How we use AI

How OneCoach uses AI

Last updated 1 September 2026

Artificial intelligence sits in the core of OneCoach. This page is the product disclosure for that system: what it does, what data it is shown, what it is forbidden to be, and how it fails. It is written to be read with the Terms of Service and the Privacy Policy. The Terms control liability. The programming ideas behind a block are described in Methodology and the science behind OneCoach. Those pages are education, not a warranty that every output follows every citation.

1. What the AI is

OneCoach AI is software. It is a stack of large language models, speech-to-text models, validation rules we write, and product logic that assembles your profile, plan, and logs into a prompt. It is not a human coach, not a clinician, and not a separate legal person. When the app says “coach”, it means this system.

We use contracted providers to run inference and transcription. They process a request as our processors, then return text. We do not name every vendor here because vendors can change. The legal relationship and transfer safeguards are in the Privacy Policy.

2. Plan generation

After onboarding, the system reads the answers you gave: goals, experience, injuries, equipment, priority lifts, session length, training days, and any imported notes. It then generates a 12-week programme. We constrain generation with programming rules (periodisation, overload, injury exclusions, equipment filters, volume caps). Those constraints are our attempt to stop obvious nonsense. They are not a guarantee.

The model can still include an exercise you should not do, a load that is too heavy, or a split that does not match a gym you actually walk into.
“Validated” means our software ran checks. It does not mean a chartered S&C coach signed the PDF.
You can rewrite the plan in chat, swap lifts, change days, or abandon it. You should, if it does not match your body that week.
A later rewrite can invalidate an offline copy sitting on your phone. Check the week before you train.

3. Coaching chat

In-session and out-of-session chat can see relevant parts of your plan, workout history, personal bests, nutrition logs, and health metrics you have connected. It can propose exercise swaps, load changes, meal logs, and answers to training questions. It reasons over text. It cannot see the bar, your brace, or the person next to you.

Replies can be wrong, stale, or internally inconsistent with last week’s plan.
The coach may agree with an unsafe request (“add 20 kg”, “train through this sharp pain”). You must refuse those requests yourself.
The coach may invent a citation, a muscle, or a percentage. Do not treat tone as evidence.
If a reply conflicts with a clinician, the clinician wins.

4. Voice

Voice is streamed for transcription, cleaned, then sent to the coaching model as text. We do not store the audio file. We do store the transcript if it becomes a message or a log.

Transcription mis-hears numbers, “left” versus “right”, injury words, and food names. Read the text before you load a bar or eat to a target.
Gym noise, other people’s voices, and wind on a walk will degrade accuracy.
Do not dictate another person’s health data or anything you are not allowed to record in that room.
Microphone permission is optional. If you refuse it, type instead.

5. Nutrition and shopping lists

Natural-language meal logging asks a model to estimate energy and macros from a description. That is an estimate, not a weighed analysis.
Barcode scan matches a code to a product database. Databases lag real recipes. Store-brand items are often wrong.
Shopping lists and UK price hints are assembled from sources available at run time. They are not an offer to sell groceries and not allergen labelling.
Do not use these tools as medical nutrition therapy, insulin dosing, or eating-disorder treatment.

6. Imported plans

If you paste or upload an existing programme, a model tries to structure it into the app. Parsing can drop accessories, misread intensity, or assign the wrong laterality. You warrant you have the right to upload that text. We do not adopt the imported author’s coaching duty.

7. Health and readiness

If you connect Apple Health or similar sources, models and rules may use sleep, heart rate, or related metrics to colour a readiness view or to suggest backing off. Those scores are wellness labels. They are not a diagnosis of overtraining, illness, anaemia, or cardiac risk. A green score does not clear you to ignore chest pain, dizziness, or a clinician.

8. Offline use

A copy of your current plan can be used without a signal. Logging may queue until you reconnect. Anything that needs a model (chat, a rewrite, a meal estimate, a new barcode match) needs a network. If the queue fails, you may need to re-enter a log. We are not responsible for a missed session because the phone was offline.

9. What data the model sees

For a given request we assemble the smallest context we reasonably can: the user message, and slices of profile, plan, logs, nutrition, and health data that the product thinks are relevant. That packet goes to the inference or transcription provider. Community posts you did not include in the request are not automatically dumped into every prompt. Your password and your card number are never sent to a model. Apple holds checkout.

Full categories, retention, and your rights are in the Privacy Policy.

10. We do not train models on you

We do not use your training logs, chats, voice transcripts, health metrics, or community posts to train a foundation model.
We instruct inference and transcription providers not to use OneCoach customer content to train their models.
We may compute aggregated product metrics (for example, how often a rewrite is requested) that cannot reasonably identify you.
A human may read a conversation if you open a support ticket about it, if we investigate abuse or a safety event, or if the law requires it. That is support and safety, not model training.
If we ever run a research project that needs identifiable training data, we will ask for a separate, specific consent. The science inbox is not a back door into your account.

11. Hard limits and known failure modes

We try to stop the system from doing certain things. You must assume the stop can fail.

We design it not to

Diagnose disease or prescribe medicine.
Ignore an injury you clearly typed, on purpose.
Programme equipment you said you do not have, on purpose.
Pretend to be a named human on the OneCoach team.

It still might

Miss a buried injury in a long message, or a voice typo of that injury.
Hallucinate a load, a tempo, a supplement, or a study.
Encourage volume that is too high for a bad week, a beginner, or a return from time off.
Give form cues that are wrong for your structure or your sport.
Produce a meal estimate off by a large margin.
Conflict with itself across two messages.
Be unavailable when a provider is down, including mid-set.

12. What the AI must not be used for

Emergency care. Call 999 or your local emergency number.
Clinical diagnosis, imaging interpretation, or medication changes.
Physiotherapy or post-operative rehab unless your clinician has approved the exact plan.
Eating-disorder treatment, crash dieting, or insulin dosing.
Coaching a minor, a client, or a team as if OneCoach were your insured professional practice.
Doping, banned substances, or instructions to harm yourself or someone else.
Legal, employment, or insurance evidence that you “were under a coach”.

13. Humans, support, and safety review

A person reads support email. That is not the same as a human rewriting every programme. We do not watch each session. If the system produces something that looks like a safety event, we may review the thread, restrict features, or terminate an account as the Terms allow.

Priority support on Pro means we aim to reply faster. It is not a clinician on call and not a guaranteed SLA.

14. Transparency

You are talking to an artificial system. We will not brand it as a named human sitting in a call centre. If a real person emails you from hello@getonecoach.app, that email will come from us as a company, not from “your AI”.

15. Changes to models

We can change providers, models, prompts, and rules without notice. The coach you used in January may not answer like the coach in June. This page will be updated when the way we use AI materially changes. Questions: hello@getonecoach.app.