Marketers have become very good at measuring what AI saves
Hours saved. Content produced. Campaign variants created. Research accelerated. Cost per asset reduced.
We are much less accustomed to measuring what AI consumes.
Every prompt, generated image, automated report, personalised message and AI-assisted decision requires computing power. That means electricity, data-centre infrastructure, cooling, networking and hardware.
At the scale of one prompt, the impact may be tiny. At the scale of a modern marketing operation, it becomes a different conversation.
The International Energy Agency estimates that data centres consumed around 415 TWh of electricity globally in 2024 and expects this to rise to approximately 945 TWh by 2030, with AI contributing significantly to that growth.
So perhaps it is time to add another metric to our AI dashboards.
Stop measuring AI by prompts
A prompt is a poor unit of environmental measurement.
One short request to summarise a paragraph bears little resemblance to generating 30 high-resolution images, analysing a large customer dataset or running an automated personalisation workflow thousands of times.
The footprint depends on factors including the model being used, the length of the input and output, how many times the task is repeated, the type of content being generated and the infrastructure running underneath it.
For marketers, the useful measurement is often closer to the finished business output.
That could mean emissions per:
- approved campaign asset
- 1,000 personalised messages
- reporting run
- completed customer interaction
- incremental conversion
This immediately makes sustainability part of marketing effectiveness rather than a separate ESG exercise.
Bigger AI is not automatically better AI
Marketing teams are rapidly gaining access to increasingly capable models, and it is tempting to use the most powerful option for everything.
Many tasks simply do not need it.
Classification, extraction, basic summarisation and simple data processing may be perfectly suited to smaller models. More computationally demanding systems can then be reserved for complex reasoning, strategic analysis and genuinely difficult creative work.
There is a useful commercial benefit here too.
Efficient AI use can reduce both environmental impact and operating cost.
Watch the creative department
Generative creativity can become particularly resource-intensive.
Producing a single useful visual is one thing. Generating 80 alternatives because producing another version takes only a few seconds creates a very different workload.
The same principle applies to copy.
AI makes variation almost frictionless, which makes it remarkably easy to create content that nobody really needed.
Creative teams can introduce some simple discipline:
Define the concept before generating large numbers of assets. Use lower-resolution imagery during exploration. Limit unnecessary retries. Agree approval criteria before production begins. Avoid automatically creating dozens of audience variations unless the performance benefit justifies them.
Less wasted generation also means less wasted marketing effort.
Personalisation needs a business case
AI-driven personalisation is one of marketing’s most promising opportunities, although scale changes its economics.
Imagine generating personalised content for 500 customers.
Now imagine 500,000.
If the additional computation produces meaningful improvements in conversion, retention or customer experience, the investment may make sense.
If a complex model creates a fractionally better prediction or an almost indistinguishable piece of copy, the additional processing may add little value.
This gives Marketing Managers another question to include when assessing AI workflows:
What incremental marketing value are we getting from the additional compute?
Ask suppliers better questions
Much of marketing’s AI footprint sits outside the organisation.
Your team may know how many images it generated or how many API calls it made without knowing which processors were used, where the data centre was located or what electricity mix powered it.
Supplier conversations therefore need to become more specific.
Ask AI platforms, agencies and technology providers about product-level energy estimates, model efficiency, hosting regions, cooling and water use, hardware emissions and the methodology behind any sustainability claims.
A company-wide sustainability report tells you relatively little about the footprint of the AI product your marketing team is actually using.
Practical changes marketers can make now
- Inventory your AI use – Map where AI is being used across content, research, analytics, CRM, advertising, customer service and automation.
- Measure outputs, not just prompts – Choose functional measures that relate AI activity to useful marketing outcomes.
- Match the model to the job – Use appropriately sized models for straightforward tasks and heavier models where their capabilities genuinely add value.
- Reduce unnecessary generation – Shorter prompts, bounded outputs, fewer retries and clearer creative briefs can all reduce wasted processing.
- Put controls around image and video generation – High-resolution visual generation can require substantially more computation than simple text generation.
- Add sustainability to procurement – Include environmental evidence alongside privacy, security, accessibility, performance and cost when evaluating AI suppliers.
- Connect carbon with performance – Track environmental impact alongside the commercial value produced by the workflow.
One number should come with a warning
Google estimated that a median Gemini Apps text prompt in 2025 consumed around 0.24 watt-hours of electricity, produced approximately 0.03 grams of CO2e and used around 0.26 millilitres of water.
It is an interesting benchmark.
It is not a universal conversion rate.
Different models, workloads, locations and infrastructure can produce very different results. Image generation, video and complex agentic workflows may require substantially more computation.
Marketers should therefore be wary of simplistic claims such as “one AI prompt equals X grams of carbon”.
The useful numbers will increasingly be the ones connected to our own workflows.
AI efficiency now needs a wider definition
AI can absolutely make marketing more efficient.
It can reduce travel, physical production, repetitive administration and unnecessary manual processing. It can also encourage us to generate thousands of things simply because we can.
The opportunity for marketers is to become more selective.
Use the right model. Generate what has a purpose. Measure the workflow. Challenge unnecessary scale. Ask suppliers for evidence. Connect environmental cost with marketing value.
We already measure whether AI makes marketing faster.
Is the next stage understanding whether it also makes marketing genuinely more efficient?
