Racing & Strategy ·

Training Plans, Apps and AI Coaches: What Adaptive Plans Can and Can't Replace

Static plans, adaptive apps and AI chatbots can handle load arithmetic and structure, but expert-rated AI plans come out mediocre without detailed input. What the thin evidence shows, and where a human still matters.

Training Plans, Apps and AI Coaches: What Adaptive Plans Can and Can't Replace

Sprint Summary

The short version — read this if you're short on time.

Static plans, adaptive apps and AI chatbots are now how many age-group triathletes train. They are good at the arithmetic of training: spreading load across the week, building gradually, scheduling recovery and keeping you consistent. But the evidence on how well they work is thin. When coaching experts rated ChatGPT-generated six-week running plans, they judged them sub-optimal, though quality improved as more detail about the runner was supplied. Expert ratings of AI-generated strength plans came out moderate, and in one evaluation some plans for fictional patients with type 2 diabetes had serious safety issues.

Adaptive training has a little more evidence. A meta-analysis of 8 studies with 198 participants found that adjusting endurance training day to day using heart rate variability improved submaximal fitness markers compared with a fixed plan, but not performance. Where a human still matters is judgement with context: injury and illness, a complicated week, a race goal that needs rethinking, and the relationship that makes you report honestly. No trial has compared app-coached, AI-coached and human-coached triathletes, so the right choice depends on your experience, injury history and how complicated your life is.

Safety & Context

An app or AI chatbot is not a clinician. Don't use one to decide whether to train through pain, return from injury, or exercise with a heart, lung, metabolic or other medical condition. In one evaluation, some ChatGPT-generated plans for fictional patients with type 2 diabetes had serious safety issues. Ask your doctor or physiotherapist.

Readiness or recovery scores from apps and wearables are estimates, not diagnoses. Chest pain, palpitations, unusual breathlessness, fainting or dizziness need medical attention whatever your score says.

If an AI tool suggests calorie, weight or body-composition targets, or very low fuelling, don't follow it without a registered dietitian. TriForward's "Under-Fuelled, Not Undertrained" explains why under-fuelling is a health risk.

Evidence here is thin and fast-moving: expert ratings of generated plans, fictional profiles, general adults and lab cyclists, not age-group triathletes training with these tools. Treat any claim that an app or AI coach is as good as a human as untested.

Full Distance

The complete research and analysis.

Ten years ago, the choice was simple: buy a plan from a book or a website, or hire a coach. Now there's a spectrum. Static plans still exist. Training platforms adjust your sessions based on completed workouts and test results. Wearables give you a daily readiness score. And large language models such as ChatGPT will write a 20-week Ironman plan in seconds if you ask.

The marketing for some of these tools implies that a human coach is now optional. That might be true for some athletes. But it's a claim the research hasn't tested, and the studies that do exist are mostly experts rating plans on paper, not athletes training with them. This article covers what the evidence says about AI-generated plans, adaptive training and app-based programmes, what these tools are genuinely good at, and where a human with context still adds something.

Four kinds of plan

It helps to separate four types of tool, because they work in different ways:

  • Static plans: a fixed 12- or 20-week programme. It doesn't change unless you change it.
  • Rule-based adaptive apps: platforms that adjust upcoming sessions based on what you completed, your test results or your ratings of how sessions felt.
  • Data-driven adaptive tools: platforms or wearables that adjust training using physiological data such as heart rate variability (HRV), sleep or resting heart rate.
  • Large language model (LLM) coaches: chatbots that write plans and answer questions in natural language, either general tools or ones built into training platforms.

Commercial adaptive algorithms are mostly proprietary and haven't been independently tested in published trials. So most of the direct evidence below is about LLM-generated plans and HRV-guided training, with app research from general physical activity.

What the research on AI-generated plans shows

Running plans: sub-optimal, better with more detail

The most relevant study for triathletes is a 2024 paper by Düking and colleagues in the Journal of Sports Science and Medicine. They asked ChatGPT to produce three six-week running plans, each with more information about the runner than the last. Coaching experts then rated each plan against 22 quality criteria drawn from the literature, on a 1 to 5 scale.

  • The plans were rated sub-optimal overall.
  • More input produced better plans. A median rating of 3 was given 0, 6 and 13 times for plans 1, 2 and 3 respectively.
  • The plan with the least input was rated significantly lower than the most detailed plan on 15 criteria.
  • The authors advised against using ChatGPT-generated plans without feedback from an expert coach.

The abstract doesn't give the number of raters. These were fictional profiles rated on paper. Nobody ran the plans.

