Blog · Review Essay · Last reviewed June 16, 2026

Human-Machine Reconfigurations and Situated Action

Lucy Suchman's Human-Machine Reconfigurations is a central book for anyone trying to understand why AI systems misread work, conversation, agency, and control when they treat human action as plan execution. Situated action, for this review, means that human activity is produced inside local arrangements of people, tools, bodies, records, timing, tacit norms, institutional authority, and repair. A plan, prompt, workflow, policy, or model output can be useful, but it is a resource inside that arrangement, not a complete specification of what happens.

The practical test is direct: when an AI system claims to understand a task, look for the situated labor it depends on. Who supplies context, repairs ambiguity, absorbs exceptions, authorizes action, contests error, and carries the cost when the system's formal version of the world is wrong?

The Book

Human-Machine Reconfigurations: Plans and Situated Actions was published by Cambridge University Press as the second edition and sequel to Suchman's 1987 Plans and Situated Actions: The Problem of Human-Machine Communication. Cambridge Core lists the book under the 2006 print publication year and describes it as a study of how agency is figured at the human-machine interface. Lancaster University's research directory records the 2007 publication, 326 pages, Cambridge University Press, and ISBN 9780521675888.

Suchman is not writing from outside computing. Her Lancaster profile says her dissertation became the 1987 book, that she spent two decades at Xerox PARC, and that her work critically engaged artificial intelligence, human-computer interaction, ethnographies of everyday practice, feminist science and technology studies, and the politics of technology design. That location matters. The book is not an abstract complaint that machines are bad. It is a close argument about what designers miss when they model people too thinly.

The second edition keeps the original argument about plans and situated action, then extends it through chapters on ordering devices, agency at the interface, humanlike machines, and the reconfiguration of human-machine boundaries. Cambridge's table of contents makes the arc visible: from interactive artifacts, plans, situated actions, and human-machine communication to later work on scripts, agencies, AI, robotics, and the demystification of the humanlike machine.

Situated Action, Defined

Situated action is not a claim that people act randomly or that planning is useless. It is a claim about where the intelligence of action is located. The intelligence is not contained wholly inside a plan, a rule, a mental representation, a prompt, or an interface state. It is distributed across the setting in which people notice what matters, coordinate with others, use available artifacts, repair failures, and decide what counts as an adequate next move.

That definition matters because digital systems often reverse the relation. They treat the official representation as the real activity and treat local practice as noise, deviation, or edge-case cleanup. Suchman's book asks designers and governors to inspect the opposite possibility: the local practice may be where the real competence lives, while the formal plan is only a partial public handle on that competence.

The distinction is not academic hair-splitting. A benefits rule, hospital workflow, moderation policy, classroom rubric, security playbook, or agent instruction can coordinate action only because people know when to follow it, when to question it, when to supplement it, and when the situation has outrun the representation. If that judgment is hidden, then automation can appear more capable than it is because human repair has been absorbed into the background.

Plans Are Not Action

Suchman's most durable contribution is deceptively simple: a plan is a resource for action, not the hidden engine that determines action. People do not move through the world by executing a complete internal program. They orient, repair, improvise, interpret, ask, gesture, look around, reinterpret the situation, and use plans when plans help.

That distinction was aimed partly at older AI planning assumptions, but it now lands directly on model-mediated work. Many institutions still imagine human activity as a workflow waiting to be encoded: intake, triage, risk score, recommendation, approval, denial, escalation. The workflow can be useful. The error begins when the ordering device is mistaken for the work itself.

Cambridge's listing for the chapter "Plans, Scripts, and Other Ordering Devices" is enough to name the target. The relevant ordering devices now include model cards, workflow diagrams, prompt templates, task graphs, service-level metrics, risk scores, agent tool policies, and compliance checklists. None of them acts alone. Each has to be interpreted, enacted, refused, repaired, or made accountable inside a specific setting.

This is why the book belongs beside discussions of legibility. Institutions make people legible by turning messy situations into categories, forms, and sequences. Suchman shows the reciprocal problem: the institution also makes itself blind when it forgets how much situated work is required to keep the category, form, or sequence connected to reality. The failure is not merely "bad UX." It is a breakdown in institutional perception.

The Interface as Field Site

The interface is not just a surface where a user sends commands to a machine. It is a field site where assumptions about personhood, competence, communication, control, failure, and agency become operational. A system tells the user what kind of action it recognizes. The user discovers what kind of person the system expects them to be.

