CaseloadEvaluateNotesLiteracyMySLPPricingSecurityFor districtsResources
Log inGet EASI at Northern Speech
White paper

The Energy Footprint of an AI Drafted SLP Report

How EASI estimates the energy and carbon of one AI drafted SLP report: the method, the range, what drives it and everyday comparisons, all sourced.

Lois Jean Brady, M.A., SLP, with Matthew Guggemos · November 2, 2026

Everyday items laid out on a light wooden table for scale: an electric kettle, a latte, running shoes, a water bottle, a closed laptop and a small plant.

Download the white paper (PDF)

1. Summary

EASI is a clinician controlled AI platform for speech language pathologists. It transcribes an English language sample and drafts reports that the clinician reviews, edits and signs. An estimate from our engineering team puts one EASI report, from a 15 to 20 minute language sample to the finished draft, at most likely 40 to 60 watt hours (Wh) of electricity and about 15 to 20 grams of CO2. The full range is 15 to 130 Wh and about 5 to 46 grams of CO2.

For scale, 40 to 60 Wh is about 30 to 45 minutes of streaming video, and 15 to 20 grams of CO2 is about one thirtieth of the lifecycle carbon of one café latte. This is an estimate built from our code and published energy rates, not a measurement. This paper shows how we built it, what moves it up or down, how it compares with everyday things, and where the comparisons fall short. A shorter version is in our companion blog post, How Much Energy Does One EASI Report Use?.

2. Why this question matters, and why students are asking it

People are worried about the energy and water that AI uses, and that concern is reasonable. It comes up with working clinicians, and it comes up even more with college students and graduate SLP students, who are being asked to use AI tools while also being asked to think about climate, cost and ethics.

We respect that concern. We also think the best response to it is not reassurance and not alarm, but numbers with sources. A good scientific habit is to look at scale rather than gut reactions, make assumptions explicit, and ask what the numbers actually allow us to conclude. This paper tries to follow that habit, including where it makes EASI look less tidy.

3. What EASI does when it drafts a report

To estimate the energy of "one EASI report," we first have to say what one report includes:

  • Recording. The clinician records an English language sample, typically 15 to 20 minutes.
  • Transcription. EASI transcribes the English audio into text. The clinician reviews and attests each transcript line before any scoring.
  • Drafting. With EASI report drafting (in beta), EASI works through a series of drafting steps. Each step reads the full transcript and drafts part of the report.
  • Running the app. Servers, storage and the database that hold the session and the draft.
  • Clinician review. The clinician reviews, edits and signs. EASI drafts; the clinician decides.

The key point for energy: one EASI report is not one short question to a chatbot. It is transcription plus many drafting steps, and each step reads a long transcript.

4. Method: how we built the estimate

4.1 Boundaries

The estimate covers operational electricity: the energy used while EASI transcribes, drafts and runs the app for one report. Like the published AI figures in section 6, it does not include training the AI models, manufacturing chips and servers, or building data centers. Those are real costs, and we name them so no one thinks they were quietly left out.

4.2 The method in plain terms

Our engineering team read EASI's code to count what happens in one typical report, then applied published energy rates to each part:

Energy per report = transcription + each drafting step (energy for the text it reads + energy for the text it drafts) + the app's ordinary computing (servers and database)

Carbon per report = energy per report × the carbon intensity of the electricity grid

  • Reading and drafting text. AI models use energy for the text they read and, at a higher rate per unit, for the text they draft. Published research gives ranges for both, and we applied low and high values.
  • Ordinary computing. The app itself uses a small amount compared with the AI steps.
  • Carbon. We converted energy to carbon with the US EPA eGRID2023 US average grid factor, about 0.350 kg CO2e per kWh (770.884 lb/MWh; EPA eGRID2023 summary data). The math: 40 Wh × 0.350 = about 14 g and 60 Wh × 0.350 = about 21 g, rounded to about 15 to 20 g. For the full range, 15 Wh × 0.350 = about 5 g and 130 Wh × 0.350 = about 46 g. California's grid is cleaner, about 0.195 kg CO2e per kWh (CAMX subregion, 429.983 lb/MWh, same EPA data), so the same report there would come to less, about 8 to 12 g at the most likely estimate.

4.3 What this estimate is and is not

It is an estimate from our engineering team, built from code and published rates. It is not a meter reading. We now log usage per feature, and one set of real runs should let us narrow the range considerably. We will publish the update when we have it.

5. Results: the range, and what drives it

Scenario Electricity per report Carbon per report
Low end about 15 Wh about 5 g CO2
Most likely 40 to 60 Wh about 15 to 20 g CO2
High end about 130 Wh about 46 g CO2

All figures are for one report from a 15 to 20 minute sample to the finished draft. Estimate from our engineering team.

Why the range is wide. The high end is close to nine times the low end. Five things drive it:

  • Length of the language sample. A longer recording makes a longer transcript, and every drafting step reads the whole transcript. Length matters more than once.
  • How much the report drafts. More sections, more detail, or additional documents mean more drafting, and drafting text costs more energy per unit than reading it.
  • Redrafting. Asking EASI to draft a section again repeats that step's energy.
  • Published energy rates vary. Independent estimates for AI energy per unit of text differ widely, so we carried a low and a high value instead of picking one.
  • The grid. The same electricity produces less carbon on a cleaner grid and more on a grid that burns more coal.