Strength plans: moderate quality

A 2025 study by Havers and colleagues in Biology of Sport had 10 experienced coaches, each with at least a bachelor's degree in exercise science and two years' coaching, rate hypertrophy and maximal-strength plans from three LLM tools against 27 criteria.

  • Overall quality was moderate.
  • Differences between tools were small, significant on only 1 or 2 of the 27 criteria.
  • One tool was the most inconsistent in its ratings.
  • The authors concluded that the plans can serve as an initial framework, but that expert supervision remains crucial, because LLMs can't account for individual responses, safety or the adaptation processes an experienced coach observes.

Detail helps, but the output still varies

A 2026 study by Yang and colleagues, also in Biology of Sport, tested consistency. They entered a simple and a detailed prompt into GPT-4 four times each, producing 8 programmes, which 11 experts rated blind.

  • Detailed prompts scored higher for personalisation, safety, feasibility and overall quality.
  • Detailed prompts gave more stable outputs.
  • But the same prompt, entered repeatedly, still produced structurally and qualitatively different programmes.

Inference: if you use an LLM to plan training, the plan you get depends on how much you tell it, and on chance. Asking twice may give you two different plans.

Safety: where it can go wrong

In a 2025 discussion paper in Sports, Akrimi and colleagues asked ChatGPT-4o for exercise plans for three fictional patients with type 2 diabetes, using three prompts each. Three coaching experts reviewed them.

  • Some plans had serious safety issues, especially for patients with secondary diseases or complications.
  • Most included key training principles but had deficits such as poor feasibility.
  • More detailed prompts tended to give better plans.

This was preliminary, with fictional patients and three reviewers. But it's a warning for any athlete with a medical condition, injury history or medication that affects exercise.

Answering questions: often better than expected

Not all the findings are negative. In a 2026 study in the Journal of Sports Science and Medicine, D'hoe and colleagues collected the most frequently asked questions from 9 qualified personal trainers, along with their answers, and put the same questions to ChatGPT 3.5. 18 personal trainers and 9 topic experts rated the answers.

  • ChatGPT outperformed the trainers on 6 of 9 questions overall.
  • It scored higher for scientific correctness on 5, comprehensibility on 6 and actionability on 5.
  • None of the trainers' answers scored higher than ChatGPT's on any question or measure.

Inference: LLMs may be good at explaining general training principles, the kind of question you'd find answered in a textbook. That's different from writing a plan for one person, which depends on individual information the tool may not have, or may get wrong.

What the evidence on adaptive training shows

The closest research to adaptive apps is on HRV-guided training, where sessions are adjusted day to day based on morning heart rate variability. A 2021 systematic review and meta-analysis by Düking and colleagues in the Journal of Science and Medicine in Sport pooled 8 studies with 198 participants and 9 interventions.

  • HRV-guided plans usually included fewer moderate and high-intensity sessions than the predefined plans.
  • Submaximal physiological markers improved more with HRV-guided training (Hedges' g = 0.296, statistically significant).
  • Performance didn't improve significantly more (g = 0.079), nor did VO2peak (g = 0.171).
  • There were fewer non-responders for performance with HRV-guided training.

Inference: adapting training to daily data may help you get similar results with fewer hard sessions, and may reduce the number of people who don't improve. It isn't a proven performance boost. TriForward's "Reading the Trend, Not the Number" covers how to use HRV as a trend rather than a daily verdict.

Commercial adaptive platforms use their own rules, often combining completed load, test results and ratings. None of these, as far as published research shows, has been tested in trials against human coaching or static plans in triathletes.

What the evidence on apps and supervision shows

Two broader bodies of research give some context, though neither is about trained athletes.

Apps and staying active

A 2019 systematic review and meta-analysis by Romeo and colleagues in the Journal of Medical Internet Research included 9 randomised trials with 1,740 adults, testing smartphone apps as the main physical activity intervention.

  • Apps produced a non-significant increase of 476.75 steps a day compared with control conditions, from 6 studies.
  • Apps were more effective over less than 3 months than over longer periods.
  • Apps targeting activity alone worked better than those also targeting diet.

These were general adults, not athletes. Inference: the novelty of an app may help in the short term, but on its own it isn't a reliable long-term motivator. Most triathletes already have the motivation; what they need from a plan is structure and good decisions.

Supervised vs unsupervised training

A 2017 meta-analysis by Lacroix and colleagues in Sports Medicine compared supervised and unsupervised balance and strength programmes in 11 studies of healthy adults aged 65 and over.