That is the political bite of Suchman's work. Design is not neutral simplification. A machine can require the world to become more machine-readable before it can help. It can force workers to narrate their judgment in approved fields. It can turn exception handling into invisible labor. It can ask users to adapt to the model, then treat the adaptation as proof that the model understood the user.

This is especially important for AI agents and assistants. When a system proposes tasks, extracts intent, summarizes meetings, routes cases, writes decisions, or simulates support, it is not simply helping a preexisting human plan unfold. It is shaping what the plan becomes. The interface can narrow the situation before anyone notices that narrowing has happened.

Suchman gives a way to resist that flattening. Start from practice. Watch what people actually do, including the pauses, repairs, workarounds, conflicts, informal cues, embodied knowledge, and local accountability that never fit cleanly into the official process map. Then design around the reality that action is produced in relation, not downloaded from a plan.

Current Context

As of June 16, 2026, the book reads less like a historical intervention in human-computer interaction and more like a governance manual for AI deployment. Generative systems are used as copilots, assistants, triage aids, drafting tools, tutors, customer-service layers, and increasingly as agents that can call tools or initiate actions. The more these systems move from producing text to shaping work, the more Suchman's question matters: what practical competence has been moved into the interface, and what practical competence has merely been hidden from view?

The current standards and legal context points in the same direction. NIST's AI Risk Management Framework describes AI risk management as work across the lifecycle and organizes it around govern, map, measure, and manage functions. Its online Core emphasizes context, human-AI configurations, oversight, documentation, engagement with affected actors, and continuous risk management. That is a standards-body version of a situated-action lesson: risk cannot be understood from the model artifact alone.

NIST's 2026 AI Agent Standards Initiative and the NCCoE project on software and AI agent identity and authorization also make the shift concrete. Agentic systems raise ordinary security and governance questions in a sharper form: What identity does the agent act under? What is it authorized to do? Which tools can it use? Who can revoke access? What logs show the chain from instruction to action? Suchman helps explain why those are not only technical controls. They are the infrastructure that keeps agency socially accountable when a system acts across many settings.

In the European Union, the AI Act's high-risk provisions similarly connect transparency and oversight to use context. Article 13 requires information that helps deployers interpret outputs and use them appropriately; Article 14 requires human oversight measures proportionate to risk, autonomy, and context of use, including the ability to understand limits, watch for over-reliance, override output, and interrupt operation. This page is not legal advice, and jurisdiction matters, but the governance direction is clear: "human oversight" has to be more than a person placed beside a machine.

The AI-Age Reading

Read in 2026, Human-Machine Reconfigurations is a direct challenge to the fantasy of total delegation. AI systems are increasingly sold as planners, copilots, autonomous agents, customer-service representatives, tutors, analysts, recruiters, managers, and companions. The sales pitch often assumes that a task can be captured as objective, context, instruction, tool call, and success condition. Suchman helps explain why that is fragile.

Most consequential tasks contain hidden situated judgment. A benefits case is not just a document classification problem. A clinical note is not just a summary. A moderation queue is not just policy application. A student essay is not just an output to score. A grief conversation is not just sentiment and next-best response. The practical meaning of the action depends on who is present, what history matters, what cannot be said safely, what evidence is missing, and who will bear the cost if the system gets the situation wrong.

The book also punctures a common AI governance mistake: treating "human in the loop" as sufficient. A human reviewer placed after a machine decision may have little power if the interface has already framed the case, selected the evidence, ranked the options, timed the work, and made dissent administratively expensive. Situated action means accountability has to be designed into the whole sociotechnical arrangement, not stapled onto the last click.

For generative AI, the lesson is even sharper. A fluent model can generate a plan that looks complete because language makes completeness easy to perform. But a good-looking plan may erase the local practices that make action possible: relationships, discretion, timing, embodied skill, institutional memory, labor politics, and obligations to people who are not represented in the prompt. The map can become more convincing precisely as it becomes less accountable to the terrain.

Governance and Safety

The governance implication is not "never automate." It is that automation must be evaluated against the situated work it will reorganize. Before deploying a model or agent into a real process, an institution should map the work practice, not only the workflow diagram. That means identifying who currently handles exceptions, who notices missing context, who has authority to pause, who documents uncertainty, and who can explain a decision to an affected person.

