All the Unwritten Processes

A good day for the robots at the regional Rust Belt airport: 25% of the route with no human intervention. The remaining 75% — breakdowns, recalibrations, hardware reconfigurations, high-intensity peripheral monitoring — appears in no job description, no performance indicator, no sales pitch. Workers call these moments saturated with invisible labour downtime. At the recycling facility in the same ethnographic study, a bad day for the AI sorting robots means a 50% grasping success rate; but that figure does not capture the parallel work of the sorters who compensate, recalibrate, restructure the flows so the system stays functional. Fox et al. (2023) observed both sites over several months and gave it a name: patchwork. They define it as "the human labour that fits in the space between what AI claims to do and what it actually accomplishes."[1]

This paper is about industrial robots. But the concept of patchwork does not stop at conveyor belts. It names a structural mechanism that runs through any organisation that has deployed an algorithmic system: the labour of monitoring, compensation and continuous adaptation that makes the system functional, without which it would collapse, and which systematically disappears from the accounts organisations give of their own functioning. In the agentic era, this mechanism acquires an extra dimension: before the patchwork of use can even begin, organisations must make a massive upstream investment in articulation, documentation and formalisation — to render their workflows legible to agents that can only operate on what they have been told. That prior debt is no less invisible than the patchwork. It is simply even more rarely named.

This work has a precise name that deployment roadmaps avoid using: process flattening. Not the tidying up of what is already written down, but the transcription of what never needed to be — the implicit procedures, the contextual heuristics, the tacit decisions that make good outcomes reproducible. An AI agent assumes this work was done before it arrived. That assumption is almost always contradicted by the facts. And it is rarely the agent that pays the difference.

The structural floor — what agents cannot absorb

Michael Polanyi put it in 1966 in a sentence that has not aged: "We can know more than we can tell."[7] (The Tacit Dimension, p. 4) That formula contains, in germ, the central problem of agentic deployment. An AI agent can only operate on what has been transmitted to it as digital artefacts — documentation, procedures, structured data, decision histories. Yet a professional organisation knows considerably more than it has ever written down. The practical question is: where does what can be told end, and where does what cannot begin?

Harry Collins (2010) offered the most rigorous taxonomy of that dividing line.[6] He distinguishes three forms of tacit knowledge: relational knowledge (not yet articulated, but articulable in principle — it can be made into a protocol if one takes the trouble); somatic knowledge (embodied, transmissible through practice and imitation); and collective knowledge (irreducibly social, acquired through socialisation in a community of practice, untransmissible to an entity that is not a member of that community). It is this third type that sets the structural floor: no effort of codification, however meticulous, can make an organisation fully legible to an agent, because the most valuable part of its knowledge does not exist as artefacts. It exists in shared practices, implicit judgments, the collective micro-decisions that make a consulting firm this consulting firm and not another, that make a radiology department know what it knows.

Böhm and Durst (2026) formalised this limit in the framework they call GRAI (Generative, Receptive Artificial Intelligence).[3] Decomposing Nonaka and Takeuchi's classic model of organisational knowledge creation into eight fields of human-machine interaction, they show that GenAI can now actively participate in seven of them — combining, externalising, internalising, storing, transmitting partially or fully formalised knowledge. One field alone structurally resists them: socialisation, understood as tacit-to-tacit transmission through co-presence and co-practice. When an agent interacts with a human to understand a complex situation, it operates, in their terms, on an "indirect reconstruction of reality from the artefacts of the digital domain."[3] It does not receive the knowledge — it receives its digital trace, which is always partial.

The empirical proof of this floor is in radiology. Lebovitz, Lifshitz-Assaf and Levina (2022) observed three departments of a large American hospital over ten months, all equipped with the same diagnostic AI tool, with comparable attitudes toward the technology.[4] In one department alone — thoracic imaging for lung cancer — radiologists regularly integrated AI results into their final diagnoses, achieving what the authors call engaged augmentation. In the other two, radiologists either ignored the AI outputs or accepted them without critical examination. The difference lay in a set of tacit interrogation practices the first department's radiologists had developed — knowledge about when and how to probe a divergent AI result, about which dimensions of the image to scrutinise, about how to reconcile two incommensurable viewpoints. These practices appeared in no written protocol. They could not: "AI interrogation practices — practices enacted by human experts to relate their own knowledge claims to those of the AI"[4] — are precisely what socialisation transmits and documentation cannot capture.

The floor is not transitory. It does not disappear with better SOPs or better models. It is constitutive of the nature of professional expertise in domains where contextual judgment is irreducible.

