Which tasks make up the footprint of AI, by occupation?

The figures that circulate on the electricity use of AI concern one question, one image, one video. For an organisation, the useful unit is rather a year of work: an accountant, a journalist or a developer do neither the same tasks nor at the same frequency. The AI Impact Database describes 18 occupations in this way, through a task basket: for each task, how often it recurs in the occupation, what share of it is done with AI, and the electricity it consumes at the servers.

These frequencies are assumptions made by Projet Celsius. For some of the tasks, they rely on the frequency reported by those who hold the occupation in O*NET, the occupation database of the US Department of Labor, which gives it in 7 classes, from "once a year or less" to "every hour or more". The share done with AI starts, for each task, from a value common to all occupations. It is adjusted to the occupation according to the Microsoft study of 200,000 conversations with Bing Copilot, which measures to which activities of each occupation AI applies and not how many people use it. It also relies on usage surveys: the Baromètre du numérique, the Stack Overflow survey of developers and the study by OpenAI and economists from Harvard and Duke on messages sent to ChatGPT. The transcription of interviews, which makes up half of the journalist's year and a third of the video producer's, rests on an assumption without a survey. The years on this page are therefore estimates ; they describe the current use of the database, that of the median user of an occupation who uses AI.

The video producer comes out on top with 7.4 kWh, ahead of the journalist, 4.6 kWh, and the developer, 4.3 kWh. The field occupation, technician or carer, comes last with 0.48 kWh: it mostly writes reports. Between the 2 extremes, the gap is from 1 to 15. The office worker, the generic profile of the database, sits in the middle, at 1.3 kWh; everyday personal use, outside work, at 0.75 kWh.

Each year has a range, which follows the low and high values of the electricity of each task, at equal frequencies: from 1.12 to 1.15 kWh for the lawyer, from 1.7 to 9.8 kWh for the graphic designer, depending on the model that generates the visuals. It does not cover the uncertainty on frequencies and on the share done with AI, which weighs more: the database also describes a frugal use, with about 3 times fewer tasks delegated to AI, and an intensive use, with 3 to 4 times more, which the entry for each occupation lets you choose.

The electricity of a task ranges from 0.07 Wh for a code suggestion to 340 Wh for a 10-second video clip, a ratio of almost 5,000. A quick question consumes 0.3 Wh, the transcription of a one-hour meeting with its summary 30 Wh, a 5-second clip 99 Wh, as much as 330 quick questions. An occupation weighs heavily when a task from the top of this scale recurs often in its year, or when a light task recurs tens of thousands of times.

For the journalist, 79 interviews transcribed and summarised make up half of the year: the database assumes 3 one-hour interviews per week, of which 60% go through AI. For the developer, these are 35,200 code suggestions, 200 opportunities per day, of which 80% with an AI suggestion; each weighs little, their number makes the total. The video producer transcribes as much as the journalist, 79 hours of raw footage or interviews in the year, or 32% of their year; their 29 generated clips make up 60% of it, and a single 10-second clip consumes as much as 1,100 quick questions.

In 6 of the 18 occupations, the office worker, the executive assistant, the human resources officer, the accountant, the sales representative and the communications officer, the heaviest task is web search delegated to an assistant: from 25% to 47% of the year. A search of this type consumes 2.7 Wh in the database, 9 times the 0.3 Wh that Google gave in 2009 for a conventional search: the assistant launches 2 searches and reads 5 pages before answering.

Emails, often cited, weigh little: 3 to 40 per day depending on the occupation, of which 7 to 13% written with AI in the database, at 0.23 Wh each. They make up 20% of the year of the manager, who writes the most, and less than 13% for the others. Our analysis of the actions quantifies what choosing a conventional search changes over a year, when it is enough.

The years above count an assistant that you question. With coding agents, which chain dozens of steps on their own, the order of magnitude changes: the database counts 594 hours of agent time in a developer's year, which takes it from about 4 to about 400 kWh. These hours depend on the tool that the organisation makes available; in the database, only the developer and the data scientist count them. The analysis over a year of use details the agent hour, the readings on which it relies and the formulas that count it.

With the world average electricity of 2025, 458 gCO2e per kWh, these years range from 0.2 kgCO2e for the field occupation to 3.4 kgCO2e for the video producer. If the servers ran in France, at 41 gCO2e per kWh according to Ember, they would range from 20 to 305 gCO2e; a Bilan Carbone® in France uses the ADEME value, 51.9 gCO2e per kWh. The video producer's year makes up 0.04% of the annual carbon footprint of a French person, 8.3 tCO2e in 2025; the developer's with agents, about 180 kgCO2e, makes up 2%.

For a team, the calculation is done by addition: 12 office workers represent 15 kWh per year in the database, about 7 kgCO2e. Our analysis of the Bilan Carbone® says where these emissions fall in an organisation's footprint.