The thesis of the sociotechnical bridge

This is the central argument of the book. If only one chapter were to be read, it should be this one: everything else —the method, the concepts, the applications— is its unfolding.

The thesis in a sentence

Every organization is a sociotechnical system; therefore, its technology cannot be transformed without first reading its human fabric. Whoever knows how to read both dimensions —and to build in both— occupies a position that is scarce by structure, not by chance.

The rest of these pages adds no new ideas to that sentence: it proves it, orders it and brings it down to concrete work.

The problem it names

Organizations arrive at artificial intelligence from very different starting points —which form a spectrum, not a single situation—, but they all end up in the same place.

At one extreme is apparent scarcity: the organization that believes it has nothing. Strictly speaking, pure zero hardly exists anymore —there are spreadsheets, a billing system, the chat where the real work is coordinated—; what is missing is not technology but design, because those systems grew on their own and live in people’s heads. The task is not to repair, but to build a first backbone, and the temptation is to treat that construction as a purchase.

In the middle are the islands: systems accumulated over the years that do not talk to one another, each area with its own version of the truth. The data exists, but fragmented; the task is not to have it, but to integrate it.

At the other extreme is the mature but fragmented organization, which already has solid systems and now needs to orchestrate them and take the leap —to add AI where it adds judgment, to govern it with confidence—. Here the obstacle is not scarcity, but inertia.

And crossing the whole spectrum appears the case that repeats most often: the project that did not pay off. A considerable share of digital initiatives does not produce the value they promised. It is rarely the technology’s fault; it is that it was installed on top of an organization no one read. A process that was broken was automated, a procedure that was redundant was digitized, AI was handed to a team that does not trust it, a dashboard was raised that no one looks at because it measures what is easy and not what matters.

In every case the underlying error is the same: to treat a sociotechnical problem as if it were a technical one. The question that decides the outcome is not “which model do we use?”, but “what does this group of people do, why this way, what do they gain and what do they lose if this changes, and who will resist?”. Without that question, the best engineering available amplifies the disorder instead of correcting it: automating a broken process only lets you make the same mistakes faster.

A concrete example helps, of the kind that recurs in the field. A mid-sized company decides to “put AI” into its customer service area: it buys an assistant that answers queries from its customer database. Technically, the system works. Two months later, no one uses it. The reason was not in the model: the database was filled in by three teams with different criteria —one used the “status” field for collections, another for shipping, the third left it empty—, so the assistant answered with data the staff themselves knew to be unreliable. The project did not fail because of the AI; it failed because it was built on data whose meaning depended on a human process no one had read. No improvement to the model would have saved it.

Another case, on different ground, with the same root. A city government digitizes a procedure that used to be done at the counter. The system is tidy, fast, well built. Six months later, the procedure still comes in on paper: the employees who handled it found a thousand reasonable motives —the elderly resident who cannot manage the screen, the exception the form does not contemplate— to keep the old channel alive. No one sabotaged anything openly. The system simply took away from one area the control of a process that was, without appearing on any org chart, its source of power and part of its reason for being. It had been read as a software problem when it was, from end to end, a problem of who decides. The technology worked; what had not been looked at was whom it displaced.

The other axis: the adoption ladder

The spectrum says where an organization starts from. The other axis is missing: toward where it leaps. Because adoption is not a switch that turns on all at once, but a ladder climbed step by step. At the bottom is the everyday office work. One step up, the systems that order the operation —an ERP that integrates administration, a CRM that gathers the customer relationship—. Higher up, predictive AI, which uses accumulated data to anticipate and not only to record the past. And on today’s step, the agentic layer: software that no longer only suggests, but executes tasks end to end.

What matters for the thesis is not the inventory of steps, but its slope: the higher the step, the dearer the cost of omitting the human reading. In office work, a badly-read system is simply avoided —people go back to their spreadsheet—. In the agentic layer, a process no one understood is not executed more slowly: it is executed on its own, at scale and with no witness —an agent that approves refunds, routes complaints or prioritizes case files according to a rule no one stopped to audit—. The error stops being a misfiled paper and becomes an automatic decision, repeated thousands of times before anyone looks at it. Technical sophistication does not soften the sociotechnical demand: it multiplies it. That is why the thesis does not age with the technology; each new step makes it more urgent, not less.

Why it is true (and not a comfortable opinion)

It is not a consulting intuition. It is an empirical finding with seventy years of backing, sustained by disciplines that do not talk to one another and that, nevertheless, reach the same conclusion.

The origin lies in the Tavistock Institute’s studies of the mechanization of the English coal mines. The new technology —the longwall method— was objectively superior, and even so it lowered productivity and raised absenteeism. The reason was not in the machine: it was that mechanization broke apart the social organization of the work. Before, the miners operated in small, self-regulating teams that shared out the tasks, covered absences and sustained morale underground; the new method fragmented them into specialized shifts dependent on one another, with nothing put back to fulfill that function.¹ The lesson founded a discipline: the technical and the social subsystem are not optimized separately, they are optimized together or fail together. Improving only one —however real the improvement— can degrade the whole. The later tradition of sociotechnical design carried that principle into method: a design that does not involve those who will use the system produces systems that people sabotage, avoid or misuse. The participation of users is neither courtesy nor a search for consensus: it is requirements engineering, because only they know the real process the system will have to support.²

The sociology of technology adds the mechanism that explains why it happens. An artifact does not “work” by its internal properties, but when it manages to translate the interests of the actors around it: to enroll them, to align them enough that each one recognizes themselves in it and finds a reason of their own to sustain it. Where that translation fails, the artifact is neither good nor bad: it is, simply, ignored.³ Adoption, then, is neither a step that comes after design nor a training problem to be solved later: it is the very proof of whether the design understood anyone.