The patchwork — the work roadmaps never write down

This structural floor has an immediate consequence: every agentic deployment generates, from its launch, a gap between what the agent can do and what there is to do. Fox et al. (2023) name the labour that fills that gap patchwork.[1] They break it into three forms: manual compensation (doing by hand what the agent failed to do), peripheral monitoring (maintaining an omniscient attention over a system supposed to run autonomously), and material reconfiguration (continually adapting the physical or informational environment so the agent can accomplish what it is supposed to). The Reciclo patchworkers — almost exclusively Spanish-speaking women, precarious workers without solid union rights — embody the typical profile of who absorbs this cost. For Fox, this is not a marginal finding: it is the central one.

What Fox observes in waste labour, Kellogg, Valentine and Christin (2020) document in the organisational literature under the name algorithmic articulation work.[2] Their systematic review of several hundred studies of algorithms at work produces a constant result: "most computational tools are not 'turnkey' or 'off-the-shelf' technologies, despite the dominant discourse — they require considerable work to be developed, refined, implemented, maintained and modified over time."[2] Articulation work is defined in all its precision through their recourse to Star and Strauss: it is "not the work of designing a system or producing a product, but the surrounding work that makes it possible — planning and coordinating who will do what, when, where and how, as well as managing unassumed responsibilities and unfinished work, so that projects do not collapse."[2]

This work has always existed. Its novelty in agentic contexts comes down to two factors. First, the unpredictability of failures: "previous technologies tended to fail in relatively predictable ways, but machine-learning algorithms often fail in ways that are difficult, if not impossible, to anticipate."[2] Fox's robots break down; LLM agents hallucinate, over-generalise, produce plausible-but-incorrect outputs in cases their deployment documentation had not anticipated. In both situations, articulation work cannot be planned in advance — it must be improvised continuously by those at the interface. Second, the scale of the demand: the McKinsey Global Institute, cited by Kellogg et al. (p. 779), estimated that in the United States alone, the demand for "algorithmic translators" — a convenient term for articulation work renamed a skill — would reach two to four million jobs by 2026.[2] That figure describes a market demand. It does not describe a wage.

What lets these two bodies of research — patchwork and articulation work — describe the same reality from two angles is a shared feature: structural invisibility. Patchwork is not acknowledged in robot vendors' sales pitches, any more than articulation work appears in agent launch presentations. Fox et al. observe that machine "downtime" — the moment when robots are stopped — is organisationally designated as a loss of productivity, when in reality it is "saturated with multiple forms of high-intensity human labour, performed in the spatial interstices when the machines break down."[1] This designation is not an oversight. It is functional: naming patchwork as work would require valuing it, paying for it, writing it into deployment budgets. The roadmap that accounted for this cost would be less attractive. So the roadmap does not account for it.

The loop that closes — the debt worsens by itself

Agentic deployment without prior investment in legibility does not only produce surface patchwork. It erodes the very structures that would allow legibility to be produced afterwards. Baygi and Huysman (2025) describe this mechanism as the rewiring of the organisational social fabric.[5] They define that fabric as "the way the paths of an organisation's members continually intertwine over the course of their daily work" — who crosses whom, when, for what, with what exchanges of expertise, trust, collegiality. An ungoverned GenAI rewires these intersections: it "does not merely improve employee productivity — it reroutes the paths along which resources circulate within organisations."[5]

Left ungoverned, this rewiring produces three disruptive archetypes: the polymath (the AI-augmented generalist who bypasses experts in domains outside their real competence, reducing the exchanges that sharpen specialists' expertise); the oracle (the disseminator of plausible but unverified information that circulates through the organisation without anyone taking charge of verification); and the siren (the employee isolated in their personalised AI assistant, cut off from the collateral interactions that transmit collective tacit knowledge).[5] These three archetypes converge on the same result: an organisation that accelerates its consumption of collective tacit knowledge without rebuilding the social fabric through which that knowledge is normally transmitted and renewed.

The "book people" of Fahrenheit 451 are, in fiction, the symmetrical answer to the central problem of this essay. In a world where books are burned, Ray Bradbury imagines individuals who memorise the works to save them from the fire — not because the texts could not be recopied or printed elsewhere, but because knowledge must be carried, embodied, alive in a human memory to remain available. Polanyi's formula — we can know more than we can tell — is the theoretical version of what Bradbury formalises in parable: some knowledge cannot be archived. It is carried. What Bradbury's book people preserve cannot be entrusted to an external medium without loss: it is the living, social transmission between bearers that gives it value. AI agents, which can only operate on digital artefacts, cannot be book people.

