The Duolingo Trap: Why Your AI Coaching Tool Looks Broken (And It's Not the Tool's Fault)

My Dad had surgery recently. General anaesthesia, the works. He woke up groggy, could barely see straight, and the first thing he did was grab his phone and complete his Duolingo streak.

That's what a product with real pull looks like. Nobody told him to do it. Nobody mandated it. He just wanted to.

Now imagine a company buys 300 people access to an AI coaching tool. Free, no strings, no manager breathing down their neck. Three weeks later, the adoption dashboard looks terrible.

But here's what most people would miss: that tool had incredible pull with the people who actually needed it. The deployment model just buried it.

The product wasn't the problem. The way B2B organisations roll out software was.

The numbers that almost fooled me

Let me walk you through what happened, because the data tells two very different stories depending on which lens you use.

300 people signed up. That bit was easy. It was free, and I'd positioned it well. Of those 300, just over 100 came back more than once. About 40 used it enough to hit the free tier limit. And the number who converted to a paid plan off the back of all that?

Barely any. Close to zero.

If I'd been reporting this to a board, or to an enablement leader who'd just bought the tool for their team, the conclusion writes itself: low adoption, poor engagement, failed initiative. Rip it out, try the next vendor.

But that's the wrong story.

The right story is that 100+ people came back voluntarily. Nobody told them to. 40 people used it so much they ran out of free credits. And across roughly 10,000 interactions, only 7 were roleplays.

Think about that for a second.

10,000 events. 7 roleplays.

These people weren't logging in to practise. They were logging in because they were stuck. Stuck on a deal. Stuck on an email. Stuck prepping for a call that was 20 minutes away. They weren't training. They were working.

And that distinction, between training and working, is where the whole measurement falls apart.

Welcome to the Duolingo Trap

Here's the analogy I keep coming back to.

Imagine your company buys everyone a Duolingo subscription. Leadership announces it at the all-hands. "We're investing in language skills! Download the app, start your streak, let's go."

Three months later, someone pulls the usage data. 15% have opened it more than twice. Daily active users are in single figures. The average streak is 4 days. Leadership looks at the numbers and says: "Duolingo doesn't work."

But Duolingo has over 100 million monthly active users. It's one of the most successful learning apps ever built. My Dad maintains a streak that survived surgery and general anaesthesia. That's genuine pull.

So what happened?

Nobody was trying to learn a language. They didn't have a trip booked to Barcelona. They weren't moving to Tokyo. There was no problem to solve, no urgency, no personal stake. You gave them a brilliant tool and zero reason to use it.

That's the Duolingo Trap. And every single centrally-purchased AI coaching tool falls straight into it.

The push problem

The enablement industry is built on push. Company buys tool. Company announces tool. Company mandates usage (or, more commonly, "strongly encourages" it, which everyone correctly interprets as optional). Company measures logins. Logins decline. Company blames tool. Rinse, repeat.

I've lived this cycle from both sides (buying the tools and building them). And the maths never works. Let me prove it.

A sales team of 50 reps. You roll out an AI coaching platform. Let's be generous and say 60% log in during the first week (that's actually high; most vendors will tell you privately that Week 1 activation is closer to 30-40%). By Week 4, you're down to 20% logging in at least once. By Month 3, you've got maybe 8-10 regulars.

The enablement leader looks at this and sees failure. The vendor looks at this and blames the enablement leader for not driving adoption. The reps look at both of them and think: "I've got 47 emails to send and a pipeline review in an hour. I don't have time to role-play with a chatbot."

Everyone's right. And everyone's wrong.

The tool probably works fine. The reps probably would benefit from it. The enablement leader probably did their best to roll it out. But the deployment model was broken from the start, because it relied on reps doing something they didn't ask to do, at a time that wasn't convenient, for a benefit they couldn't see yet.

That's push. And push doesn't scale.

The coaching gap is real. But not where you think.

MySalesCoach published their State of Sales Coaching report this year, and the numbers are grim.

41% of sales reps say they're never or rarely coached. 45% rate the quality of coaching they receive as below average, up from 29% the year before. Meanwhile, 64% of sales leaders believe they're coaching more than ever.

