Generative AI is everywhere - summarising sustainability reports, parsing bond issuances, untangling ESG disclosures. And with that ubiquity comes a fair question from climate professionals:

"Am I inflating my footprint by using these tools?"

Short answer: yes, there's an impact.

Longer answer: it's relatively small, and may not the right target for your climate efforts. The impact is comparable to small behavioral changes like switching to LEDs or hang-drying clothes in summer. You may want to target higher-impact activities first and not risk falling behind in technology in your professional context.

Good news: AI is getting more optimised, cleaner and leaner over time (all the way from clean energy used, training + inference of model providers, and the efficiency of ClimateAligned's assessment tools).

This article covers the environmental impact of AI usage in professional sustainability and climate finance work - from individual document analysis to large-scale assessment workflows using the ClimateAligned AI tools.

---

The paradox: huge growth, and lots and lots of prompts

Two truths live together here:

• AI is driving up global energy needs. Usage has exploded - ChatGPT hit billions of daily prompts faster than any tech in history, and AI is now baked into search, coding, writing, and analytics. The aggregate energy use is large.
• Individual prompts are minuscule. Data centres are processing oceans of very small sips. Each query is fractions of a watt-hour, but multiplied by billions, it adds up.

This isn't unique to AI. When the internet took off, total data-centre consumption rose - not because websites were wasteful, but because billions of people suddenly spent hours online.

Even now, total data centre electricity use only accounts to ~1.5% of global electricity according to the International Energy Agency - astonishingly efficient for something the whole planet relies on. While this percentage is projected to grow significantly as AI adoption accelerates, the efficiency gains in both hardware and software continue to improve the energy-per-computation ratio.

---

The per-prompt reality

Google recently did the most rigorous production-fleet measurement yet in their comprehensive study "Measuring the environmental impact of delivering AI at Google Scale".

Here's how much, the median text prompt consumes according to Google's study:

• 0.24 Wh energy
• 0.03 gCO₂e
• 0.26 mL water (on-site consumptive cooling only)

That's about 9 seconds of TV power and 5 drops of water.

However, for a more complete picture, we've adjusted the water figure to include indirect consumption. Google's 0.26 mL covers only on-site consumptive cooling (WUE applied to IT energy). Using the LBNL "2024 United States Data Center Energy Usage Report" factor for indirect water consumption from electricity generation (4.52 L/kWh US national average), the extra off-site water per 0.24 Wh prompt is approximately 1.085 mL.

This gives us an adjusted median-prompt water consumption of ~1.35 mL - about 5.2x Google's on-site-only figure, or roughly 25 drops of water per prompt. While this mixes scopes (on-site + off-site), it provides a fuller "water-through-the-meter" view of AI's water impact.

So when we scale up AI workflows, we're working with very small building blocks.

---

The per-page reality

Whether we like it or not, everyone still uses pdfs and word docs. But how much do these file types actually use when it comes to AI?

To be conservative, we'll assume the median prompt uses 50 tokens according to various AI benchmarking studies (see references at the end). While these studies report averages rather than medians, the datasets are described as "heavy-tailed" - meaning a few very long conversations skew the average upward. This suggests the median would be lower than the average. The WildChat dataset (296 tokens average) could be an anomaly compared to typical Google Gemini usage patterns. Thus, 50 tokens is a fair conservative estimate for the median, ensuring our energy calculations err on the side of caution.

A token is a piece of text, like a word or part of a word, that AI uses to process language. And there are about ~1,000 tokens per page on average based on typical document analysis.

So one page ≈ 20 prompts.

Thus our baseline numbers to use for every per page analysis:

• 4.8 Wh energy
• 0.60 gCO₂e
• 27 mL water (includes indirect water use)

That's about the equivalent of microwave heating an espresso in Belgium.

---

Real-world document median lengths

Now, let's shift our attention to some documents that we may be using as analysts. Annual reports are the longest whilst bond documents are often the shortest.

(we use median for the purposes of "typical" - note that these represent the most relevant pages for analysis, excluding boilerplate text and graphics)

• Annual Reports: ~160 pages
• Sustainability Reports: ~80 pages
• Climate/TCFD: ~30 pages
• Bond documents: ~12-16 pages

Document lengths vary significantly across sustainability reporting types

---

Case studies

Here's a breakdown of how much different workflows consume, from daily tasks to quarterly reports, with real-world equivalents for perspective. These examples represent typical document evaluation workflows for extraction and assessment purposes in sustainability and climate finance for financial professionals.

