There's a pattern in endurance tech right now: every platform announces "AI-powered training." You open the feature, and it's a ChatGPT wrapper where you type "what workout should I do today?" and it generates a generic interval session that has nothing to do with your actual fitness.
That's not what we built at Baseline. And I want to be transparent about what we did build, why we built it that way, and how we think about the line between useful AI and gimmick AI.
This is not a press release. It's an honest explanation of our product decisions, including where we said no to certain AI features.
The three things Baseline AI actually does
Baseline uses large language models (specifically Google Gemini) in exactly three places. Each one was designed to solve a specific problem our users had. None of them is a chatbot.
1. Daily AI insight card on the dashboard
This is the most-used AI feature in Baseline. Every morning, when you open the dashboard, you see a card at the top:
"Your HRV is 8% below your 30-day baseline. Your CTL has risen 5 points in the last 10 days. Your resting heart rate is stable. This pattern is consistent with productive training load accumulation - you're in a build block. Consider a recovery day if you feel fatigued, but the data alone doesn't indicate overreach."
Or:
"Your weight trend is down 2.5% over 14 days while training load has increased 15%. Your HRV is 14% below baseline. This combination of signals suggests possible low energy availability. Consider increasing caloric intake, particularly carbohydrate, and monitoring for 5–7 days."
Hundreds of variations, personalised to your data, generated fresh every morning.
How it works: We have a set of detection rules that scan your data for patterns - HRV deviation from baseline, CTL ramp rates, weight trends, sleep changes, HRV-RHR correlation shifts, and about 20 other signals. When a pattern crosses a threshold, the detection rule passes the relevant data points to Gemini with a structured prompt: "Here's the athlete's HRV trend over 14 days, here's their CTL trend, here's their sleep efficiency. The pattern is [X]. Generate a 2–3 sentence insight that explains what this means and suggests one actionable step."
The model is not given access to your identity, your exact workout names, or any identifying information. It receives anonymised trend data and returns text.
Why this isn't a chatbot: We don't have a text box where you type questions. The insight card appears automatically. You read it in 15 seconds. You either take the suggested action or you don't. It's push, not pull. The cognitive load is near zero.
2. Weekly summary email
Every Sunday, Baseline sends a summary email (opt-in, of course). The AI generates a narrative summary of your week:
"This was your highest training load week of the month - 785 TSS, up 18% from last week. Your CTL rose from 62 to 67. Your HRV dropped 7%, which is an expected response to increased load. Your sleep was stable. Key insight: your form (TSB) dropped to -18, the lowest in 30 days. Consider scheduling a recovery day Tuesday if you're planning another hard week."
The email also includes a section of notable trends - "You visited 6 new running routes this week" or "Your fastest 10K of the season was Thursday" - generated from simple data aggregation, not AI.
The AI is used only for the narrative framing. The numbers, trends, and recommendations are rule-based. The AI translates the rules into readable English.
3. Ask Baseline (the one exception)
The one feature that looks most like a chatbot is "Ask Baseline" - a natural language interface to your data. You can ask:
- "Show me my HRV trend for the last 30 days"
- "What's my best 10K time this year?"
- "Compare my training load from this month to last month"
- "How many miles did I run in March?"
Ask Baseline translates these questions into database queries - not into training advice. It won't tell you what workout to do today. It won't write your training plan. It won't tell you if you're overtrained. It retrieves data from your account and presents it clearly.
The reason: we don't believe a general-purpose language model can give reliable individualised training advice. The liability, the accuracy risk, and the potential harm from bad advice are too high. A runner who asks "should I race this weekend?" should not get an AI-generated answer based on a training forum's collective wisdom. They should talk to their coach, read their own data, and make a human decision.
Ask Baseline is a data retrieval tool, not a coach. This distinction matters deeply to us.
What AI doesn't do in Baseline
This list is as important as what AI does.
AI does NOT write training plans
We don't have a "generate your week" feature. There are good platforms for structured training plans (TrainingPeaks, Final Surge, TrainerRoad, TriDot). We're not trying to replace them. Our users typically have a coach or follow a structured plan elsewhere. Baseline is where they see the full picture - training, recovery, health, nutrition - in one place.
AI does NOT give medical advice
If your data looks concerning (sustained HRV suppression, erratic sleep, large unexplained weight changes), Baseline will suggest talking to a healthcare provider. It will not diagnose anything. The insight cards are carefully prompt-engineered to avoid diagnostic language. We use "suggests possible" and "consider checking" language deliberately.