The link with Collins is direct. Collective tacit knowledge — the floor agents cannot absorb — does not live in individual heads. It lives in the shared practices that organisational intersections keep alive. To reduce those intersections is to progressively impoverish the very material agents would need to learn to read. The vicious loop runs thus: agentic deployment without legibility → rewired social fabric → eroding collective tacit expertise → legibility even harder to produce → need for even more intense patchwork → even heavier costs for peripheral workers. Each turn of the loop makes the next one harder, without that escalation being visible in short-term performance metrics.

Lebovitz et al. (2022) provide the most concise formula for what this loop produces in practice: "what looks like augmentation on paper is much closer to automation."[4] In the two radiology departments without tacit interrogation practices, radiologists did not verify the AI outputs — they accepted or rejected them without real engagement. This non-integration was invisible to adoption metrics (the radiologists were indeed "using" the tool) while being functionally equivalent to automation without human control. Organisations that deploy agents without investing in the conditions of real cognitive engagement produce exactly that: a theatre of augmentation whose decoupling from operational reality only becomes visible at moments of failure — too late, too expensively.

Böhm and Durst (2026) note themselves — with a candour one does not always expect in pro-adoption papers — that "the field is developing at a high pace, leaving little time to study the effects, particularly from the perspective of knowledge-intensive activities with the human in the loop."[3] That observation about pace is the admission of an impasse: agentic deployment occurs within timeframes that structurally forbid the accumulation of use competence and the production of the legibility needed for its real effectiveness.

The expense the roadmaps do not write down

Organisations that deploy agents without accounting for the legibility debt are not making a correctable estimation error. They are making a structural choice about who will absorb that debt. The same redistribution hits junior workers. Alavi, Leidner and Mousavi (2024) note that GenAI deployment tends to short-circuit junior professionals — traditionally tasked with gathering information and drafting syntheses, these workers are now bypassed when their supervisors address the AI directly: "this shift denies junior staff the chance to expand their knowledge and skills."[8] Sambasivan et al. (2021) document the same phenomenon under the name data cascades in the context of training-data preparation: "everyone wants to do the model work, not the data work"[9] — 53 AI practitioners in high-stakes environments reporting that preparation work is structurally invisible, under-paid, concentrated on peripheral workers. The convergence of Fox (waste labour), Kellogg et al. (gig and knowledge work) and Sambasivan (AI data preparation) on this same distributional result owes nothing to chance. It is not an accidental feature of some deployments — it is a structural feature of the mode of value attribution in organisations that integrate algorithmic systems.

The recoding of this debt as "skills" is the final move in the process. When the OECD (2024) observes that AI formalisation increases the demand for workers able to specify, document and audit AI behaviours, the formulation is technically correct: demand does rise. What it does not say — what the vocabulary of skills cannot say — is that this demand is a demand on the workers least positioned to negotiate it. Nor does it say that the productivity gains agents make possible flow to the orchestrators, while the legibility costs fall on the articulators.

To this day there is no study that has directly measured the legibility cost — in person-hours, distribution by hierarchical level, financial cost — of rendering a professional workflow legible for agents. The closest is Sambasivan et al. (2021) on training-data preparation in high-stakes contexts. This measurement gap is not a research lacuna waiting to be filled. It is an indication that the question is not being asked by those with the resources to fund it. Like the patchwork itself, the debt is easier to ignore than to measure.

The central limit of this essay must be named: the cumulative mechanism — structurally insufficient legibility and an eroding social fabric simultaneously — is the synthesising argument of this essay. Each component is documented separately (the Collins / Böhm-Durst floor; the Fox / Kellogg patchwork; the Baygi rewiring); the articulation of the two dynamics into a loop of mutual degradation is an analytical inference, not a direct measurement.

What agents assume before they begin is that someone has already done the work of rendering them legible in the flow they are about to traverse. That assumption is rarely written into roadmaps as a prerequisite — and more rarely still as a cost. It is simply there, in the foundations, invisible and structurally assured of remaining visible only to the people whose job it is to satisfy it.

The implication is not to give up codification. It is to treat it for what it is: skilled, real, unequally distributed work that deserves to be named, compensated and bounded. A competent agentic deployment explicitly maps the boundary between what can be codified and what cannot before promising any delegation; makes visible the articulation cost that field workers absorb; and treats the Polanyi floor not as a technical obstacle to solve but as a structural constraint to respect in the very architecture of the delegated workflow. Organisations that ignore this cost do not eliminate it — they displace it downstream, where it will manifest as failures no one will be able to explain.