Case 1: One bond doc (10 pages*)

*Real-world page usage skips boiler-plate text and graphics.

• 48 Wh energy
• 6 g CO₂e
• 270 mL water

≈ 1/4 of a full kettle's energy, 1/6 of boiling a kettle's CO₂e, 1/6 of a full kettle's volume

Case 2: Annual 100-bond impact report (small fund-level analysis)

• 4.8 kWh
• 600 g CO₂e
• 27 L water

≈ 1/8 of daily household energy, driving 2 miles, 1/5 of daily Londoner water use

Case 3: One company assessment* (~50 relevant pages)

*Real-world page usage skips boiler-plate text and graphics. Retrieval-augmented generation is used to focus only on pages with relevant content.

• 240 Wh energy
• 30 g CO₂e
• 1,350 mL water

≈ a bit more than a full kettle's energy, less than boiling a kettle's CO₂e, 3/4 of a full kettle's volume

Case 4: 100 company assessments

• 24 kWh
• 3 kg CO₂e
• 135 L water

≈ 3/5 of daily household energy, driving 12 miles, nearly a day's Londoner water use

Daily use vs Quarterly use Comparisons

Daily vs quarterly AI usage patterns compared with real-world equivalents

---

Lifestyle equivalents: silly numbers edition

500 million users generate approximately 2.5 billion prompts daily (Axios report) - averaging 5 prompts per person per day. Over the course of a year, this amounts to 1,825 prompts per person, resulting in an estimated 54.75g CO₂e emissions annually.

Here's some silly equivalents to put that number into perspective:

What you'd give up for a year of AI prompts: surprisingly little!

References: Average banana CO2e: source | Boiling a kettle CO2e: source | Shower CO2e: source | Sending emails CO2e: source | London Tube usage CO2e: source

---

Heavy AI usage by a sustainability professional compared to other lifestyle choices

Annual AI footprint: professional sustainability work

For someone using AI daily in sustainability work, here's a realistic annual breakdown:

• 10 company assessments per week (50 pages each): 520 x 30g = 15.6 kg CO₂e (per year)
• Quarterly bond report (100 bonds): 4 x 600g = 2.4 kg CO₂e (per year)
• 100 median prompts daily (general use): 365 x 100 x 0.03g = 1.1 kg CO₂e (per year)
• Other workflows (document summaries, analysis): ~0.9 kg CO₂e (per year)

Total: ~20 kg CO₂e annually

---

So… should you quit AI for the planet?

No. The footprint of your personal AI usage is quite modest — comparable to small behavioral changes like switching to LEDs or hang-drying clothes in summer and may not be the right target for your climate efforts.

What does matter:

• Pushing vendors toward cleaner grids and better water efficiency.
• Supporting standards and transparency in reporting.
• Using AI where it amplifies evidence-based work (instead of wasting time doomscrolling).

Closing callback: If you really want to shrink your climate footprint, focus on bigger wins like transport choices, home energy efficiency, or supporting clean energy policy before turning luddite.

Fun fact: The average per-capita carbon footprint in the UK is about 23kg CO₂e per day. Even with extensive AI usage, your daily AI footprint is likely under 100g CO₂e—that's less than 0.5% of your total daily emissions. Probably not the best place to start looking for impact.

---

Warning: Call yourself a nerd? For those who really want to go deep on the numbers, start here.

---

Technical Deep Dive

Constants & Assumptions

Core AI Impact Data (Google Production Measurements):

Per "median prompt" (Google production telemetry across serving stack):

• 0.24 Wh energy
• 0.03 gCO₂e
• 0.26 mL on-site cooling water (Measuring the environmental impact of delivering AI at Google Scale)

Other relevant conversion details:

• Median prompt size (conservative estimate): 50 tokens
• Page conversion: 1,000 tokens per page → 1 page = 20 prompts (x20 multiplier)

Water Scope Adjustment:

• Indirect water from power generation: 4.52 L/kWh (US national average)
• Additional water per 0.24 Wh prompt: 1.085 mL
• Total water per prompt: 0.26 + 1.085 = 1.35 mL (LBNL 2024 United States Data Center Energy Usage Report)

Comparison Baselines:

