Authors and currents: what the thesis rests on

This chapter is not a literature review. A literature review orders a field to show that one knows it; what follows orders a field to show what an argument hangs from. There are eight currents, and from each one two things matter: its core idea —reduced to what it contributes, without the deference of the academic summary— and, above all, the exact place it occupies in the structure of the thesis. Whoever is looking for the bibliographic record with its verification status will find it in the Bibliography; what is offered here is the scaffolding.

Before taking them one by one, it matters to see the shape of the whole, because it is that shape that gives the argument its force. The eight currents come from disciplines that do not keep company —postwar industrial sociology, the anthropology of science, information engineering, the economics of innovation, the social psychology of organizations, among others— and were written over the course of seven decades, with no dialogue among them. And yet they converge. That traditions so distant should reach the same point —that the performance of a technology is decided in its human fabric, not in its technical merit— is what turns that assertion into a finding and not a comfortable opinion. The scaffolding is not a collection of supporting citations: it is that convergence, turned into structure.

The foundation: the organization as a sociotechnical system

At the origin is the current that gives its name to everything else. Eric Trist and Ken Bamforth, at the Tavistock Institute, studied in the early fifties the mechanization of English coal mining and ran into an anomaly that founded a discipline: an objectively superior technology lowered productivity.¹ The explanation was not in the machine, but in that mechanization had dismantled the social organization of the work —the small, self-regulating teams that shared out tasks, covered absences and sustained morale underground— without putting anything in its place to fulfill that function. Hence the principle that orders the discipline: every productive organization has a social subsystem (people, roles, culture, relations) and a technical one (tools, processes, technology), and performance does not depend on optimizing each one separately but on their joint optimization. The two subsystems are optimized together or fail together; improving only one, however real the improvement, can degrade the whole.

The later tradition of sociotechnical design turned that finding into method. Enid Mumford, with the ETHICS method, added the operative corollary that the principle had left implicit: if the social subsystem is part of the system, the design must involve those who are going to use it.² Not out of courtesy nor in pursuit of consensus, but because only they know the real process the system will have to support; the participation of the users is requirements engineering. This current is the foundation of the argument: it justifies the central assertion —every organization is a sociotechnical system— and the first rule of the method, read the social before touching the technical.

The word “social” invites a misunderstanding, that of the soft, and it must be deactivated from the start. The social subsystem is not the climate or the motivation: it is the structure of how the work is organized —who decides, who covers for whom, how autonomy is shared out—. The finest proof came from later comparisons within the same mining. One same longwall technology performed differently depending on how it was organized: fragmenting the tasks, it fell; preserving the group’s self-regulation, it did not. The variable that moved the result was not the machine, identical in both cases, but the design of the organization of the work. It is the cleanest demonstration that the social subsystem is an engineering variable, not a residue to be administered.

The mechanism: why an artifact “works”

Sociotechnical theory diagnoses that there are two subsystems, but does not explain why the articulation between them is achieved or lost. That mechanism is contributed by actor-network theory, of Bruno Latour, Michel Callon and John Law. Its thesis is more radical than the sociotechnical one: the technical and the social are not two spheres that must be coordinated, but a single network of actors, human and non-human. An artifact does not “work” by its internal properties; it works when it achieves translation —enrolling the actors around it, aligning their interests enough that each one recognizes itself in it and finds a reason of its own to sustain it—.³ Success, then, is not an intrinsic property of the technology: it is a relational achievement, and where translation fails the artifact is neither good nor bad: it is left ignored.

This current puts a name to the verb of the thesis —to translate— and founds one of its concepts of its own, translation in flesh and blood: the translation is done by a person, not by a document. It also closes off an illusion of method, that adoption is a step subsequent to design or a training problem that gets solved afterward. If “to work” is to have translated, adoption is not an epilogue of the design: it is the very test of whether the design understood someone. Added together, the sociotechnical and the actor-network compose the thesis of the bridge: the problem is of two subsystems (Tavistock) and the solution is to translate between them (Latour).

Translation, moreover, is not a single act but a process with moments, which Callon described in his founding study: to problematize —to define the matter in such a way that the other actors need to pass through one—, to interest —to fix each actor in that role—, to enroll —that they accept it— and to mobilize —that they act accordingly—. Naming those moments matters for the method because it makes diagnosable where an adoption breaks: an initiative may have problematized well and fail at the enrollment, or enroll a few and mobilize no one. Translation is neither won nor lost as a block; it is won or lost in stretches, and each stretch is a concrete place to look.