  • Supervised training improved muscle strength and power more than unsupervised training (SMD 0.51).
  • It also improved several balance measures more.
  • The advantage was larger compared with fully unsupervised programmes than with programmes that included a few supervised sessions.
  • The authors cautioned that there were few studies and their quality was moderate.

This is a very different population and type of training. Inference: some supervision, even a small amount, seems to add something beyond the programme itself. For a triathlete, that might mean an occasional session with a coach or a swim technique review rather than full-time coaching.

What plans and apps do well

Drawing these together, these are reasonable strengths of good plans and apps. They are inference, not tested findings:

  • Load arithmetic. Gradual progression, recovery weeks and a sensible spread of intensity across swimming, cycling and running.
  • Consistency. A session waiting in your calendar removes a decision.
  • Data handling. Tracking load over weeks better than memory can.
  • Explaining principles. As the question-answering study suggests, LLMs can explain general concepts clearly.
  • Cost and availability. Much cheaper than one-to-one coaching, and available at 10pm.

Where a human still matters

And these are areas where a human with context is likely to add something:

  • Injury and illness. Deciding whether a niggle needs a lighter week, a physio or a complete break. No app can examine you.
  • Context the data can't see. A new job, a sick child, a bereavement, a hot week, a bike that doesn't fit.
  • Judgement about goals. Whether a race target is realistic, or whether to drop to a shorter distance this year.
  • Technique. Swim technique in particular benefits from someone watching. TriForward's "Filming Your Stroke: What Swim Video Analysis Can and Can't Fix" covers what video can and can't do.
  • The relationship. Accountability and honest reporting, covered in TriForward's "What to Expect From a Coach, and What a Coach Should Expect From You".

Note that human coaches don't always use data well either. In a 2024 pilot study covered in TriForward's "When Your Coach's 'Easy' Isn't Your Easy: Intended vs Perceived Training Load", age-group triathlon coaches rarely acted on weekly load reports. The comparison isn't perfect app against perfect coach.

Using these tools well

  1. Give them detail. Every LLM study found better plans with more information. Include your training history, available hours, injuries, test results and race.
  2. Check the plan against basics. Does it build gradually? Are there easy weeks? Is most of the training easy? Does it fit your real week?
  3. Don't regenerate until you like it. If the same prompt produces different plans, the one you prefer isn't necessarily the best.
  4. Feed back honestly. Adaptive apps can only adapt to what you log. Rate sessions honestly.
  5. Keep medical decisions human. Pain, illness and conditions go to a clinician.
  6. Buy targeted help. A one-off review, a technique session or a few months of coaching before an A-race may add more than full-time coaching or none at all.

A triathlon scenario

The following example is illustrative, not drawn from a study. A 46-year-old new triathlete with a full-time job asks a chatbot for a 16-week plan for her first standard-distance race. She gives only her race date. The plan has six sessions a week including two hard runs, and she can only manage four.

A better approach:

  • She re-prompts with detail: four sessions a week, 6 hours maximum, a previous calf strain, a weak swim, and a 25-metre pool.
  • She compares the result with a structured app plan and picks the one that fits her week.
  • She logs every session with a 0 to 10 effort rating.
  • When her calf tightens in week 6, she sees a physio rather than asking the chatbot whether to run through it.
  • She books two swim technique sessions with a club coach.

Common mistakes

  • Treating a generated plan as a prescription rather than a draft.
  • Giving the tool almost no information and expecting a personal plan.
  • Asking an app or chatbot about pain, illness or medication.
  • Following a readiness score over how you feel, or ignoring how you feel because the score is high.
  • Changing apps every few weeks, losing the consistency that's their main strength.
  • Assuming a human coach is always better, or always unnecessary.

What we don't know

  • Outcomes in triathletes. No trial has compared app, AI and human coaching in triathlon.
  • Commercial algorithms. Most adaptive platforms haven't published independent validation.
  • Real-world training. The LLM studies rated plans on paper. Nobody followed them.
  • Speed of change. Tools improve quickly; studies of older versions may not reflect today's.
  • Injury. Whether app- or AI-planned training changes injury risk is unknown.