• UK grid carbon intensity: 211 gCO₂e/kWh (Our World in Data)
• Kettle energy: 0.11 kWh/L → 1.7L full kettle = 187 Wh (Heatable)
• Daily household energy: ~40 kWh/day (Ofgem TDCVs: 2,700 kWh electricity + 11,500 kWh gas/year) (Ofgem)
• Daily Londoner water use: ~144 L/person/day (Greater London Authority)
• UK car emissions: 0.167 kgCO₂e/km → 10 miles = 16.1 km ≈ 2.7 kgCO₂e (National Rail Assets)

Per-page baseline (50-token median): 4.8 Wh, 0.60 gCO₂e, 27 mL water (including indirect water use)

---

Detailed Calculations

Case 1: Single Bond Document (10 pages)

Assumptions:

• 10 relevant pages (skipping boilerplate text and graphics)
• 1 page = 20 prompts (1,000 tokens ÷ 50 tokens/prompt)

Calculations:

• Prompts: 10 pages x 20 prompts/page = 200 prompts
• Energy: 200 x 0.24 Wh = 48 Wh
• CO₂e: 200 x 0.03 g = 6 g CO₂e
• Water: 200 x 1.35 mL = 270 mL

Comparisons:

• Energy: 48 Wh vs full kettle 187 Wh (≈ 1/4 kettle)
• CO₂e: 6g vs kettle CO₂e 39g (0.187 kWh x 211 g/kWh) (≈ 1/6 kettle)
• Water: 270 mL vs full kettle 1,700 mL (≈ 1/6 kettle volume)

Case 2: 100-Bond Impact Report

Assumptions:

• 100 bonds x 10 relevant pages per bond

Calculations:

• Prompts: 100 bonds x 10 pages x 20 prompts/page = 20,000 prompts
• Energy: 20,000 x 0.24 Wh = 4.8 kWh
• CO₂e: 20,000 x 0.03 g = 600 g CO₂e
• Water: 20,000 x 1.35 mL = 27 L

Comparisons:

• Energy: 4.8 kWh vs daily household 40 kWh (≈ 1/8 daily household)
• CO₂e: 600g vs driving 10 miles 2.7 kg (≈ driving 2 miles)
• Water: 27L vs daily Londoner 144L (≈ 1/5 daily water use)

Case 3: Single Company Assessment (50 pages)

Assumptions:

• Input documents could be 100-500 pages in total
• 50 relevant pages used (skipping boilerplate text and graphics, using retrieval-augmented generation to focus only on pages with relevant content)
• Comprehensive sustainability analysis

Calculations:

• Prompts: 50 pages x 20 prompts/page = 1,000 prompts
• Energy: 1,000 x 0.24 Wh = 240 Wh
• CO₂e: 1,000 x 0.03 g = 30 g CO₂e
• Water: 1,000 x 1.35 mL = 1,350 mL (1.35 L)

Comparisons:

• Energy: 240 Wh vs full kettle 187 Wh (≈ bit more than full kettle)
• CO₂e: 30g vs kettle CO₂e 39g (≈ less than boiling kettle)
• Water: 1,350 mL vs full kettle 1,700 mL (≈ 3/4 kettle volume)

Case 4: 100 Company Assessments

Assumptions:

• 100 assessments x 50 relevant pages each

Calculations:

• Prompts: 100 x 50 pages x 20 prompts/page = 100,000 prompts
• Energy: 100,000 x 0.24 Wh = 24 kWh
• CO₂e: 100,000 x 0.03 g = 3 kg CO₂e
• Water: 100,000 x 1.35 mL = 135 L

Comparisons:

• Energy: 24 kWh vs daily household 40 kWh (≈ 3/5 daily household)
• CO₂e: 3 kg vs driving 10 miles 2.7 kg (≈ driving 12 miles)
• Water: 135L vs daily Londoner 144L (≈ nearly day's water use)

Annual Professional AI Usage Breakdown

Realistic annual breakdown for sustainability professional:

• 10 company assessments per week (50 pages each): 520 x 30g = 15.6 kg CO₂e/year
• Quarterly bond report (100 bonds): 4 x 600g = 2.4 kg CO₂e/year
• 100 median prompts daily (general use): 365 x 100 x 0.03g = 1.1 kg CO₂e/year
• Other workflows (document summaries, analysis): ~0.9 kg CO₂e/year

Total: ~20 kg CO₂e annually

---

Training Cost Adjustments (x1.5 Factor)

Our analysis focused on inference (serving) energy. To include training costs, apply a x1.5 multiplier based on lifecycle energy analysis showing ~60% inference / ~30% training / ~10% prep & fine-tuning for deployed LLMs (Electricity Demand and Grid Impacts of AI Data Centers).