The motor: leading the change, not decreeing it

The thesis of the bridge, posed this way, is a static diagnosis: it says there are two subsystems and that success is to align them, but it does not say how one intervenes in the social subsystem to achieve it. That verb is contributed by organizational change management, and that is why this current is the motor of the thesis: if the sociotechnical diagnoses, change management is the therapeutics.

The canonical skeleton comes from Kurt Lewin: unfreeze —create the need and uninstall the inertia—, change —the real transition— and refreeze —stabilize the new so that it does not revert—. John Kotter made it operative in eight steps, from urgency and the guiding coalition to the early wins and the anchoring in the culture.⁴ And Juan Díaz Barrios contributed the substance beneath the procedure: an integral change is sustained on delegation, communication, collaboration, participation and learning, not on the announcement.⁵ The common thread of the three is that the obstacle to change is human —habit, fear, meaning—, not technical, which links it immediately back to Mumford’s participation and to the enrollment of translation: getting the actors to recognize themselves in the new is, in practical terms, to unfreeze and freeze again.

Here it is necessary to stop and be honest, because this current is the most exposed to passing off as science what is codified practice. The “Lewin model” as it circulates is in good part a later construction: Lewin never wrote “refreeze” as a closed model —the idea remained in a minor subsection of his work—, as the rereading of Cummings, Bridgman and Brown documents.⁶ And Kotter’s eight steps are prescriptive literature of managerial practice, not research with a control group. The argument uses them as useful maps of a process, not as strong empirical theory; the empirical weight of the thesis is borne above all by the sociotechnical current and the field evidence of the closing, not this one. Change management contributes the verb, not the proof —and better to say it before a sharp-eyed reader points it out.

There is a corollary of this current that the argument uses more than the models themselves: the rereading of resistance. In ordinary language, “resistance to change” names an irrational obstruction that must be overcome. Read from the social subsystem, resistance is information: it says what the change threatens —a control that is lost, a competence that was ceasing to be valuable, a source of power written in no org chart—. Whoever treats it as an obstacle to remove loses the data; whoever reads it as a symptom finds the precise place in the social subsystem the design did not contemplate. Well-led change does not crush resistance: it deciphers it.

The descent into the AI era: human-centered AI

The three previous currents were born before contemporary artificial intelligence; one is needed that translates the sociotechnical principle to the era of the models. That is human-centered AI. Ben Shneiderman proposes to combine high automation with high human control and shows that they are not a trade-off: the best-designed systems are at once more autonomous and more governable, not one thing at the expense of the other.⁷ Virginia Dignum systematizes responsible AI —transparency, accountability, values built into the design—, and Cathy O’Neil documents the reverse: the concrete harm of opaque models that hide biases at scale and become unauditable just when they decide most about people’s lives.⁸

The core idea of the current is that AI must amplify, control and give confidence to people, not isolate or replace them. It is the direct translation of the sociotechnical thesis to the AI era, and the underpinning of the fourth link of the method and of two concepts of its own: AI as an extension of judgment and value in the person, not in the model. It is also the point where the theoretical scaffolding touches the normative frame: the auditability and the human control that HCAI poses as good design are, on the regulatory plane, what instruments such as the EU AI Act or the NIST AI Risk Management Framework require. Here theory and governance ask for the same thing in two different tongues.

The most precise form of Shneiderman’s proposal is a change of axis. Where common sense imagines a single lever —more automation is less human control, and inversely—, his frame separates that lever into two independent axes: the degree of automation of the machine and the degree of control of the person. The desirable quadrant is not that of much automation and little control, nor the inverse, but that of both high: very capable systems the person still governs. That separation of axes is what dissolves the false dilemma between power and control, and it is the technical translation of the sociotechnical principle of joint optimization: one does not choose between the machine and the person, one optimizes them together. The condition for that to be possible is explainability —a system that cannot be explained cannot be controlled and, therefore, is not adopted: it is tolerated, and at the first doubt it is abandoned—.

