The real challenge is:
How can we help AI understand why a workout turned out the way it did?
This was one of the central questions we considered while developing Fit2AI 2.1.
In this version, we began bringing COROS data more deeply into Fit2AI’s AI analysis workflow.
This means more than adding another data source.
More importantly, we began trying to move AI from “seeing one workout” toward “understanding a period of training.”
One workout does not tell the whole story
Suppose you completed a 10km threshold run today.
Average pace: 3:40/km.
Average heart rate: 175.
If you give AI just these numbers, it can certainly analyze them.
But questions immediately arise:
Is 3:40/km fast or slow for this person?
Is a heart rate of 175 normal for them?
Does slowing down in the second half indicate a lack of fitness or fatigue from recent training?
How does this workout compare with last week’s?
Is the user currently building a base, increasing intensity or preparing for a race?
Without historical data, many judgments lack context.
That is why we believe:
What really matters in AI workout analysis is not how many metrics the model sees, but how much useful context it has.
COROS provides more than a single activity
COROS contains a wealth of data that can help us understand training.
An individual activity record is important, of course.
Pace, heart rate, cadence, distance, elevation and changes across the session help us understand what happened in that workout.
But things become more interesting when the data spans multiple workouts.
AI can begin to compare:
How has training volume changed recently?
Has the quality of similar workouts changed?
Has heart rate at the same pace changed?
Is there enough recovery time between high-intensity sessions?
Are threshold-related metrics changing?
Is an unusual performance an isolated event or part of an ongoing trend?
At this point, AI is no longer looking at a single workout record.
It is looking at a training history.
We cannot just send all the data to AI
There is also a very practical engineering problem:
There is a lot of workout data.
A one-hour run can contain a large number of individual samples. Sending all the raw data from dozens of past workouts to a model would be expensive and would not necessarily produce better results.
More data does not automatically mean more information.
One important task for Fit2AI is therefore to turn:
Workout data
into:
Training context that AI can understand effectively.
Different kinds of questions need different data.
If the user asks:
“Why did I slow down over the last two kilometers today?”
The focus may be on pace, heart rate and changes between splits in that workout.
If the user asks:
“Why has my threshold pace declined recently?”
Then today’s data alone is clearly not enough.
The system needs to expand its analysis to a period in the past and find relevant workouts and changes in metrics.
And if the question is:
“Have I been pushing my training too hard lately?”
We may need to examine the relationship between training frequency, intensity distribution, training load and recovery.
So we do not want Fit2AI to simply send the entire database to the model.
The real problem to solve is:
Once the user asks a question, what should we show AI?
From data queries to training context
We have gradually come to see Fit2AI’s AI analysis as several distinct stages.
First comes the user’s question.
Then the system determines which training information the question needs.
Next, it retrieves relevant workout data and organizes the raw records into a structure better suited to analysis.
Finally, this information and the user’s question form the context for AI.
In simple terms:
User question
Identify the data needed
Retrieve COROS training information
Organize the training context
AI analysis
Ask follow-up questions
The biggest difference from simply uploading a FIT file is that data can now be retrieved in response to the question.
Users do not have to find a workout from a particular day and upload records one by one.
The training history itself becomes part of the conversation.
Conversation matters more than a workout score
Many fitness products end by giving a workout a verdict:
Good.
Average.
Insufficient recovery.
Training load too high.
These results have value, of course.
But we especially want Fit2AI to retain one important capability:
Keep asking.
If AI says:
“Your heart rate rose noticeably in the second half.”
The user can follow up:
“Is the weather a likely cause?”
Then ask:
“How does it compare with that 15km threshold run last month?”
And continue:
“Should I still do the same workout next week?”
When training data becomes context for a conversation, AI analysis starts to move beyond a one-off report into a problem-solving process that can go progressively deeper.
This is one of the biggest differences between Fit2AI and a traditional workout data dashboard.
AI needs to know what it does not know
The more data we connect, the more clearly we see something else:
Workout data is never complete.
Your watch knows your heart rate.
But it may not know why you only slept four hours last night.
It knows your pace dropped today.
But it may not know it is 32°C outside.
It can see your training volume increasing.
But it does not know that today actually felt very easy to you.
So Fit2AI’s goal is not to create the impression that “AI knows everything.”
On the contrary, we want the system to distinguish:
Which conclusions come from the data.
Which are only possible explanations.
Which information is still missing.
When important information is missing, the best response from AI is not to force an answer, but to ask the user more questions.
From recording data to using it
Over the past decade or more, sports devices have solved an important problem:
How to record training.
We now wear powerful sensors on our wrists, and fitness platforms hold ever-longer personal training histories.
This raises a new question:
What else can we do with the data that already exists?
Bringing COROS data into Fit2AI 2.1 is just one step in that direction for us.
We want training history to be more than an activity list that keeps getting longer.
It should be searchable, comparable and open to analysis—and to questions.
Ultimately, when you ask:
“How has my training really been going lately?”
AI should see more than today.
It should know how you got here.
That is where we believe AI can truly change the experience of using workout data.