For AI assistants and decision-support systems, the safety floor should include a clear boundary between draft and binding action; visible evidence, uncertainty, and source provenance; audit trails that preserve prompt, retrieval, tool-use, output, human review, override, and final action; and appeal paths for people affected by the system. The reviewer must have time, domain context, authority, and organizational protection to disagree. Otherwise oversight becomes theater.

For AI agents, the same point becomes a permissions problem. An agent should have a distinct identity, least-privilege access, scoped tools, revocable credentials, transaction limits, escalation thresholds, and incident review. The key question is not whether the agent "understands" the organization. The question is whether the organization can trace, constrain, contest, and recover from the agent's actions when the formal plan meets an unanticipated situation.

This is also a procurement issue. Vendors often sell general capability, but Suchman's argument pushes buyers toward local evidence: Has the system been tested in this setting? What failure modes appear when users are rushed, undertrained, multilingual, disabled, distrustful, or dealing with high-stakes exceptions? What labor is created for frontline staff? What duties are shifted to the affected person? What happens when the model output is plausible but procedurally wrong?

Where the Book Needs Friction

The book is conceptually dense. Readers looking for quick policy prescriptions, product checklists, or a simple "AI should do this" argument will have to translate. Suchman's vocabulary moves through ethnomethodology, anthropology, feminist STS, agency theory, robotics, and interface studies. That density is part of the value, but it can make the book less accessible than its practical importance warrants.

The other limit is that the AI landscape has changed. Large language models, foundation-model platforms, agent toolchains, data-center politics, benchmark culture, synthetic media, and model-mediated social life arrived after the book's core empirical world. Suchman gives the theory of situated action and human-machine configuration, but readers still need contemporary work on platform power, labor extraction, surveillance, procurement, data governance, and model evaluation.

The book also should not be turned into a veto on formalization. Formal representations can be useful, protective, and democratically important. Rules can prevent arbitrariness. Documentation can support appeal. Standard procedures can reduce discretionary abuse. Suchman's point is subtler and more useful: representations are partial, and their adequacy depends on the arrangements that keep them answerable to practice.

Still, the argument scales well. The more powerful the system, the more dangerous it is to confuse formal representation with practical competence. The more autonomous the agent, the more urgent it becomes to ask which humans, institutions, bodies, and environments are being folded into the action while the system is described as acting on its own.

What This Changes

The book's practical test is this: when a machine claims to understand a human activity, look for the missing situated labor.

In AI workplaces, that means looking for the workers correcting summaries, overriding recommendations, feeding edge cases back into the system, and carrying liability for model-shaped errors. In public services, it means looking for the people forced to translate their lives into categories the system can process. In companion and therapeutic interfaces, it means looking for the outside relationships and care institutions that are made less visible by a frictionless conversation.

Suchman's deeper warning is about reconfiguration. Human and machine are not fixed categories that meet at a clean boundary. They are enacted through practice. A dashboard can make a manager more machine-like. A chatbot can make a user more confessional. A workflow can make a worker into an exception handler for an automated system. An agent can make an institution less willing to maintain human expertise because the machine appears to have absorbed it.

That is why this book remains essential. It teaches readers to distrust the fantasy that cognition, agency, and responsibility can be cleanly moved into the machine. Human-machine systems do not replace situated action. They rearrange it. The governing question is who gains power from the rearrangement, who becomes legible, who becomes invisible, and whether the people inside the system can still contest the frame that names what they are doing.

Source Discipline

This review separates four kinds of claims. The bibliographic claims come from Cambridge Core and Lancaster University. The reading of Suchman's argument comes from the book's publisher records, chapter listings, and the review's own interpretation of the situated-action tradition; it does not reproduce the text of the book. The contemporary governance claims are grounded in NIST materials and the EU AI Act service pages cited below. The practical audit recommendations are this site's synthesis, not a claim that Suchman, NIST, or the European Commission endorses them.

The current governance sources also have limits. NIST's AI RMF is voluntary guidance, not a binding statute. The NIST agent standards work is active and standards-oriented, not a finished global rulebook. The EU AI Act applies by jurisdiction, role, system type, and implementation date. Those limits are part of the lesson: source discipline prevents a governance page from becoming another overconfident plan detached from the settings where it will be used.

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