That change of axis resolves, in passing, the objection that the era of agents makes inevitable: if AI already executes on its own, what remains of “high human control”? The answer is that control does not require intervening in every decision —that, at scale, is impossible—, but governing the rules with which the agent decides: what it resolves on its own, when it escalates to a person, what it leaves recorded for audit. Control moves from the decision to the design. The more autonomous the system, the higher the other axis, not lower: autonomy without governance is not Shneiderman’s desirable quadrant, it is its opposite disguised as progress.

It is worth marking here a tension with actor-network theory, because the two currents hold up the argument from opposite assumptions. The actor-network is symmetrical by method: it treats humans and non-humans with the same vocabulary and refuses to privilege some over others in advance. Human-centered AI is asymmetrical by principle: it puts the person at the center as the ultimate criterion of value. Far from cancelling each other out, they share out the work. The actor-network describes how a system stabilizes in fact —who enrolls, what interests get translated—; human-centered AI dictates toward what configuration it should stabilize —the one that gives judgment back to the person—. One says how the world is; the other, toward where it is worth moving it. It is the same division that later regulates the use of Rogers: description and norm are different tools, and it is best not to confuse them.

The material requirement: data as fitness for use

A thesis on AI cannot stay within the organization and judgment: AI eats data, and data have a current of their own that the argument needs. Richard Wang and Diane Strong established that data quality is not technical accuracy but fitness for use in a context, and they decomposed it into dimensions that go far beyond correctness: intrinsic, contextual, of representation and of accessibility.⁹ Thomas Redman contributes the economic and management side: he documents that poor data quality is above all a problem of processes —not of technology— and that it costs most organizations a far from trivial fraction of their revenue, on the order of 15% to 25%.¹⁰

This current holds up the third link of the method and the concept of the data that lies: a piece of data can be technically impeccable and still lie, because its meaning depends on a human context that at some point was lost. That quality be defined by fitness for use, and not by accuracy, is what displaces the problem from the database to the organization: a piece of data is of poor quality when it does not serve the decision someone has to make with it, and that is not fixed with a script. At bottom, data quality is a social problem that manifests itself in a table.

Two consequences of defining quality as fitness for use deserve to remain explicit, because they are what make it a management problem. The first: quality is relative to the decision. The same set of data can be excellent for one decision and dreadful for another, without a single value changing, because the use changed; there is no “quality” data in the abstract. The second: quality is determined upstream, in the process that generates the data, not downstream, in the database that stores it. A piece of data is the trace of a human act —someone filled in a field with a criterion—, and if that act is badly designed or badly incentivized, no later cleaning recovers what was never recorded. That is why data quality is not attacked with data quality tools: it is attacked by intervening in the process that dirties them.

The prior condition: maturity and absorptive capacity

Before choosing what technology to incorporate, an organization has to be in condition to absorb it; that is the current of maturity. (A clarification of order: the scaffolding is laid out in sequence of construction —foundation, mechanism, motor—, but in the method this current goes first, because measuring maturity before buying is rule number one of the thesis.) Jan Jöhnk, Malte Weißert and Katrin Wyrtki systematized AI readiness as a multidimensional phenomenon: from a qualitative study they distilled 18 factors of readiness grouped into five categories —strategic alignment, resources, knowledge, culture and data—.¹¹ The lesson is that it is not enough to have technology; one has to have the organizational conditions to sustain it. Those factors ground the concept of sociotechnical maturity, the rule of measure maturity first and the theoretical basis of the IMIA model.

But the list of factors describes a state, and maturity is a process. The dynamic mechanism that Jöhnk lacks is contributed, thirty years earlier, by Wesley Cohen and Daniel Levinthal with absorptive capacity: an organization’s ability to recognize the value of external knowledge, assimilate it and apply it depends on the knowledge it already has, and is therefore cumulative and path-dependent.¹² From there follows something the snapshot of factors does not show: an organization with no history of data or analytics cannot “buy AI” and absorb it all at once, because it lacks the prior knowledge that would make it legible. Maturity is not the inventory of factors present —the snapshot—, but the learning capacity accumulated over time —the film—. That is why the diagnosis that opens the method reads trajectory, not only state.

Cohen and Levinthal’s mechanism moreover hides a property that inverts the usual intuition about adoption: absorptive capacity is self-reinforcing. Since it depends on prior knowledge, each investment in learning widens the base that cheapens the next learning; and, symmetrically, its absence perpetuates itself, because whoever did not accumulate cannot recognize the value of what they lack and so does not invest in acquiring it. Hence maturity cannot be skipped with a purchase: what is missing is not the tool but the trajectory that makes it exploitable, and that is not sold separately.