References

  • Düking P, Sperlich B, Voigt L, Van Hooren B, Zanini M, Zinner C. ChatGPT generated training plans for runners are not rated optimal by coaching experts, but increase in quality with additional input information. Journal of Sports Science and Medicine, 2024;23(1):56–72. doi:10.52082/jssm.2024.56
  • Havers T, Jelonnek C, Masur L, Isenmann E, Sperlich B, Geisler S, Düking P. A professional assessment of training plans for muscle hypertrophy and maximal strength developed by generative artificial intelligence. Biology of Sport, 2025;42(4):353–380. doi:10.5114/biolsport.2026.152350
  • Yang Z, Zhang X, Li H, Ye J. More details, less variability? A crossover design study on the impact of information granularity on ChatGPT's training program stability. Biology of Sport, 2026;43:379–392. doi:10.5114/biolsport.2026.154148
  • Akrimi S, Schwensfeier L, Düking P, Kreutz T, Brinkmann C. ChatGPT-4o-generated exercise plans for patients with type 2 diabetes mellitus: assessment of their safety and other quality criteria by coaching experts. Sports, 2025;13(4):92. doi:10.3390/sports13040092
  • D'hoe B, Kirk D, Boone J, Colosio A. ChatGPT outperforms personal trainers in answering common exercise training questions. Journal of Sports Science and Medicine, 2026;25(1):235–261. doi:10.52082/jssm.2026.235
  • Düking P, Zinner C, Trabelsi K, Reed JL, Holmberg HC, Kunz P, Sperlich B. Monitoring and adapting endurance training on the basis of heart rate variability monitored by wearable technologies: a systematic review with meta-analysis. Journal of Science and Medicine in Sport, 2021;24(11):1180–1192. doi:10.1016/j.jsams.2021.04.012
  • Romeo A, Edney S, Plotnikoff R, Curtis R, Ryan J, Sanders I, Crozier A, Maher C. Can smartphone apps increase physical activity? Systematic review and meta-analysis. Journal of Medical Internet Research, 2019;21(3):e12053. doi:10.2196/12053
  • Lacroix A, Hortobágyi T, Beurskens R, Granacher U. Effects of supervised vs. unsupervised training programs on balance and muscle strength in older adults: a systematic review and meta-analysis. Sports Medicine, 2017;47(11):2341–2361. doi:10.1007/s40279-017-0747-6

Frequently asked questions

Can ChatGPT or another AI write a good triathlon plan?

It can write a reasonable first draft, but studies where coaching experts rated AI-generated running and strength plans found them sub-optimal or moderate. Quality improved when the AI was given more detail about the athlete. No study has tested AI-written triathlon plans on real triathletes.

Are adaptive training apps better than a static plan?

Possibly for some measures, but the evidence is limited. A meta-analysis of 8 studies found that adjusting endurance training using heart rate variability improved submaximal fitness markers compared with a fixed plan, but not performance or VO2peak. Commercial adaptive algorithms haven't been independently tested in triathletes.

What does a human coach do that an app can't?

Judgement with context: spotting that a pattern of niggles, poor sleep and a stressful month needs a different week, handling injury and illness, adjusting race goals, and the relationship that makes you report honestly. Supervised training beat unsupervised training in older adults, though that's a different population.

Should I use an app, an AI chatbot or a coach?

It depends on your experience, budget and how complicated your life and injury history are. Many age-groupers do well with a structured app plus honest self-monitoring, and use a coach or physio for specific problems. If you're injured, stuck, or stepping up a distance, a human with context is more likely to help.

The evidence behind this guide

1 graded principle, 8 sources. Open one to see the sources.

AI-generated training plans are rated sub-optimal to moderate by experts and improve with detailed input, and adaptive (HRV-guided) training improves submaximal markers but not performance; none of this has been tested in triathletesLimited evidence8 sources

In practice: Use apps and AI tools for structure, load progression and general explanations, give them detailed information, check plans against basic principles, and keep injury, illness and medical decisions with a clinician or experienced coach. Inference from expert ratings and other populations, not tested in age-group triathletes.

Peer-reviewed studyIndependently verified
ChatGPT generated training plans for runners are not rated optimal by coaching experts, but increase in quality with additional input information ↗
Düking P, Sperlich B, Voigt L, Van Hooren B, Zanini M, Zinner C · 2024 · Journal of Sports Science and Medicine
Coaching experts rating three AI-generated plans for a fictional runner

Three ChatGPT six-week running plans with increasing input detail, rated by coaching experts on 22 criteria (1-5 scale). Plans were sub-optimal; a median rating of 3 was given 0, 6 and 13 times for plans 1-3; the least-informed plan was rated lower than the most detailed on 15 criteria. Authors advise against using ChatGPT plans without expert feedback. Paper ratings of a fictional profile; number of raters not in abstract.

