Can an AI training coach get you race-ready in 8 weeks?

“Hi Charlie, it looks like you didn’t train as much as planned last week. Maybe this was a tough time for training. No worries, let’s replan – I know you can do this!”

Six weeks into my eight-week training block for a duathlon, and I’ve let Hugo, my coach, down. The previous week had involved travel for work, and although this included a two-hour ride into the outskirts of Copenhagen for a bike launch (the perks of being a triathlon journalist), a relentless itinerary and tight article deadlines meant the running trainers I’d packed had remained firmly in my suitcase.

I’m not that worried that I’ve disappointed Hugo, though. In fact, I don’t think it would be possible. While there are plenty who would describe their coach as emotionless, mine literally doesn’t have any. That’s because Hugo’s not human.

Hugo is the faceless avatar of training platform HumanGO, and has been setting, analysing and tweaking my training as I build towards my first the standard distance duathlon at Hever Castle. And with two weeks to go, my prospects of drawing a respectable line in the sand have taken a hit.

How to train for a duathlon: everything you need to know

Coach in the machine

HumanGO is just one of a number of training platforms that have introduced artificial intelligence (AI) into the world of coaching.

Rather than the one-size-fits-all plan that you might find in some places, an AI coach promises a more personalised, tailored approach – one closer to a real-life human.

The app creates schedule of workouts based on tests like functional threshold power and maximum heart rate, while it adapts them as required as you build towards a goal event.

“The AI is driven by a few things,” explains Pav Bryan, HumanGO’s master cycling coach. “The individual – your threshold, your availability, stuff like that – and then it’s also driven by the goals you’ve set.” He adds that it estimates what your goal’s racing intensity will be based on you and your ability, and plans a periodised schedule of workouts created by real-life coaches.

“In an ideal world, you’d start with nine months to go, and you could get a good base phase, and move into build and peak, and that’s where you would start to do more intensity like the goal itself, and it’s all adapted around your availability.”

I only had eight weeks, but I was keen to see how it would prepare me for my duathlon debut.

An experienced runner and cyclist with a marathon PB of 2:43 and a 1:05 bike leg in a flat Olympic triathlon, I’ve dabbled in various structured plans in my time – from following free ones found online to working with a coach – and I was also interested to see whether it was an improvement on the former, and a potential replacement for the latter.

A strong start

Getting started was simple. On firing up the app, I was asked how much training experience I had and was prompted to enter my Chronic Work Load or Strava Fitness score before I created my ‘goal’ – in this case training for a race, although there were other options like improving general fitness and performance if there are no events on the horizon.

After selecting my type of race, the distance and date, I set the target of ‘challenging myself’ given that I had no PB to aim for but can’t do anything by halves. Finally, it asked for recent threshold pace and power numbers.

A couple of minutes later, an eight-week plan was generated that included a base, build, peak and taper phase. With a further bit of customisation (I’m a new dad and the days of 10-plus-hour training weeks are on pause for the time being), I had a realistic-yet-achievable workload based around two rest days, three runs and three rides per week, including one brick session.

Digging into the plan, it looks like anything a coach would create, mixing low intensity sessions with an increasing number of high intensity intervals as it progresses towards race day. It also has a neat feature – goal stats – that highlighted compliance, readiness and projected potential in a simple infographic.

In two months, it forecasted I would go from 60% to 78% readiness – a moveable target that could also increase with better-than-predicted training performances.

Computer says no

The first few weeks of training went as planned. Four months since I’d got round the London Marathon on little training or sleep, it was nice to have structure to my workouts and some extra motivation to force me out of the front door (or onto the turbo trainer).

I was getting on well with Hugo too, who would provide weekly updates on my progress. It would also explain the physiological benefits of the training I’d been doing during the base phase, and what I could expect during the build.

The honeymoon period ended though with a holiday during my fourth week of the plan. Although it was in the UK, logistics (a car full of baby paraphernalia) meant I wouldn’t be able to take my triathlon bike along with me.

Opening the message function, I asked if I could swap out any bike workouts for runs instead. “It seems there was a misunderstanding,” parped the bot. “It seems that I cannot assist with the workout swap directly.”

In what would be a routine conversation with a real-life coach, complete with a tailor made solution to boot, I was left a bit stuck.