The pace: adoption as a social process

Up to here the argument has explained why adoption is achieved or lost; what is missing is what explains at what pace and in what order it occurs. Everett Rogers showed that the adoption of an innovation is not an event but a social process that unfolds over time and draws an S-curve, and he offered two lenses the argument uses again and again. The first distributes the population into five categories according to when they adopt —innovators, early adopters, early majority, late majority and laggards—, each with different motives. The second explains the speed: it depends on five perceived attributes of the innovation —relative advantage, which is the best predictor; compatibility; complexity; trialability and observability—.¹³ The decisive thing is the adjective: they are attributes perceived by the adopter, not objective technical properties of the artifact. It is the thesis of the actor-network said in another key: what Latour posed as relational, Rogers makes precise as perceived —the weight is not in the artifact but in the head of whoever decides to adopt it—.

The value of this lens is that it turns “the adoption fails” into something diagnosable: an AI initiative is not held up because the technology is bad, but because for its real users it scores low on compatibility —it clashes with how they work—, on complexity —they do not understand it— or on observability —no one sees the result—. And it explains a persistent pattern, which Geoffrey Moore named as the chasm: the pilot enthuses innovators and early adopters and dies on crossing toward the early majority, because it was designed for whoever tolerates friction and not for whoever demands that the thing simply work.¹⁴ Hence a rule of the method: design for the early majority from the start, not for the enthusiast.

No current demands as much caution as this one, because it is the one that most easily slides from description to prescription. Rogers is a diffusionist: he tends to assume that to adopt is good and that the only problem is to accelerate. The thesis of the bridge does not accept that assumption —the Kenya experiment and O’Neil’s work show that adopting without judgment can amplify harm—. Rogers is used in a descriptive way —how and at what pace something diffuses— but never normatively —to adopt is not an end in itself—. The normative filter is set by the human criterion of the previous current; without it, the S-curve may be diffusing harm at scale, and faster. (And it is not advisable to over-interpret the most-cited datum of Rogers: his synthesis of studies attributes to the five perceived attributes between 49% and 87% of the variance in adoption rates, but that is an aggregate range of the literature, not a constant measured in an experiment.)

Rogers contributes, lastly, two ideas the method treats as design levers and not as passive descriptions. The first is reinvention: adopters rarely take an innovation just as it comes, they modify it while adopting it, and that adaptation —far from being a deviation to correct— is usually the condition for the adoption to be sustained. The second is that two of the five perceived attributes, observability and trialability, are not fixed properties of the artifact but things the design can manufacture: a result that was hidden can be made visible, a thing can be left to try without commitment. Read this way, Rogers’s attributes cease to be a thermometer of why something is not adopted and become a list of interventions to get it adopted; each low attribute is a place where the design can work before the S-curve decides on its own.

The substrate: tacit knowledge and the construction of meaning

The seven previous currents form a flow, but there is an eighth that is not a step of the flow but the ground on which they all rest: that of tacit knowledge and the construction of meaning. Michael Polanyi formulated the idea with a phrase that became famous —“we know more than we can tell”—: much of expert knowledge is tacit, embodied in practice and impossible to codify fully in a manual or a document.¹⁵ Karl Weick contributed the complement: organizations do not find the meaning of facts, they construct it —retrospectively, socially and guided by plausibility rather than by accuracy— and they enact the environment they later measure; the categories with which they record the world end up molding the world they see.¹⁶

These two ideas are the epistemic godparents of the concepts of its own, and that is why the current runs beneath the others. Translation in flesh and blood is Polanyi carried to the bridge: reading an organization rests on tacit knowledge that is not transferred by document, so that the translation has to be done by a person and not by a deliverable —and it is also the reason the actor-network’s translation must be embodied—. The data that lies is Weick: a piece of data does not tell the truth on its own, it signifies within a situated act of construction of meaning; when that context is lost —or when the organization badly enacted the category it measures— the data remains impeccable and still deceives. Against Wang and Strong, who give the normative side —what quality data is—, sensemaking gives the descriptive side —why its meaning is fragile—. Anchoring the vocabulary of its own in these two currents takes it out of metaphor: it ceases to be a happy image and becomes a concept that can be defended.

