CatalystLTI
← Field notes
AI & LeadershipJuly 5, 2026· 6 min read

AI made the soft skills the hard part. We still don't train them.

Two reports landed this summer from opposite ends of the field, and they said the same thing.

The first ran in Forbes, reading PwC’s latest jobs data, and it cuts against the whole “AI is coming for the jobs” script. As AI spreads, the market is paying more for the human part of the work. Leadership-focused roles have seen 42% faster wage growth since 2021. In AI-exposed entry-level jobs, 52% of the newly demanded skills are the kind we used to call “senior”: judgment, communication, owning a problem nobody has scoped yet. And in the detail that stuck with me, Ford rehired 300 engineers after deciding that automated quality without human judgment isn’t actually quality.

The second ran in Communications of the ACM, making the same case from inside software itself. You’d expect agentic coding, the kind that writes and runs large chunks of the code for you, to make the human skills matter less. Their conclusion was the opposite. When the machine handles the mechanics, what’s left is communication, framing, and judgment: saying clearly what you want, catching where the AI is wrong, making the call when the data’s incomplete. Agentic development, they argued, may make those skills matter more, not less.

Put the two together and you get a conclusion that should change how engineers train, and mostly hasn’t. The part of the job everyone filed under “nice to have” just became the scarce, expensive, decisive part. AI automated the thing we spent years getting good at, and left us holding the thing we never practised.

“Soft skills” is the worst name in tech

Start with the name, because the name is half the problem. Call something a “soft skill” and you’ve already told everyone it’s optional, innate, and a little unserious: the stuff you either have or you don’t, filler next to the real work of shipping. Nobody puts “executive presence” on a sprint board.

But look at what actually decides an engineer’s career after the first few years. Whether the budget survives the review. Whether the senior engineer takes the feedback or digs in. Whether the room believes you when you say the date will slip. None of that is soft. It’s just hard in a way that doesn’t compile. You find out you were bad at the exec update in the exec update, and there’s no test suite that would have caught it.

The two reports are saying the same thing in different accents: the market has started pricing these skills like they’re hard, because they are, and because they’ve gotten rare.

We drill the automatable part and rehearse none of the rest

Here’s the asymmetry that gets me. For the part of engineering AI is busy automating, we built an entire training industry. LeetCode, system-design courses, mock interviews, thousands of hours of deliberate reps at exactly the work a model can now do in seconds.

For the part that just became decisive — the hard conversation, the pushback, the call under pressure — the plan is “read a book and hope.” We treat leadership as a personality you grow into instead of a skill you rehearse. So engineers walk into the highest-stakes conversations of their careers cold, having practised the thing being automated and nothing else.

Reading about deadlifting is not deadlifting. Everyone knows this about the gym and forgets it about leadership.

Catalyst is where you take the reps

That gap is the reason Catalyst exists. It’s a private simulator where you rehearse the high-stakes conversation out loud, before it happens for real, against an AI actor who plays the other side and pushes back the way a real person does: the principal defending the system you want to replace, the CFO who wants a number and won’t take a story, the report you have to put on a plan.

You talk. Not type. The actor stays in character. And there’s no score on the screen while you’re in it. A live scoreboard makes you perform for the meter instead of staying in the room. The measurement comes after.

When the session ends, Catalyst reads back what you actually did, on the two signals that quietly decide how you land:

Presence Index asks whether you command the room, rolled up across six dimensions: composure, clarity, authority, engagement, responsiveness, and conviction. It’s built from how you actually spoke, and more to the point, how steady you stayed when the actor turned up the pressure.

Decisiveness Score asks whether you make the call: commitment language, ownership, and whether you close with a clear next step or trail off into “maybe we could look into it.”

Both numbers come from deterministic rules, not an AI guessing your potential. Every point traces back to a moment in your transcript: the composure dip at 1:47 when the CFO cut you off, the recommendation you closed without an owner. Then it hands you one drill aimed at your weakest signal, you run it, and you re-measure. Rehearse, measure against the evidence, drill the gap, watch the line move. It’s closer to LeetCode for the human part than to a course you sit through, and that’s the point, because the human part is the one nobody built a LeetCode for.

AI didn’t kill the soft skills. It removed the excuse.

For years you could get away with being the brilliant engineer who’s rough in the room, because the code was the job and the code was enough. That deal is expiring. When the model writes the first draft of the code, what’s left of the job is the conversation around it: the framing, the disagreement, the judgment call. That’s the part you’re now measured on.

The reports call it a paradox. I think it’s simpler than that. The automatable part got automated. The human part didn’t, and it got more valuable because everything around it got cheaper. The engineers who win the next decade won’t be the ones who prompt best. They’ll be the ones who practised the part everyone else still thinks you can’t.

You can. It just takes reps, and a place to take them where getting it wrong costs nothing.

Catalyst — measured leadership rehearsal for engineers. Practice the conversation before it costs you.

Sources

Next station

Take your first rep.

Free trial — 15 days, about 2 hours of conversation. No card.

Start practising