We revised our own earlier number. An earlier estimate on our website was much lower. When our team rebuilt it more carefully, the old figure proved too low, so we replaced it with this range.

6. How one report compares with one AI request

Item Energy Carbon Source
One median short AI text prompt 0.24 Wh 0.03 g CO2e Google Cloud, Aug 2025
One typical chatbot query about 0.3 Wh not given Epoch AI, Feb 2025
One AI query with a long input about 2.5 Wh not given Epoch AI, Feb 2025
One AI query with a very long input about 40 Wh not given Epoch AI, Feb 2025
One EASI report (most likely) 40 to 60 Wh about 15 to 20 g CO2 Estimate from our engineering team

At its most likely value, one EASI report uses about as much energy as 170 to 250 median short AI text prompts. That is the honest consequence of reading long transcripts many times. When you see a headline about "one AI question," remember that the size of the input changes the answer a great deal.

7. Everyday context

These comparisons are context, not judgment. Every item below is part of normal life.

7.1 Energy (electricity in use)

Everyday item Energy One EASI report (most likely 40 to 60 Wh) equals Source
One hour of streaming video about 77 Wh about 30 to 45 minutes of streaming IEA, Dec 2020
Boiling 1 liter in an electric kettle 108 to 115 Wh about a third to a half of one liter boil Durand et al., 2022

At the high end (130 Wh), a report is about an hour and forty minutes of streaming, or a little more than one liter boiled.

7.2 Carbon

Everyday item Carbon One EASI report (most likely 15 to 20 g) equals Scope Source
One café latte about 550 g CO2e about one thirtieth of a latte Whole life Nab and Maslin, 2020
Driving 1 mile, average US car about 400 g CO2 driving about 200 to 260 feet Tailpipe only US EPA, 2023
One pair of running shoes about 14 kg CO2e 700 or more reports per pair Whole life Cheah et al., 2013
One 500 mL plastic water bottle, bottle only 72 to 109 g CO2e about one seventh to one quarter of the bottle Partial life (before filling and transport) US NIH NEMS

At the high end (about 46 g), a report is still about one twelfth of a latte.

7.3 Where these comparisons fall short

  • Whole life versus electricity in use. The latte, the shoes and the bottle are counted from farm or factory onward. The EASI estimate, the AI figures, streaming and the kettle are mostly electricity at the moment of use, and they leave out building chips and data centers. So the carbon comparisons give a sense of scale, not an exact match.
  • Different kinds of energy. Bottle production is mostly fuel energy, not electricity, which is why we compare the bottle by carbon only.
  • Grids vary. Section 4.2 shows how much the grid factor alone can change the carbon figure.
  • Ages of sources. The streaming figure is from 2020; the shoe study is from 2013. Newer studies may differ.
  • Company reported figures. The median prompt figures from the major provider are measured by that provider and, by its own note, not checked by an independent third party. The independent research group's figures are estimates.

8. How AI energy per request has changed

AI energy per request is a moving target, and mostly a falling one.

  • Big efficiency gains are being reported. One major provider reported that over a recent 12 month period, the energy used by its median AI text prompt fell by a factor of 33, and its total carbon fell by a factor of 44 (Google Cloud, Aug 2025). These are the company's own figures.
  • Short requests are now in the same range as old web searches. A 2009 estimate put one search at about 0.3 Wh. Recent estimates for one short AI text request are 0.24 to 0.3 Wh. These measure different things in different years, so treat this as a rough marker, not a match.
  • But longer, richer tasks use more. The same independent analysis shows long inputs using roughly eight to over a hundred times the energy of a short query. Tools like EASI that read whole transcripts sit on that side of the curve.
  • Our own estimate changed too. As section 5 explains, we raised our number when we looked more closely. Estimates should change when better information arrives.

9. What clinicians and students can do

When using EASI:

  • Record the sample length your assessment actually calls for. Longer samples use more energy at every drafting step.
  • Draft the sections and documents you need, and edit a draft directly instead of asking EASI to redraft the whole thing.
  • Remember the clinician is in charge: you decide what is drafted and what is kept.

When evaluating any AI energy claim (ours included):

  • Ask what is being counted: one short question, or a whole task with long inputs?
  • Ask about the boundary: electricity in use only, or the whole life of the hardware?
  • Ask which grid, and which year.
  • Ask whether it is a measurement or an estimate, and who checked it.
  • Compare at scale, with sources, and notice when a comparison mixes whole life numbers with electricity in use.

For students, these questions make a good class exercise: pick any AI energy headline and run it through the list.

10. What we will do next

  • Narrow the range with real usage data, and publish the update here.
  • Ask our providers for per request energy and carbon data we can cite.
  • Revisit the comparisons as newer studies appear.

11. Contact

Questions or corrections are welcome. Learn more about EASI at NorthernSpeech.com/EASI.

12. References

Download the white paper (PDF)

Your caseload shouldn't take over your life.

EASI connects evaluations, reports, progress notes, planning and scheduling, with the SLP in control.

Get EASI at Northern Speech More resources
The Energy Footprint of an AI Drafted SLP Report | EASI