← All posts

AItoolsresearch

Toolobesity? When writing a tool costs nothing

Generative AI makes writing internal tools almost free, and maintaining them too. The cost doesn’t vanish: it shifts to whoever has to find them.

Since spring, I’ve been building the internal tools of a small industrial company on my own, with an AI. I eventually counted what I had produced. The result surprised me: 535 features, for ten users, in under eight months. I thought I had built far fewer.

That number made me want to understand what had happened, and I’m drawing a research article in information and communication sciences from it. Here is the idea, ahead of the results.

What AI makes free, and what it leaves untouched

It is often said that generative AI cuts the cost of building software. That is inaccurate. It cuts the cost of writing it. Recent work on AI-assisted programming describes this shift: the developer’s job becomes a mediation of intent (Meske et al., 2025). Someone who can say precisely what they want gets in a few hours what would have taken weeks.

But a tool is more than its code. Before writing it, you have to know what people need, which means talking to them, gathering documents, understanding the rules of the trade. After writing it, someone has to actually use it. These two costs, the cost of the conversation and the cost of appropriation, have not moved. The sociology of design showed this long before AI: an object only helps people design together if it is built together, like a draft. Delivered finished, it prescribes (Vinck, 2009).

The result is mechanical. For a few months, you produce very quickly everything you can specify alone. Then the pace drops, not from fatigue, but because the remaining needs call for exchanges that AI does not speed up. That is what happened in my case, and I call this episode tool inflation.

The idea of inflation is not new for management tools. Detchessahar and Journé (2007) already spoke of an inflation of instrumental devices, and Kalika, Boukef Charki and Isaac (2007) of a millefeuille in which media pile up without replacing one another. What changes is the speed, and the fact that one person is enough.

A stock that nothing limits

The most counter-intuitive part comes next. In classic development, every function added will have to be maintained, and that cost slows accumulation. You think twice. But a stock written by an AI costs almost nothing to maintain, since the same AI maintains it.

So the brake disappeared along with the cost of writing. Nothing on the producer’s side limits the number of tools any more.

The cost has not gone away, though. It has moved to the person who searches. Faced with several hundred functions, the difficulty is no longer having the tool, it is knowing that it exists and where to find it. This is exactly what happened to information: it costs nothing to store and costs to find. We called that infobesity, a word coined in the early 1990s by analogy with obesity (Sauvajol-Rialland, 2014).

Hence a word to put to the test, with its question mark: toolobesity. It would describe the state left behind by tool inflation, when the cost of a stock of tools has passed entirely from the person who builds to the person who receives.

Grafted tools, soilless tools

A distinction helps explain why some tools catch on and others do not.

A grafted tool lives inside software that people already use: a macro in their spreadsheet, a button in their email client. It requires no move. A soilless tool lives elsewhere: you have to open another application, learn where things are, change a habit.

The difference matters, because a soilless tool shows its cost before its benefit. You immediately see what it will demand, and you cannot yet imagine what it might bring. An old law of documentation, Mooers’ law, puts it bluntly: a system goes unused when using it seems more painful than doing without (recalled by Boudokhane, 2006). The word is meant in the gardeners’ sense, not as a jab: a soilless crop grows very well, provided you supply everything the soil used to provide. For a tool, that means landmarks, documentation, guides. We often build soilless for a simple reason: existing software does not lend itself to this kind of addition, and it is simpler to build alongside.

What I take from it for now

The question I started with was: how do we get all these tools adopted? I now think it is badly posed. It assumes the problem lies with those who receive. Work on non-use invites us to turn the question around: not using something may respond to excess, rather than betray a delay (Jauréguiberry, 2012).

The problem lies rather in a gap in speed. Writing now goes much faster than exchange, and nothing automatically closes that gap. The responses that come spontaneously, documenting, guiding, adding an assistant, are themselves more tools.

The article in progress measures this production, describes what it leaves behind, and discusses these responses. I will publish it here early next year. Until then, if you also build internal tools with an AI, I am curious to know whether you recognise this picture.

Use of AI

The tools discussed here were written with AI assistance. This post was drafted with the help of Claude (Anthropic), based on exchanges with the author, who reviewed and validated its content. The cited references were read in the original text.

References

  • Boudokhane, F. (2006). Comprendre le non-usage technique : réflexions théoriques [Understanding technical non-use: Theoretical reflections]. Les Enjeux de l’information et de la communication, 2006(1), 13–22.
  • Detchessahar, M., & Journé, B. (2007). Une approche narrative des outils de gestion [A narrative approach to management tools]. Revue française de gestion, 174, 77–92.
  • Jauréguiberry, F. (2012). Retour sur les théories du non-usage des technologies de communication [A review of theories of non-use of communication technologies]. In S. Proulx & A. Klein (Eds.), Connexions : communication numérique et lien social (pp. 335–350). Presses universitaires de Namur.
  • Kalika, M., Boukef Charki, N., & Isaac, H. (2007). La théorie du millefeuille et l’usage des TIC dans l’entreprise [The millefeuille theory and the use of ICT in the company]. Revue française de gestion, 172, 117–129.
  • Meske, C., Hermanns, T., von der Weiden, E., Loser, K.-U., & Berger, T. (2025). Vibe coding as a reconfiguration of intent mediation in software development: Definition, implications, and research agenda. IEEE Access, 13, 213242–213259. https://arxiv.org/abs/2507.21928
  • Sauvajol-Rialland, C. (2014). Surcharge informationnelle en entreprise : l’infobésité, réalité ou illusion ? [Information overload in companies: Infobesity, reality or illusion?]. Cahiers de la documentation, 2014(1), 5–10.
  • Vinck, D. (2009). De l’objet intermédiaire à l’objet-frontière : Vers la prise en compte du travail d’équipement [From intermediary object to boundary object: Towards taking equipment work into account]. Revue d’anthropologie des connaissances, 3(1), 51–72.