AI does NOT modify your training load metrics
CTL, ATL, TSB, HRV baselines, training load - all of these are computed deterministically from your data. AI is not involved in any calculation that affects your performance metrics. The AI operates only on the interpretation layer - turning numbers into sentences.
AI does NOT write social content (sharing, posts, etc.)
Some platforms use AI to generate Strava descriptions or social media posts. Baseline doesn't. That feature felt wrong to us - the voice should be yours, not ours.
Privacy: the part that matters most
I want to be direct about this because it's the number one question athletes ask us about AI.
Your data is never trained on.
The Gemini API calls we make do not contribute to training data. We have a signed business agreement with Google that explicitly prohibits the use of API traffic for model training. Your health data - your HRV, your power numbers, your sleep, your weight - never enters a training corpus.
Beyond the AI feature, Baseline's data privacy architecture:
- All athlete data is stored in per-user encrypted buckets
- The AI only receives the specific data points needed for the current insight (your last 14 days of HRV, not your entire history)
- Data is not shared between users - LLM is stateless per request, and we don't persist the conversation context in a way that cross-contaminates users
- You can export all your data at any time (full JSON download)
- Delete your account and all data is permanently deleted within 30 days
We don't sell data. We don't share data with advertisers. There are no third-party trackers on the Baseline dashboard. The business model is subscription revenue from athletes who find the product valuable - not monetization of your health information.
Why Gemini specifically
We chose Google Gemini over OpenAI and open-source models for three reasons:
-
Latency. Gemini is fast - responses return in 300–800ms for the short prompts we use. OpenAI's comparable models were 2–3x slower when we tested. For a daily insight card that appears when you open the dashboard, speed matters.
-
Structured output reliability. Gemini's JSON mode is excellent. Our insight generation pipeline works as: detection rules → structured data → Gemini → JSON response → render card. The structured output needs to parse reliably every time. Gemini's JSON mode delivered 99.8% parseable responses in our testing.
-
Data handling agreements. Google's GCP data processing agreements were the clearest on the AI-training exclusion. For a health data product, that clarity matters.
We may switch models as the landscape evolves. The AI features are model-agnostic behind our internal API - we could swap Gemini for Claude or GPT tomorrow if a better option appears. Right now, Gemini is the best fit for our specific use case.
Where the line is
We've thought a lot about where AI is helpful vs harmful in a training context. The framework we use internally:
Helpful AI: Surfaces patterns in your data that you wouldn't see yourself. Explains those patterns in plain English. Suggests one actionable step. Takes 15 seconds to consume.
Harmful AI: Pretends to know you better than you know yourself. Generates training plans without understanding your context. Answers medical questions. Replaces the judgment of a human coach.
Everything in Baseline is designed to stay on the helpful side of that line. We've turned down feature requests for AI-generated training plans, AI coaching, AI race predictions, and AI nutrition plans. Not because we couldn't build them, but because we don't believe they'd be good enough to trust - and the cost of a bad recommendation in training is measured in lost fitness, missed goals, or injury.
What's coming next
The AI features will get better in three ways:
-
Pattern library expansion. We're building more detection rules - currently about 20 patterns, targeting 50+ by year end. Each new pattern is a new type of insight the AI can surface.
-
Longer context windows. Currently daily insights look at ~14–30 days of data. Longer windows (3–6 months) will reveal seasonal patterns: "Your HRV tends to drop in July every year - this is consistent with your heavy summer training block, but you might consider a mid-summer recovery week this year."
-
User preference tuning. Over time, the AI will learn which types of insights you find useful and which you ignore, adjusting the frequency and emphasis accordingly. No data leaves your account for this - it's a per-user preference model.
What we won't add: a chatbot training coach. Ever. If that changes, I'll write a post explaining exactly why, with the same transparency I've tried to show here.
The bottom line
AI in Baseline does three specific, limited things: a daily insight card, a weekly summary, and a data retrieval interface. None of them is a chatbot. All of them are designed to show you patterns in your data that you'd miss otherwise.
We don't think AI replaces coaches, training plans, or your own intuition. We think it highlights the connections between your training, recovery, and health data - connections that are real but invisible when your data lives in six different apps.
The insight card isn't the future of training. But it is a really useful second opinion, delivered in 15 seconds, based entirely on your data, every morning. That's enough.