For many organizations the starting point is not an excess of technology but its scarcity: management lives in scattered spreadsheets, data sits locked in silos that do not talk to each other, processes are manual, and decisions are made without the evidence in view. The opportunity is real and, for the first time, within reach of companies and institutions that could not afford it before: to order that information, connect it, and take the management and analysis of the business to another level with data architecture, agents and artificial intelligence.
But that opportunity hides a trap, the same one for those starting from scratch as for those who already tried to digitize and got no return: treating it as a technical problem. Buy the tool, connect the data, switch on the model. What decides whether the investment pays off or evaporates is not the technology chosen, but how much the organization was understood before installing it. Starting from little or no digitization does not reduce the risk: it concentrates it, because it forces you to build the technical subsystem and the human one at the same time. And the cost of getting it wrong does not show up in the code: it shows up in the budget and in the adoption that never comes.
This essay holds a thesis and draws its consequences for whoever decides: the problem of digital transformation is sociotechnical, not technical, and understanding it that way is the condition —not the ornament— of the return.
A sociotechnical problem, not a technical one
The underlying question is not “which model do we use?” but “what do these people do here, why this way, what do they gain and lose if this changes, and who will resist?”. Every organization is a sociotechnical system: a social subsystem (people, power, culture, incentives) and a technical one (processes, data, software) that only perform if they are designed together.
This is not a consultant’s hunch; it is a finding with seventy years of empirical support. The Tavistock Institute studies (Trist and Bamforth, 1951) showed that an objectively superior technology lowered productivity when it destroyed the social organization of work without replacing it; Enid Mumford’s ETHICS method turned it into a design discipline. The Human-Centered AI literature (Shneiderman, 2020; Dignum, 2019) restates the pattern for the age of algorithms. And actor-network theory (Latour, 2005) names it precisely: a system stabilizes —“works”— when it manages to translate and align the interests of the actors involved, not by virtue of its technical perfection.
The gap where projects die
The market, however, is split in two. On one side, those who understand people —processes, change management— but do not build the system. On the other, those who build —developers, data engineers, architects— but rarely read the organization: they receive a requirement already translated, and badly, by a third party. Between them lies a translation gap, and that is where the project is lost: each side blames the other, and both are partly right.
Closing that gap is not a matter of placing an intermediary between the two sides —that is coordination— but of holding a dual competence: reading the human fabric and writing the system, so that the knowledge does not degrade in the handover. A profile scarce by structure, not by chance.
The method: continuity without handoffs
The discipline I propose unfolds in four functions that form a loop that runs from the person to the data and from the data back to the decision. Their value is not in covering four specialties, but in that between them there is no handoff: the same reading that originates the diagnosis is the one that models the data and designs the AI. In a conventional team, information degrades at every handoff between specialists who do not understand each other; here, the hypothesis about the organization is preserved end to end. The same path serves both to fix what already exists and to build from little digitization: in that case there is no prior system to repair, but silos to connect and information foundations to lay before putting anything on top.
| Function | What it contributes | Foundation |
|---|---|---|
| Organizational reading | Maps power, culture, incentives and real processes; tells the stated problem from the real one. | Sociotechnical tradition (Tavistock, Mumford); Díaz Barrios (2005). |
| Software engineering | Turns that reading into auditable, maintainable systems people use every day. | Baxter & Sommerville (2011): iterative sociotechnical design, with users. |
| Data engineering | Guarantees reliable data that is meaningful in its context: without it, AI only automates errors. | Wang & Strong (1996); Redman (2008): quality is contextual, not only technical. |
| AI architecture | Designs solutions that amplify human judgment and return value to the decision, not models for fashion’s sake. | Dignum (2019); O’Neil (2016): responsible, auditable AI. |
What the evidence says
Prudence is not a posture; it is what the data show, on two planes. At the individual level, a field
experiment with 640 entrepreneurs in Kenya (Otis et al., 2024) showed that access to a generative-AI
assistant improved the performance of high performers (+15%) but worsened that of low performers
(−8%): without the human judgment to filter it, the same tool widens the gap instead of closing
it. What decides is not the model, but the human system around it.
At the aggregate level, the asymmetry repeats across organizations. UNCTAD (2025) warns of a “compute divide” that excludes the smallest organizations, and AI penetration in Latin America does not reach 4% —against more than 20% in Europe—, with internal distances just as large (in Brazil, 41% of large firms versus 11% of SMEs; CEPAL, 2024). The implication for whoever decides is direct: the return on AI does not depend on adopting it, but on adopting it where the organization is ready to sustain it.
Measure before investing
That is why the first deliverable is not technology but a diagnosis: a sociotechnical reading that separates the stated problem from the real one and rules where AI pays off and where it only destroys value. Automating low-value work frees up hours; measuring maturity first avoids the spending that does not pay back.
The diagnosis does not curb ambition: it orders it into a sequence. First the silos are connected into an information architecture that lets the data talk to each other; on top of it a base of governed, reliable data is built, meaningful in its context; and only on top of those foundations are agents and AI introduced, and only where there is already someone to govern and use them. Order matters: putting the agents before the foundations does not speed up the return, it compromises it. This way, taking management and analysis to another level stops being a catalog promise and becomes a construction with a foundation.
The same principle holds in the public sector, where judgment becomes a condition: what cannot be audited cannot be used. There, AI does not replace the civil servant; it amplifies their capacity to manage and serve, on interoperable data and auditable models that incorporate the perspective of those who use them. Institutional modernization is not bought in the form of software: it is built by strengthening capabilities and putting the recipient —customer or citizen— at the center.
The foundations
Five currents sustain this argument; each contributes a single idea:
- Sociotechnical systems (Trist and Bamforth; Mumford): the social and the technical subsystems are optimized together or fail together.
- Actor-network theory (Latour; Callon): technology stabilizes when it translates and aligns the interests of the actors.
- Human-Centered AI (Shneiderman; Dignum; O’Neil): AI must amplify, control and earn the trust of people, not isolate them.
- Contextual data quality (Wang & Strong; Redman): quality is fitness for use; without context, data lies.
- AI readiness (Jöhnk et al., 2021): readiness is strategy, culture, talent and governance, not just infrastructure.
The implication for whoever decides
The conclusion is not that AI should be avoided, but that its value is conditional: it appears when the organization is understood before it is programmed. Applied well, it does not replace people —it frees up hours, improves decisions with evidence and closes gaps instead of widening them—; applied badly, it does the opposite, faster and at greater scale. The difference between one case and the other rarely lies in the model: it lies in the discipline with which the human system was read before touching the technical one.
Far from making this demand obsolete, agentic AI —software that no longer suggests but acts on its own— makes it more urgent: the more the tool decides by itself, the costlier the process no one stopped to understand, because it runs at scale and without a witness. Human judgment does not disappear; it moves to the design and audit of the rules the AI decides by.
Diagnosing the organization, modeling the data, building the system and governing the AI as a single path —without the meaning getting lost at the edges— is the practice I offer to companies and institutions that want their AI to deliver, not just to function.