Read those together. Leaders think they're coaching more. Reps say the coaching is getting worse. Someone's lying, or (more likely) nobody's measuring the same thing.

And here's the kicker: only 34% of sales leaders have ever received any formal training on how to coach.

So we've got untrained coaches, coaching more often, delivering worse results. And the industry's answer? is to buy AI tools and push them at reps who already feel over-coached and under-helped.

Sound familiar?

Pull changes everything

When I looked at the data from those 300 users properly (not the dashboard summary, the actual usage patterns) something clicked.

The people who became power users, the ones who hit the limit, who came back day after day, almost all started the same way.

They had a stuck deal.

Not a training goal. Not a manager telling them to log in. Not a certification deadline. A deal. A real one. With a real buyer who'd gone quiet, or a discovery call they'd botched, or an email they couldn't get right.

They came to the tool with a problem. The tool helped them solve it. They came back with the next problem. And the next.

That's pull. And in the data, pull looks nothing like push.

Push adoption looks like a cliff. High activation in Week 1, steep decline, long tail of nobody logging in. Pull adoption looks like a staircase. Low initial numbers, but each user who arrives actually stays, and usage compounds.

The irony is that pull looks worse in the first month and better in every month after that. But most companies measure at Month 1 and make their decision.

The deployment model is the product

SalesHood published research last week showing that standalone AI roleplay "falls short" of driving real behaviour change. They're right. But I think they've spotted the symptom whilst missing the disease.

Roleplay works. But roleplay deployed as a push activity ("go practise three times this week") creates compliance, not capability.

The framing is backwards.

Instead of: "Here's a coaching tool. Go use it."

It should be: "You've got a deal that's stuck. Let's fix it. Oh, by the way, the thing that just helped you? That's your coach."

The first framing is Duolingo-for-the-office. The second is Google Maps. Nobody needs to be told to use it, because it's solving a problem they have right now.

At Replicate Labs, this is what we've been building towards. We call it "Show Me, Let Me, Coach Me." On Day 0, the AI does the heavy lifting. It writes the email, preps the call, structures the deal review. The rep gets immediate value. Over time, as the rep builds capability, the AI shifts from doing to coaching. The rep barely notices the transition, because they were never "in training." They were always working.

The deployment model isn't a feature of the product. It is the product.

So what do you actually do about it?

If you're an enablement leader who's bought (or is about to buy) an AI coaching tool, here's what I'd do differently knowing what I know now:

1. Kill the launch event. Don't announce it at the all-hands. Don't send the "exciting new tool!" email. Every one of those signals screams "this is optional and you should ignore it."

2. Start with stuck deals. Find the 10 reps with the most stalled pipeline. Give them the tool with one instruction: "Tell it about a deal you need help with." That's it. No onboarding. No training on the tool. Just: here's a problem, here's something that might help.

3. Measure the right thing. Stop measuring logins. Start measuring outcomes. Did the deal move? Did the email get sent? Did the call go better? If the tool is solving real problems, usage will follow. If it's not solving real problems, no amount of manager nudging will save it.

4. Build the workflow, not the habit. The best AI coaching doesn't require a new habit. It shows up where the rep already works: in Slack, in the CRM, in their email. If your tool requires a separate login and a separate workflow, you've already lost.

5. Let the pull spread. When one rep uses the tool to unstick a deal and tells their mate about it at lunch, that's worth more than 50 enablement emails. Create the conditions for pull to happen and then get out of the way.

The uncomfortable truth

The AI coaching tools aren't broken. Most of them are genuinely good. I've seen what these vendors have built (the coaching engines, the roleplay, the call analysis) and a lot of it is impressive.

But a brilliant tool deployed badly is indistinguishable from a bad tool.

And right now, the industry is measuring itself on push metrics (activation rates, login frequency, module completion) whilst ignoring the only metric that actually matters: did this change what the rep did on their next call?

That's the Duolingo Trap. Brilliant product. Wrong deployment model. And the product gets the blame every time.

Stop buying tools. Start solving problems.

---

If you've lived through the push-adoption cycle and have a horror story (or a success story), I'd genuinely love to hear it. Drop me a message. I respond to every one.