Here too one must declare the position, because resting on Polanyi’s tacit installs a known tension with the knowledge management of Nonaka, for whom much of the tacit can indeed be made explicit and codified. The thesis takes sides knowingly: what is codifiable gets codified —hence the systems, the documentation, the models—, but the core of organizational judgment remains tacit, and hence the translation is done by a person. The codification is not denied; what is denied is that it exhausts the knowledge.

Each of these ideas has an internal structure, and it is there —not in the famous phrase— that it ceases to be image and begins to operate. In Polanyi, tacit knowledge has a “from-to” form: one relies on particulars one cannot state —the touch of the hand, the reading of a gesture— in order to attend to a whole one does recognize; that is why one knows more than one says, and that is why that knowledge is transferred by living alongside the practice, not by reading it. In Weick, the construction of meaning is a cycle: the organization enacts —acts on the environment and so partly creates it—, selects a plausible interpretation of what happened and retains it as a frame for the next time. The consequence for the data is severe: an organization does not measure a world that was there, it measures the categories it enacted, and afterward those categories give it back a world that confirms what it already believed. The data can lie not only by losing context: it can lie by having been constructed not to see.

And here this current touches O’Neil’s, a connection worth making explicit. If an organization measures the categories it enacted —with their blind spots included—, a model trained on those data does not learn reality: it learns the enacted categories, and afterward reproduces and amplifies them at scale. The bias of the model rarely is born in the algorithm; it is born in the prior sensemaking that fixed what is measured and what is ignored. The data that lies, automated, becomes a bias that scales. Hence a consequence for the method that is not obvious: the sociotechnical reading of the first link is also bias control, because it audits the categories before the model inherits them without asking where they came from.

The evidence that confirms it

The scaffolding would be a theoretical edifice if it did not rest on field evidence. It is not theory, but it is part of the argument, and three pieces suffice. The Kenya experiment showed that generative AI helped the high-performing entrepreneurs —around 15% more— and harmed the low-performing ones —around 8% less—, because the latter followed generic advice without the judgment to filter it: the field proof that AI without a human criterion to orient it can amplify inequalities instead of closing them.¹⁷ The Technology and Innovation Report 2025 of UNCTAD documents that the opportunity of AI for development coexists with a “compute divide” that excludes the smallest: inclusion is not automatic, it has to be designed.¹⁸ And the data of CEPAL show that the penetration of AI in Latin America is below 4% against more than 20% in Europe, but also that within the region the intra-country gap is enormous —in Brazil, 41% of large firms use AI against 11% of SMEs—.¹⁹ That gap between the large firm and the SME, as wide as the gap between regions, is precisely the terrain where the bridge works.

How they are articulated (the scaffolding in one image)

   sociotechnical (Tavistock/Mumford) ── the problem is of TWO subsystems
              +
   actor-network (Latour)             ── success is to TRANSLATE and align actors
              =  THESIS OF THE BRIDGE
              ↓ enacted by leading the change
   change (Lewin/Kotter/Díaz B.)      ── the bridge is crossed, not only drawn
              ↓ applied to the AI era
   HCAI (Shneiderman/Dignum/O'Neil)   ── AI amplifies the person or harms them
              ↓ which imposes three requirements
   data quality (Wang&Strong)         ── without faithful data, AI lies fast
   maturity (Jöhnk; Cohen&Levinthal)  ── without capacity to absorb, AI subtracts
   diffusion (Rogers/Moore)           ── adoption is a social process, not an event
              ↓ confirmed by evidence
   Kenya / UNCTAD / CEPAL             ── without judgment or design, AI widens gaps

Beneath this whole structure run Polanyi’s tacit knowledge and Weick’s construction of meaning: they are not a step of the flow, but the epistemic substrate that founds the concepts of its own —translation in flesh and blood, the data that lies— that run through every link. Read this way, the chapter does not enumerate eight authors: it shows a single conclusion held up by seven paths that did not know one another and an eighth that founds them from beneath. That is the force of the scaffolding, and the reason the thesis does not depend on any of the currents in particular.

See also: The agentic extension · The thesis of the bridge · The four links · Concepts of our own · The bridge applied to SMEs · Bibliography