Can AI actually help you learn a new language while running a company?

Learning a language with AI as a busy founder: what works

AI can meaningfully shorten the path to a solid B1 conversational level in about 100 focused hours for a busy operator, but only if you use it to accelerate the boring parts and preserve the effortful retrieval. Cal Newport’s Deep Work argument applies: friction at the moment of recall is where memory gets wired.

The short answer is that AI has genuinely changed what is possible for a busy operator trying to learn a new language, and I have used it hard on Estonian and Japanese over the last year. But most of the "learn a language in 30 days with AI" content misses the actual mechanism, and if you copy those routines you will burn a month and quit. What has worked, both in my own practice and in the current cognitive-science literature, is a specific pairing: AI as an infinitely patient conversation partner and cue-generator, alongside a small, boring, daily habit that AI cannot replace. The tools have improved dramatically in 2026 — apps like Lingochunk, which launched on Hacker News in June, turn native audio into flashcards and shadowing drills automatically, and Claude and ChatGPT will now hold long, correcting conversations in almost any language at a level a private tutor would charge $80 an hour for. What has not changed is your brain.

The mechanism that matters, and that most AI-driven language pitches ignore, is retrieval spaced over time. Language learning is not information transfer — you know intellectually that dom means home in Russian after seeing it once. It is retrieval strength: how reliably you can pull that word out of memory under mild pressure, weeks and months later. Every serious learning-science result of the last twenty years, including the ones Daniel Kahneman touches on in Thinking Fast and Slow when he distinguishes System 1 recall from System 2 effort, points to the same protocol: short, frequent, effortful retrieval, spaced across days. AI does not remove the need for effort; it removes almost every other bottleneck around it. You no longer need to find a tutor. You no longer need to make your own flashcards. You no longer need to guess whether your pronunciation is close. But you still need to sit with the discomfort of trying to remember something you almost forgot, and no amount of tooling will do that for you.

The stack I actually use, and would recommend to a founder who has thirty to forty five real minutes a day: one AI-generated spaced-repetition deck built from the specific vocabulary of your life, one daily voice conversation with Claude or ChatGPT in the target language, and one weekly session with a real human native speaker on iTalki or Preply. The AI does the boring compounding work — the deck, the drills, the on-demand explanations, the shadowing. The human does the thing AI still cannot do reliably, which is give you honest feedback on your prosody and, more importantly, model the emotional texture of the language. Lingochunk-style tools now automate the deck-building from real audio you care about (a podcast episode, a movie scene, a work call transcript) — that turns raw immersion into retrieval material in about ninety seconds, which used to be the biggest bottleneck.

The mistake I made for the first three months was to substitute AI for the boring habit rather than layer it on top. I would have a thirty-minute conversation with Claude, feel productive, close the app, and skip the ten minutes of retrieval practice because I had "done my language work." A month later I could hold a fluent shallow conversation about seven topics I had rehearsed and could not order coffee outside that script. That is the AI trap in learning generally, and it is exactly the failure mode I wrote about in "how to stop being addicted to AI" and "how to stop relying on AI for studying" — the assistant does the visible work while the invisible work, the wiring of memory, gets skipped. Cal Newport's Deep Work argument applies almost verbatim: the substrate skill is built by effortful, focused, uncomfortable retrieval, not by frictionless exposure. AI is superb at removing friction. It is your job to add it back in the one specific place it needs to be added — the moment of trying to recall.

The other under-discussed piece for founders is what to prioritize. If you are learning a language for a real reason — an operating market, a co-founder's family, a life move — you need domain vocabulary faster than a generic app will give it to you, and this is where AI has a real edge over a course. I built a personal 200-word "operator's vocabulary" for Estonian using Claude in a single afternoon — invoicing, contract terms, small talk with a landlord, the specific verbs you need to open a bank account — and drilled it for two weeks. That got me further in real life than the previous three months of generic Duolingo. The prompt template is simple: "You are teaching me [language] for [very specific situation]. Give me the 50 most common words and phrases I will actually hear or need, ranked by frequency in that context." Then feed the output into a spaced-repetition tool of your choice. This is Naval Ravikant's specific-knowledge point applied to language — the useful vocabulary is not the general one, it is the one specific to the game you are actually playing.

The honest ceiling is that AI can get you to a solid B1 conversational level in a target language faster and cheaper than any method has ever done — my rough number is a hundred hours of focused practice over six months for a European language, maybe two hundred for a distant one like Japanese or Estonian. Beyond that, hitting real fluency still requires immersion, real relationships, and a lot of embarrassment in native-speaker contexts that AI cannot manufacture. So the answer to whether AI can help you learn a language while running a company is yes, meaningfully — but only if you use it to accelerate the boring parts and preserve the effortful ones. The founders I know who have actually learned a language in the last year are running that exact split. The ones who "were learning" and quit made AI do all the retrieval work for them, and their brains, quite reasonably, refused to remember what they never had to work for.