With training included the values could be adjusted by 1.5x:

• Single bond doc: 72 Wh energy, 9g CO₂e
• Single company assessment: 360 Wh energy, 45g CO₂e
• 100-bond report: 7.2 kWh energy, 900g CO₂e
• 100 company assessments: 36 kWh energy, 4.5kg CO₂e
• Annual professional AI usage: ~30kg CO₂e (vs 20kg inference-only)

There's limited information on model training vs. inference data use so we kept it at just inference given this wouldn't significantly change the results.

---

Key Assumptions & Caveats

Scope & Methodology:

• Scope alignment: Google's per-prompt CO₂e reflects electricity-generation emissions (market-based factor) plus amortized hardware emissions; water is on-site consumptive cooling only. We add indirect water from power generation using LBNL's US average 4.52 L/kWh for a fuller water picture; real values vary with grid/cooling technology.
• Conservative estimates: We use 50 tokens per prompt (deliberately conservative; median is likely lower) with 1,000 tokens per page.
• Upper bound approach: We estimate our current RAG implementation with heavy document usage to get an upper bound. Real usage is more modest and optimizations improve constantly.
• Grid comparisons: We use UK grid intensity (211 gCO₂e/kWh) for kettle and household comparisons.
• Provider variation: Different providers vary by efficiency, cooling, and grid mix.

Training Cost Details (for reference, not used in main text body):

• x1.5 multiplier rationale: Based on lifecycle energy analysis showing ~60% inference / ~30% training / ~10% prep & fine-tuning for deployed LLMs. Google's internal data shows ~60% inference / ~40% training, suggesting x1.67 as upper bound; we use x1.5 as balanced estimate.
• Application: Multiply Google per-prompt energy and CO₂e by x1.5 to include training. For water, scale indirect water from electricity generation by same factor.

Geographic Considerations:

• Water intensity variation: The 4.52 L/kWh is US national average; varies significantly by region based on grid mix and cooling technology.
• Grid carbon intensity: Comparisons use UK grid (211 gCO₂e/kWh); adjust for other regions as needed.

---

Sources & References

Core AI Impact Data:

• Google per-prompt measurements (energy, carbon, water): Measuring the environmental impact of delivering AI at Google Scale - Patterson et al., 2024
• Indirect water consumption from power generation: LBNL 2024 United States Data Center Energy Usage Report - Masanet et al., 2024
• Training cost lifecycle analysis: Electricity Demand and Grid Impacts of AI Data Centers - Kaack et al., 2024
• Google training/inference energy split: Good News About the Carbon Footprint of Machine Learning Training - Google Research Blog, 2022
• Global data center energy consumption: Energy and AI - International Energy Agency, 2024

Comparison Baselines:

• UK grid carbon intensity: Our World in Data - Carbon intensity of electricity - Ember data
• Household energy usage (TDCVs): Ofgem - Average gas and electricity usage - Ofgem, 2024
• London water consumption: Greater London Authority - Water Use - GLA Q&A
• UK car emissions factor: National Rail - Carbon comparison calculations - National Rail, 2024
• Kettle energy consumption: Heatable - Appliance Running Costs - Heatable, 2024

Token Studies & Prompt Analysis:

• Chatbot Arena conversations dataset: Hugging Face - LMSYS Chatbot Arena - 52 token average
• LMSYS Chat-1M dataset: Hugging Face - LMSYS Chat-1M - 70 token average
• WildChat dataset analysis: Emergent Mind - WildChat Dataset - 296 token average

Additional Context:

• ChatGPT usage statistics: Axios report on OpenAI usage - Axios, 2025
• Lifestyle carbon equivalents:
  • Average banana CO₂e: Carbon Cloud product report
  • Boiling kettle CO₂e: Beeco energy efficiency guide
  • Shower CO₂e: CO2 Data shower calculator
  • Email CO₂e: Carbon Literacy email impact
  • London Tube CO₂e: TfL Freedom of Information response