"AI is not profitable" - Before AI, think in processes

These are soft truths. They do not force the organisation to look at the place where the problem usually resides: its inability to describe, with precision, what it actually does.

• Versión española: https://www.linkedin.com/pulse/es-que-la-ia-rentable-o-antes-del-proyecto-el-proceso-beatriz-vallina-qbhcf

The obstacle is not that people do not know how to draft prompts, but that many organisations do not have explicit processes: what they really have are administrative traditions, sometimes with the appearance of process. They have habits stabilised through repetition, tolerated shortcuts because someone "knows how this works", spreadsheets functioning as oral memory and even as word processors, exceptions converted into routine, and decisions made according to criteria nobody has formalised because, at some point, they lived inside somebody's head. Where there should be workflow, there is operational folklore; and on top of folklore one can, of course, place a chatbot.

Quite simply, the conversation about generative AI goes wrong when it begins with the tool. The issue is not which model to use, which vendor has the best demo, or how many hours of training the workforce needs in order to "make the most of AI". The prior question, less marketable, is another one: what process actually exists? What are its inputs and outputs, and which decisions does it "contain"? What information does each decision require? Why, and according to which criterion, is a result deemed "correct", and which exceptions are accepted? And, of course, which part creates value and which part is merely sedimented and automated bureaucracy?

Without that prior work, the use case is theatre; for the moment, cheap little theatre. A task is selected because it looks modern, because it produces an attractive prototype, or because it fits neatly into a presentation on "quick wins"; nobody knows whether that task belongs to a relevant process, whether it addresses a real bottleneck, whether it displaces a critical decision, or whether it automates bad habits.

The shiny new toy automates error hyper-efficiently. This is not a minor matter: the danger of organisational cosmetics is that they may accelerate what was already badly designed.

The literature on organisational readiness for AI, technology adoption and business process management does not quite say this, nor with this much bite, but it points in a compatible direction: readiness is not merely technological, nor merely individual. It includes data, alignment between business and technology, the quality of information flows, learning capability and fit with existing processes. In the adoption of generative AI in software engineering, for instance, compatibility with existing workflows appears as a decisive factor. The main users of generative AI, namely developers, seem to understand this rather clearly. In business process management, AI makes sense when it is integrated into capabilities for automation, improvement, learning and redesign; it does not make sense when it is released like an aerosol over an opaque organisation.

Training, therefore, matters, but not as it is usually told. Training a person to use AI in a poorly described process, if indeed it has ever been described as such a thing, as a process, is teaching them to accelerate in the fog. It may improve their local productivity, in the best of cases; perhaps they write faster, summarise faster, search faster. We do not know whether they write well, summarise well or search well. We do not know where the work is going, according to which criterion it is decided, who validates it, what risk is accepted, and which part of the result can be repeated without depending on individual talent, which is not, incidentally, a negligible matter; individual talent is what saves even an AGI from its non-omniscience.

Competence does not replace design: let us not confuse everyday heroism with systems.

The honest itinerary has less glitter, but it is also far more powerful: map processes, atomise tasks, identify decisions, formalise criteria, redesign the workflow and, only then, introduce AI where it really makes sense (really). This means, exactly, freeing human judgement from having to sustain in memory and spreadsheets what should have been made explicit in the system.

There is a political, almost moral, difference between putting AI into an organisation and redesigning an organisation so that some tasks can be assisted by AI without destroying responsibility. The first buys a promise; the second requires the organisation to describe its work with an honesty that many institutions avoid because it reveals too much: duplications, controls nobody uses, ceremonial approvals, data that arrive late, decisions without an owner, exceptions that are no longer exceptional, and roles that survive because the process was never written down.

Little is said about this because it does not sit well in an innovation seminar: it is easier to say "we need upskilling" than "we do not know how we work". The first distributes blame politely among everyone; the second points to a design debt. But if generative AI has any serious value for organisations, it will not be the value of covering that debt with fluent text.

Cautionary note

The claim is not that training or use cases are irrelevant. The claim is that they arrive too late if there is no sufficiently good description of work in place beforehand. The available evidence supports the importance of organisational readiness, data quality, fit with existing workflows and business process management capabilities; it does not, by itself, turn process mapping into the single cause of success or failure.

References of interest

• Jöhnk, J., Weißert, M. & Wyrtki, K. (2020). "Ready or Not, AI Comes: An Interview Study of Organizational AI Readiness Factors". DOI: 10.1007/s12599-020-00676-7.

• Uren, V. & Edwards, J. S. (2023). "Technology readiness and the organizational journey towards AI adoption: An empirical study". DOI: 10.1016/j.ijinfomgt.2022.102588.

• Russo, D. (2024). "Navigating the Complexity of Generative AI Adoption in Software Engineering". DOI: 10.1145/3652154.

• Abbasi, M. et al. (2024). "A review of AI and machine learning contribution in business process management". DOI: 10.1108/BPMJ-07-2024-0555.

• Zebec, A. & Stemberger, M. (2024). "Creating AI business value through BPM capabilities". DOI: 10.1108/BPMJ-07-2023-0566.

Research by: Beatriz Vallina, PhD

Thesis Supervisors: Roberto Cervelló, Prof.PhD & Juan José Lull, PhD

Institution: Doctorate in Agrifood Economics, Universitat Politècnica de València

© 2026. All rights reserved.