Summarising an animal record
A short briefing drawn from the record your team already keeps. Reused while that record is unchanged.
Responsible AI
AI, with your team in charge.
Useful help. Clear limits. People responsible for the decisions.
A zoo owner should be able to ask three things, and get straight answers:
What it does
Small, defined jobs. Never the final word.
Drafting a SOAP note
A starting draft only. A clinician reviews it before anything is saved.
Finding relevant regulations
Answers draw on records from our regulatory library and show their sources so you can check them.
Before saving
A drafted note needs a clinician's review before it is saved. A suggested answer is a starting point to check against your records and sources.
Who checks it
Limits we set, and limits we admit.
We are continuing to develop evaluations using representative animal-care and compliance questions.
- Defined jobs
- Each feature has set instructions, a set of information it may access, and response limits configured in the application.
- Access controls
- Private record tools follow your facility's access controls. The public website assistant refers veterinary and legal questions to qualified professionals.
- People decide
- Veterinary, welfare, regulatory and staffing decisions need review by the appropriate person at your facility. AI can miss context or make mistakes.
- Checks, not certification
- Automated checks cover access boundaries, request limits, usage recording and safeguards for incorrect extraction. They catch software failures; they do not certify every answer as accurate or replace professional judgement.
Environmental impact
Recorded usage. Estimated impact.
Tokens are the small pieces of text a model reads and produces. Our providers return token counts, and we record them. We then use estimation tools and published research to translate usage into approximate electricity and carbon. Those figures are estimates, not readings from a meter in a provider's data centre.
What we record
- Feature, model, input and output tokens, and whether the request succeeded
- Cached and reasoning tokens, when the provider supplies them
- Missing usage stays marked missing, never counted as zero
- Images tracked separately
- AI used to build the software is accounted for separately from AI used by your facility
What we estimate, and what we cannot
Environmental figures are modelled, not metered. Model choice, hardware, input length and how providers batch requests all change the result.
Water remains unquantified. We do not know the serving data centre or its cooling system.
On request
Product AI runs when someone asks for help.
Cached briefings
Animal briefings are reused while the underlying record is unchanged.
Capped
Public chat requests and response length are limited.
No own model
Echidna does not train its own AI model.
An example to put the numbers in context
300,000 output tokens: about 4–15 hours of electricity for an efficient household fridge.
- Step 011,000 replies× ~300 output tokens each= 300,000 output tokens
- Step 020.16–0.60 Whper conventional query (study IQR) × 1,000= 0.16–0.60 kWh
- Step 03~0.041 kWh / hourefficient top-freezer fridge, 360 kWh/year ÷ 8,760 h≈ 4–15 hours
Read this before quoting the number.
This illustrates one research scenario. It is not Echidna's actual monthly usage, and it is not a fixed conversion for every 300,000 tokens.
- A scenario, not Echidna's actual monthly usage.
- Not a universal conversion for every 300,000 tokens.
- 0.16–0.60 Wh is the study's interquartile range, not a minimum and maximum.
- Input tokens are part of each query's estimate; they are not free.
- The fridge is a household reference, not veterinary or commercial equipment.
- Inference only: excludes model training, hardware manufacture and your own device.
- Same electricity does not mean same carbon; it depends on where and when power is generated.
See the calculation and its limits
Tokens. 1,000 short replies × about 300 output tokens = 300,000 output tokens.
Energy per query. Oviedo and colleagues model an interquartile range of 0.16–0.60 Wh for a conventional query with about 300 output tokens, using frontier-scale models on H100 hardware. Multiplied by 1,000 queries, that is 0.16–0.60 kWh. This is a bottom-up estimate, not metered hardware. The range covers the middle half of the study's results, not the lowest and highest possible values, and the per-query figure already includes reading the input.
Fridge reference. The ENERGY STAR 2024 reference for an efficient top-freezer refrigerator, as stated on our original page, is 360 kWh per year. 360 ÷ 8,760 hours ≈ 0.041 kWh per hour, averaged across normal cooling cycles. 0.16 ÷ 0.041 ≈ 4 hours; 0.60 ÷ 0.041 ≈ 15 hours.
How our own estimates work. Grouping requests as conventional or longer queries at 1,000 output tokens is an approximation. For carbon we use the EPA eGRID2022 US average of 373.3 gCO2/kWh as a proxy; it is not our providers' actual electricity supply. These are not full lifecycle figures.
Out of scope. Inference only. Model training, hardware manufacture and the device you use are excluded.
Is it offset?
No. And we will not pretend it is.
Through Wild Spaces, twelve selected nonprofits receive lifetime use of Echidna, including setup and support, with a value of up to $15,000 per organization per year. We report that contribution separately from our AI footprint.
Donating software does not cancel out emissions, and we do not call Echidna carbon neutral. Any future offset claim would need its own verified project records and retired credits.
About Wild Spaces