Skip to content

Guide

What an explainable AI workout recommendation shows the coach

A practical checklist for any AI tool that changes a client's training. If a recommendation cannot answer these four questions, check it by hand before you approve it.

By RaiNGE product teamUpdated 5 min read

Written for coaches reviewing AI-suggested workout changes before they reach a client.

Short answer

An explainable AI workout recommendation shows the coach four things: the client inputs it used, the rule or reason behind the change, exactly what changed from the original, and how to reverse it. If any of the four is missing, the coach cannot audit the recommendation and should check it by hand before approving.

  • Inputs shown: which client signals the recommendation read, and how recent they were.
  • Reason named: a specific trigger, such as a low-back flag or pain at 4/10, beats a phrase like optimized for recovery.
  • Change visible and reversible: the before and after values, and one step to undo.

Compare a standalone AI workout generator with programming software that keeps the coach's review in the loop. Compare generator types

On this pageThe four questions every recommendation should answer

The four questions every recommendation should answer

Run any AI suggestion through these four questions before approving it. Each strong answer below is specific enough to check against the client's record in a few seconds.

What a strong and a weak answer look like for each question.
QuestionStrong answerWeak answer
What inputs did it use?Right knee pain 5/10, logged this morning. Last five session RPEs: 8, 8, 9, 9, 8.Based on your client's data
Which rule or reason fired?Pain at 4/10 or more triggers a review of today's lower-body workOptimized for recovery
What changed?Split squat 4×8 at RPE 8 with 90 s rest became 3×8 at RPE 6 with 120 s restVolume adjusted
How do I reverse it?One revert restores the original sets, reps, rest, and intensity notesEdit the workout by hand
What a strong and a weak answer look like for each question.

This page is the working checklist. For the argument behind it, read the essay on explainable AI in fitness coaching.

Strong and weak explanations, side by side

A weak explanation restates the recommendation in nicer words. A strong one names the signal, the rule, and the trade-off, so the coach can disagree with a specific part of it.

The same four recommendations, explained badly and well.
RecommendationWeak explanationStrong explanation
Reduce lower-body volumeOptimized for recovery.The last five session RPEs were 8 to 9 against a target of 7, so squats and lunges drop from 4 working sets to 3 this week.
Swap the RDL for a hamstring curlReduces injury risk.The profile has a low-back flag, which keeps loaded hinges out. The curl trains the back of the thigh with no spinal load. Hinge practice is lost.
Repeat last week's loadProgression not recommended.The client missed one of three sessions and rated the last one RPE 9, so there is no evidence yet that the load is easy.
Hold the session for reviewManual approval required.Check-in pain was 8/10. Above 7/10 the client needs coach review, and professional evaluation is recommended before loading continues.
The same four recommendations, explained badly and well.

Worked example: auditing one adjustment in under a minute

A proposed adjustment arrives for a client's lower-body day. Here is the audit, in order.

  1. 1

    Check the inputs

    Right knee pain 5/10, logged 10 hours ago. Last five session RPEs 8, 8, 9, 9, 8. Both are recent enough to matter today.

  2. 2

    Find the trigger

    Pain at 4/10 or more. The proposal follows from the pain, so the reason holds up.

  3. 3

    Compare before and after

    Split squat 4×8 at RPE 8 with 90 s rest becomes 3×8 at RPE 6 with 120 s rest. Step-ups 3×10 are unchanged.

  4. 4

    Add what the system cannot see

    You know the client ran a 10 km race yesterday, which explains the high RPEs. The cut fits, and the knee is worth a question at warm-up.

  5. 5

    Decide, and know the undo

    Approve it. If you later learn the pain was logged in error, one revert restores the original. If the knee passes 5/10 during warm-up, end lower-body loading for the day.

Signs an explanation should not be trusted

Treat the recommendation as unexplained if

  • It cites a signal you cannot find in the client's record.
  • It sounds more certain than the data allows, such as promising to prevent injury.
  • It claims to know the cause of pain or how the client's technique looks.
  • Its stated goal does not match the purpose of the session.
  • It keeps no record of the original, so the change cannot be undone.

What RaiNGE does today, stated narrowly

Measured against the checklist, RaiNGE covers some items fully and some only in part. This table says which.

RaiNGE against the explainability checklist.
Checklist itemWhat RaiNGE doesWhat it does not do
InputsDrafts start from goals, training history, equipment, schedule, injury and pain notes, and readiness check-ins. The generator reads the latest progress check-in pain.Trace each exercise back to the input that chose it
Rule namedDeterministic safety rules remove exercises that conflict with flagged limitations, and cautions are shown to the coachExplain model choices that no rule governs
Per-exercise reasonEach exercise carries a short coaching note, a cue of up to about 15 wordsWrite a full rationale for every exercise
Pain-driven changePain at 4/10 or more creates a pending adjustment. The check reads pain from about the last 36 hours and the last five RPE scores.Change the plan automatically or route clients to clinicians
ReversibleApproved changes to sets, reps, duration, rest, and intensity notes apply as one change that can be reverted, and the athlete is notified after it succeedsLearn from your edits: coach changes do not retrain the system
Second lookFull-program drafts get a second AI review pass before the coach sees them, always when the client has injury notesAct as an independent check: it uses the same model and may be skipped for clean, high-scoring drafts
RaiNGE against the explainability checklist.

The coach still reviews, edits, and approves every program before it reaches the client. The case for keeping that step is made in human-in-the-loop AI coaching.

Common questions

Does an explained recommendation mean the AI is right?

No. An explanation shows how the recommendation was reached, which lets you spot a stale input or a rule that does not fit this client. A well-explained recommendation can still be wrong. Explainability makes errors cheaper to catch; it does not remove them.

How long should an AI explanation be?

Short enough to read in the minute you have per client, and specific enough to check. One line naming the input, one naming the rule, and a before-and-after view of the change is usually enough. Long generated paragraphs tend to bury the one fact you need.

Does RaiNGE learn from the changes I make to its recommendations?

No. Coach edits are not used to retrain or teach the system, and its explanations do not change based on feedback. What changes the next draft is new client context: updated injury notes, readiness check-ins, and the latest pain score. Keep your own note if you want a record of why you overrode a suggestion.

Should clients see the explanation too?

A short version, yes. Tell the client what changed (load, range, sets, or the exercise), why in one sentence, and which symptom would stop the session. Keep diagnosis out of it. The coach-facing detail, such as RPE trends and rule names, can stay with the coach.

Next step

Compare a standalone AI workout generator with programming software that keeps the coach's review in the loop.

Compare generator types

More on workout safety

The full workout safety guideBrowse all articles

Related elsewhere