Predictive, generative and agentic AI: which type actually uses the most energy, and what it means for fundraising

Not all AI costs the same to run, because the type of model you choose changes its energy footprint by a wide margin.

For fundraising teams, this matters, because you want tools that work and that you can explain to a board, a donor or a colleague. Energy use is part of that story.

Here is a plain-English guide to how three kinds of AI use energy, and how Dataro keeps its footprint low.

Three kinds of AI do three different jobs

Start with what each type actually does.

Predictive models rank and score. They look at your data and sort records by likelihood, such as who is most likely to give or lapse, and they return a list rather than an essay.

Generative models create content. They write text, draft summaries or produce images, and each answer is built one word at a time.

Agentic systems chain many model calls together. They plan, call tools, check results and try again, so a single task can trigger dozens of separate steps.

Energy scales with model size and the number of calls

Two things drive the energy a task uses: how big the model is, and how many times it runs.

Predictive ranking is light, because it uses a small model once. Generative content costs more, because the model is larger and builds each answer step by step. Agentic workflows compound the cost, because they stack many model calls on top of each other.

The research backs this up. All figures below are directional industry estimates from third parties, not Dataro measurements. They vary widely with hardware, setup and how a task is defined.

A 2026 analysis in The Climate Brink found energy per AI task spans more than five orders of magnitude. Simple text classification uses thousandths of a watt-hour. An agentic workflow making 5 to 50 frontier model calls can use 50 to 500 watt-hours for a single task.

The pattern holds in peer-reviewed work too. A study in Joule put median frontier inference at about 0.31 watt-hours per query. Turn on reasoning mode and that rises about 13 times. A Hugging Face leaderboard, reported by Singularity Hub, found reasoning models use about 30 times more energy on average than the same models without it.

Agentic systems go furthest. The Green Software Foundation found agentic coding tasks can use up to 1,000 times more tokens than a normal chat. A planning loop can turn a 2,000-token request into a 2 million-token workflow. More tokens did not mean better accuracy.

A tree model can use a fraction of a large model's energy

The gap between model types is real and large.

In one text classification benchmark, a gradient-boosted tree used under 0.01 watt-hours per run, while a large generative model needed 48.6 watt-hours for similar accuracy. That is a gap of about three orders of magnitude, so for that task the tree used roughly a thousandth of the large model's energy.

The job was the same and the accuracy was the same, but the cost was very different.

Data center demand is rising, so efficiency matters more

This is not a small or shrinking problem.

The International Energy Agency estimates global data centers used about 415 terawatt-hours of electricity in 2024, and it projects this will more than double to around 945 terawatt-hours by 2030, with AI as the main driver. The US and China account for nearly 80% of that growth. These are IEA estimates, reported in 2025, not Dataro figures.

As demand climbs, the choice of model becomes more than a technical detail, because it shapes how much energy your tools pull.

Dataro uses the lightest model that solves the problem

Our approach is simple: reduce first.

We use the lightest model that does the job well, and for most fundraising tasks that means gradient-boosted trees. They rank and score donors accurately at a tiny fraction of the energy a large generative model would use.

We reserve generative AI for the places it earns its spot, such as cited donor research where it adds real value, and we do not reach for a large model when a small one gives the same answer.

We do not claim Net Zero. Efficiency leads, and offsetting is a supporting point. The honest version is this: the biggest lever is using less, and that is where we focus.

What this means for your team

A few practical takeaways:

• Match the model to the task. Ranking and scoring do not need a large generative model.

• Ask vendors what kind of model runs under the hood, and how many calls a task makes.

• Treat agentic workflows with care. The energy cost compounds fast, and more compute does not always mean better results.

• Judge efficiency by the work done per unit of energy, not by the size of the model.

Smaller models, strong results

AI energy use is not a single number, because it depends on the type of model and how often it runs. Predictive ranking is light, generative content costs more and agentic chains compound fastest.

Dataro's answer is to use the smallest model that works, lean on predictive methods and save heavier models for where they genuinely earn their place. That approach keeps results strong while keeping energy use low.

Want the full picture? Read our whitepaper on efficient AI for fundraising, or book a demo to see the approach in action.