It was a recovery week after a three-week base phase, so it wouldn’t have been the end of the world to miss the two rides, but I didn’t want to impact my goal stats.

Rather than forgo the workouts completely, I decided to swap out the bike sessions for runs, trying to match the duration, training load and intensity as best as I could, but it was far from perfect. My projected potential took a hit as a result, dropping down to the low 70s. I started to lose a bit of faith in the machine.

Error message

With the start of the peak phase came an increase in intensity. Threshold intervals were sprinkled into runs and I experienced the rush of endorphins that come with pushing yourself harder than zone two and tempo work. I was confident in my progress, even if I did breach the upper limits a bit much for Hugo’s liking.

The bike though was another story. Switching to the turbo for all of my rides and training solely based on power, things weren’t adding up. After completing a threshold pyramid session, I hit the duration, load and intensity as planned, but my compliance was listed at 73%.

Hugo’s workout note said “don’t go too easy”, and looking at my heart rate data, I had only just tickled zone 3 for 17 seconds, with none of the 10 minutes planned at zone 4. I recorded my perceived exertion as ‘comfortable’ and noted how the training felt.

This continued for another few sessions – falling short on key build workouts with no adjustment in my training zones on the cards – before I decided to take things into my own hands.

While the simplest solution would have been to take another FTP test, I didn’t want to throw a big load into the mix off the back of other threshold work.

Instead, I modified each ride, upping the power by 10% and targeting the workout’s planned heart rate zones instead. It wasn’t seamless, but compliance was back to 100% and Hugo was generally happy.

Short of stimulus

That was until my work trip. A combination of travel overseas, long days and not a lot of free time meant the schedule went out the window. On return, other priorities came to the fore, and I only completed one out of a planned six sessions for the week. But my non-compliance felt deeper and ingrained.

Previously, I have sought out treadmills, got up an hour earlier or squeezed a session into the tightest timeframes when travelling.

But the lack of actual accountability when using an AI coach meant that I didn’t feel guilty about missing a workout. And when my readiness and projected scores dropped to the low 60s, I wasn’t concerned. If anything, it made me question its analysis.

The penultimate week was the plan’s peak phase, and I returned as a motivated, model student, while the taper dialled down the intensity pre-race day.

Good compliance meant I had managed to get my projected performance back to 71, while a smart race planning feature – where you can upload a course GPX and how hard you want to complete each leg – recommended paces and power that would get me to the line without burning through all my matches. It also projected a punchy 2h7m finish time.

Come the event, I tried to follow the plan to a tee. After a fast start on an undulating mixed terrain run with steep inclines, I found myself third. I was hitting my 3:58km pace, but the hills and adrenaline meant my heart rate was at threshold.

After a quick transition, a long drag out of Hever Castle’s grounds slowed the pace, and with the bike course being two 20km laps, the climb would have to be tackled twice, making the 1h4m projection a tough ask.

I settled into a rhythm, keeping my heart rate on the cusp of tempo and threshold, ramping things up and getting out the saddle on any spikes in the gradient. Rolling back into the event village eight minutes down on my race plan, I was happy with my performance given the rolling parcours.

This turned to joy on leaving the transition zone for a second time, I was now up to second place. Shuffling around the abridged 4.1km course, I attempted to hold my pace from earlier, but my legs felt like lead.

Nothing in my training, not even the brick sessions, had prepared me for this. Willing the race to be over, I forced screaming calves to sprint for the finish, securing second place and crossing the line at 2:16:21.

Intelligent alternative

Reflecting on my eight weeks of preparation and race day performance, the AI coach clearly worked. It provided the right training stimulus to get me in shape for the task at hand, refining my fitness to enable me to set a competitive benchmark in a new discipline.

But it was by no means flawless. A user with less training experience might have stuck with the suggested workloads, turning up to an event undercooked and unable to reach their true potential. The inability to communicate and modify training sessions based on other life commitments also seemed like an oversight, but one that could be ironed out easily enough.

The main issue though is the disconnect of ‘working’ with an AI coach. If you want structure and a personalised training plan, then it’s a great tool, but don’t expect to build a relationship and have that extra boost of motivation when the going gets tough.

Free 8-week duathlon training plan