In the era of AI the pattern worsens rather than disappears. Human-centered AI research shows that systems that isolate, obscure or replace the person generate rejection and harm, while those that amplify their judgment, give them control and earn their trust, pay off.⁴ And the field evidence confirms it: in an experiment with hundreds of entrepreneurs, generative AI helped the high performers and harmed the low performers, because the latter followed generic advice without the judgment to filter it.⁵ AI without a human reading to orient it does not level: it widens the gap that already existed.

Four perspectives, a single conclusion. The industrial sociology of the nineteen-fifties, the tradition of participatory design, the sociology of technology and contemporary AI research start from different objects —a coal mine, an information system, any artifact, a generative model— and are separated by decades; even so they converge on the same point: the performance of a technology is not decided on its technical merit, but on how it locks into the human fabric that receives it. That such distant disciplines arrive at the same thing is what makes this a finding and not an opinion. The question is no longer whether the human dimension matters, but who is in a position to read it.

The structural gap

If the problem is sociotechnical, the solution requires reading the human and building the technical. But the labor market is split into two halves that rarely touch.

On one side are those who understand people —processes, change management, organization—, but who do not build the system. They diagnose, deliver a careful report and withdraw; the code is written afterward by someone else who never read that report, nor sat down with the people it spoke of. On the other are those who build the system —developers, data engineers, architects—, but who rarely read the organization. They receive a requirement already translated, and badly, by a third party, and they set about optimizing, with all their craft, the wrong solution. The talent of neither side fails; what fails is that they never met.

Between the two halves opens a translation gap, and it is there that the project is lost. Each side attributes the failure to the other —“the business doesn’t know what it wants” against “the techies don’t understand the problem”— and both are partly right. The cost of that gap is not abstract: it is paid in systems no one uses, in data no one trusts and, measurably, in direct loss —poor data quality costs most organizations a considerable fraction of their revenue—.⁶ That liability is almost always the mark of a human process the technical side could not see and the organizational side could not correct.

Closing the gap is not achieved with an intermediary who coordinates the two parts: that is project management, and it leaves the translation in the hands of a document that degrades at every step. It is achieved with an integrated competence —the two readings exercised as a single practice— that reads the human fabric and builds the system without the meaning being lost in the handover. That practice does not translate between two specialists: it embodies the translation. It is scarce by structure, and hence its value.

The integration of two domains

Closing the gap requires bringing together two domains of competence that the market keeps apart:

• The reading of the social subsystem —the sociology of organizations: power, culture, incentives, resistances, real processes against formal processes.
• The construction of the technical subsystem —software and data engineering: auditable systems, pipelines, infrastructure, and the design and governance of AI, with the demand for auditability of a regulated environment.

Most professional profiles were trained in one of the two domains, and exercise it with competence within its limits. The paradigm calls for crossing that limit —exercising both domains as a continuum: diagnosis of the organization, modeling of the data, construction of the system and governance of the AI, each stretch tied to the previous one and answering for what the one before left behind—. That continuity, and not the sum of specialties, is what the method organizes in The four links.

Honesty criterion. The integration of competences is a structural and verifiable differential, not a proclamation: it rests on evidence —systems in production, MCP servers, the IMIA maturity model— and not on adjectives. Where a piece of data is an assumption, it is said. The paradigm is earned in the facts.

Implications

From the thesis follows, with no additional steps, everything the rest of the book develops:

1. First you measure maturity, then you choose the technology. The sociotechnical diagnosis precedes any purchase, it does not accompany it. Skipping it is the most expensive way to start: you acquire the right tool for the wrong problem, or the right one for a problem the organization is not yet in a position to sustain. That is why an instrument is needed to measure that starting point before deciding. → IMIA, the maturity instrument
2. Data quality is a social problem before a technical one. A “dirty” piece of data is rarely a data-entry error: it is usually the faithful mark of a human process badly designed or badly incentivized —three teams filling the same field with different criteria because it serves each of them for something else—. It is not cleaned with a script; it is corrected by intervening in the process that dirties it. → Concepts of our own
3. AI governance is not a brake; it is the condition of adoption. Auditability, traceability and human control are not opposed to the technology being used: they are exactly what makes people trust it enough to use it. A system that cannot be explained is not adopted: it is tolerated, and at the first doubt abandoned.
4. Value is measured in the person, not in the model. The deliverable is never a model in production nor a dashboard switched on: it is a better human decision, an hour freed from a task that did not deserve it, a gap that closes. When the metric shifts from the artifact to the person, it becomes impossible to confuse activity with result —and almost every project that does not pay off lives in that confusion—.
5. AI is a multiplier with a sign, and the bridge decides the sign. The same tool amplifies what it finds: on order and judgment, it multiplies value; on disorder and without a judgment to filter it, it multiplies the harm and widens the gap instead of closing it —exactly what the Kenya experiment showed, where AI helped those already doing well and harmed those doing badly—. The human reading is not an ethical supplement to the project: it is what inverts the sign of the multiplier, from amplifier of inequalities to closer of gaps.

Antithesis: what the thesis argues against

• Against technological solutionism: the belief that the problem is solved by buying the tool.
• Against diagnosis without construction: the organizational analysis that never reaches the system.
• Against technique without reading: the construction that does not understand whom it serves.
• Against AI enthusiasm: models adopted for fashion, measured by vanity metrics and not by value.

The position of the bridge is what remains once those four comforts are discarded.

See also: The four links · Concepts of our own · Authors and currents · Bibliography