The hidden economy of prompting
The literature on AI productivity has focused on speed of execution. Almost no study measures what happens before the task begins.
In 2023, the McKinsey Global Institute estimated that generative AI tools could add between $2.6 and $4.4 trillion annually to the global economy. The figure is widely cited; the implicit condition, far less so: that this productivity depends on teams knowing what to ask the system. A prompt is not a technical formality. It is the transfer of human thought to the machine — and when that thought does not yet exist, the transfer cannot be completed.
The iteration cycle as invisible cost
Every organisation working with generative AI has developed, without naming it, an informal trial-and-error process: a prompt is written, the output falls short, it is rewritten, iterated. What that cycle rarely accounts for are its real costs. Not only the time of whoever is iterating, but the accumulated cognitive load, the progressive degradation of judgment under the pressure of finding something usable — and, frequently, the acceptance of a mediocre result from sheer exhaustion rather than conviction.
The data in the following first chart are not surprising in their direction — everyone senses that a more precise prompt requires fewer rounds — but in their magnitude: the gap between a vague prompt and a specific one is not a marginal adjustment. It is a difference of scale that, multiplied by a team’s task count over a year, stops being an individual problem and becomes a management variable.
What mediocre output does to the metrics that matter
When AI output is published without sufficient editorial judgment, the consequences are rarely immediate. They are cumulative. Nielsen Norman Group research on digital behaviour has documented that a brand’s perceived consistency — its ability to maintain a recognisable voice and point of view over time — is one of the most determinant factors in sustained reader attention and trust. Generic output does not destroy a brand; it dilutes it.
The second chart puts numbers to this intuition. The performance gap between content generated with a generic prompt and a specific one is not marginal across any of the metrics considered. And that gap cannot be attributed solely to model quality: it reflects, above all, the quality of the brief the human formulated.
The cost the industry has preferred not to measure
In 2019, Emma Strubell and her colleagues at the University of Massachusetts Amherst published a study at ACL that for the first time quantified the energy consumption of large language processing models. Their findings — that training a model of BERT’s scale could be equivalent in emissions to over 600 kg of CO₂ — sparked debate, but also a tendency to consign the energy question to the domain of training. Inference, the everyday act of making a query, seemed negligible.
Subsequent research has qualified that impression. Luccioni et al. (2023) showed that at the scale of real-world use, inference accumulates a significant impact: small per call, but multiplied by millions of daily queries. For a team of ten who iterate nine times where two would have sufficed, the waste is not abstract: it amounts — according to the estimates in the third chart — to the CO₂ of hundreds of kilometres by car per year.
Who: the exact recipient, not the segment
A well-built brief does not say ‘millennials interested in sustainability’. It identifies a specific person: what they already know, what objection they hold, what they are reading this week. Without that level of specificity, the AI produces the lowest common denominator of the internet — which is precisely what every competitor is also getting.
What: the thesis, not the topic
There is a difference between asking for ‘an article on leadership in uncertain times’ and asking for ‘a text that argues communicative clarity is the only form of leadership that works when data is ambiguous’. The first produces content. The second produces a point of view.
How: the constraints that liberate
Editorial constraints — tone, length, what must not appear, voice references — do not restrict the AI: they orient it. A model without constraint instructions converges toward the safe, the balanced, what neither offends nor convinces. Constraints are the way of asking it to take a position.
Three questions before opening any tool
The solution requires no more sophisticated tools or more powerful models. It requires, before opening any interface, taking time to answer in writing: who exactly is this for, and what does that person already know? What position are we taking, beyond ‘giving information’? What is the one change we want to produce in whoever reads it?
These are not new questions. They are the same ones any copywriter, strategist, or designer should answer before beginning any work. The novelty that generative AI adds is not making that prior process obsolete: it is making visible, with a clarity that did not exist before, when that process has not taken place. The prompt is not the start of the creative process. It is its result.
FOR THIS WEEK...
How many iterations did you need for your last AI task?
Want to apply these ideas to your business?
Tell me what you’re working on and what you’d like to improve. We can talk it through in a free 30-minute call or, if you prefer, you can send me a message.
REFERENCES
International Energy Agency (IEA). (2023). CO₂ Emissions in 2022. Paris: IEA Publications.
Luccioni, A. S., Viguier, S., & Ligozat, A.-L. (2023). Estimating the Carbon Footprint of BLOOM, a 176B Parameter Language Model. Journal of Machine Learning Research, 24(253), 1–15.
McKinsey Global Institute. (2023). The economic potential of generative AI: The next productivity frontier. McKinsey & Company.
Nielsen Norman Group. (2022–2024). Research on content quality, readability, and digital trust. nngroup.com.
Strubell, E., Ganesh, A., & McCallum, A. (2019). Energy and Policy Considerations for Deep Learning in NLP. Proceedings of ACL 2019, 3645
Comments