Peer-reviewed studyVerified — correction applied
A professional assessment of training plans for muscle hypertrophy and maximal strength developed by generative artificial intelligence ↗
Havers T, Jelonnek C, Masur L, Isenmann E, Sperlich B, Geisler S, Düking P · 2025 · Biology of Sport
10 experienced coaches (exercise science degree, at least 2 years coaching) rating LLM plans

10 experienced coaches rated hypertrophy and strength plans from GPT-3.5 (ChatGPT, Copilot) and Google Gemini on 27 criteria: overall quality moderate, tool differences on only 1-2 criteria, one tool most inconsistent. LLM plans can be an initial framework but expert supervision remains crucial. Fictional advanced trainee.

Peer-reviewed studyIndependently verified
More details, less variability? A crossover design study on the impact of information granularity on ChatGPT's training program stability ↗
Yang Z, Zhang X, Li H, Ye J · 2026 · Biology of Sport
11 experts (mean 35.4 y, 18.1 y experience) rating 8 AI-generated programmes

Simple and detailed prompts each entered four times into GPT-4 (8 programmes), rated blind by 11 experts on a 15-item scale. Detailed prompts scored higher for personalisation, safety, feasibility and overall (not effectiveness) and gave more stable outputs, but repeated identical prompts still produced structurally and qualitatively different programmes.

Peer-reviewed studyIndependently verified
ChatGPT-4o-generated exercise plans for patients with type 2 diabetes mellitus: assessment of their safety and other quality criteria by coaching experts ↗
Akrimi S, Schwensfeier L, Düking P, Kreutz T, Brinkmann C · 2025 · Sports
3 coaching experts reviewing plans for 3 fictional patients with type 2 diabetes

Discussion paper with preliminary data: three fictional patient profiles, three prompts each, reviewed by three coaching experts. Some plans had serious safety issues, especially for patients with secondary diseases or complications; most included key training principles but had deficits such as feasibility; detailed prompts tended to rate better.

Peer-reviewed studyIndependently verified
ChatGPT outperforms personal trainers in answering common exercise training questions ↗
D'hoe B, Kirk D, Boone J, Colosio A · 2026 · Journal of Sports Science and Medicine
9 personal trainers' questions; rated by 18 personal trainers and 9 topic experts

9 level-4 personal trainers supplied frequent questions and their answers; ChatGPT 3.5 answered the same questions; 18 trainers and 9 topic experts rated them. ChatGPT scored higher overall on 6 of 9 questions (correctness 5, comprehensibility 6, actionability 5); trainers never scored higher. General information questions, not individual plans.

Meta-analysisIndependently verified
Monitoring and adapting endurance training on the basis of heart rate variability monitored by wearable technologies: a systematic review with meta-analysis ↗
Düking P, Zinner C, Trabelsi K, Reed JL, Holmberg HC, Kunz P, Sperlich B · 2021 · Journal of Science and Medicine in Sport
8 endurance-training studies, 198 participants (sex split not reported in abstract)

8 studies, 198 participants, 9 interventions. HRV-guided plans usually had fewer moderate/high-intensity sessions than predefined plans. HRV-guided training improved submaximal physiological parameters (g = 0.296, 95% CI 0.031 to 0.562) but not performance (g = 0.079) or VO2peak (g = 0.171) significantly; fewer performance non-responders.

Meta-analysisIndependently verified
Can smartphone apps increase physical activity? Systematic review and meta-analysis ↗
Romeo A, Edney S, Plotnikoff R, Curtis R, Ryan J, Sanders I, Crozier A, Maher C · 2019 · Journal of Medical Internet Research
9 RCTs, 1740 adults (general and specific health populations)

9 RCTs, 1740 adults; in 6 studies apps produced a non-significant increase of 476.75 steps per day (95% CI -229.57 to 1183.07; P = .19). Apps were more effective over less than 3 months and when targeting activity alone. General adults, not athletes.

Meta-analysisIndependently verified
Effects of supervised vs. unsupervised training programs on balance and muscle strength in older adults: a systematic review and meta-analysis ↗
Lacroix A, Hortobágyi T, Beurskens R, Granacher U · 2017 · Sports Medicine
11 studies of healthy adults aged 65 and over

11 studies in healthy adults aged 65+. Supervised balance/resistance training beat unsupervised for muscle strength/power (SMD 0.51) and several balance measures (e.g. test batteries 0.53, dynamic balance 0.35); advantage larger against fully unsupervised programmes. Few studies, moderate quality (median PEDro 5).

Full write-up of this principle

Get The Forward

Research-led triathlon guidance, straight to your inbox. No spam, no filler — unsubscribe any time.