulators, cybersecurity agencies, or official institutional pages where possible. Reviews in The Guardian and Literary Review are used only for reception and interpretive context, not as evidence for legal or technical requirements.

This review does not treat AI systems as conscious, divine, AGI, or legally autonomous. It treats them as software-mediated institutional actors whose practical significance depends on the authority, tools, records, and organizations through which they act.

Sources

Book links are paid affiliate links. As an Amazon Associate I earn from qualifying purchases.

Buy on Amazon Browse Books

Return to Blog · Return to Books