Duolingo AI-First: How It Built 148 Courses in a Year (2027)

This guide covers Duolingo's AI-first strategy: what CEO Luis von Ahn's memo said, how the company shipped 148 new courses, and how a product team can apply the same approach without Duolingo's budget. It also covers the step most AI-first stories leave out, which is checking quality with customer evidence. BuildBetter shows up later in that context, as one way to check whether AI-shipped content is causing friction for users. Everything here comes from public sources, and we say clearly where those sources stop.

The Story in Brief: Duolingo's AI-First Bet and 148 New Courses

In late April 2025, Duolingo CEO Luis von Ahn told the company “we are going to be AI-first” in an all-hands email that Duolingo then posted on its official LinkedIn page. The memo framed the change as a strategic turning point: “This time the platform shift is AI.”

A few days later, Duolingo announced 148 new courses. Von Ahn put that number next to the company's history: “Developing our first 100 courses took about 12 years.” The new batch took about a year, according to FOX Business coverage via FOX 29.

Here is a derived comparison. These are our numbers, not Duolingo's. 100 courses in roughly 12 years is about 8 courses per year. 148 courses in about one year is roughly 15 to 20 times that rate. Course scope and complexity may differ between the two periods, so treat the ratio as directional.

The rest of this guide breaks down what Duolingo actually did, the decisions it had to make, and how a product team can apply the same approach this month.

Sourcing note: We could not retrieve the full memo text from LinkedIn. Quotes come from Duolingo's LinkedIn excerpt and from press coverage by Business Today, the Pittsburgh Business Times via WPXI, and FOX Business via FOX 29. The events took place in 2025. The “2027” label refers to this edition of the playbook.

What Duolingo Actually Did: The Mechanics of Going AI-First

Duolingo turned “AI-first” into specific operating rules: contractor phase-outs, new hiring criteria, new performance-review criteria, and a headcount gate. Here is each piece and its source.

• Declared a platform shift. “This time the platform shift is AI.” (LinkedIn)
• Moved contractor work to AI. Duolingo will “gradually stop using contractors to do work that AI can handle.” (WPXI)
• Rebuilt functions instead of patching them. “Most functions will have specific initiatives to fundamentally change how they work” (WPXI). Business Today quotes the memo: “Making minor tweaks to systems designed for humans won't get us there.”
• Changed hiring and reviews. Candidates will be assessed for AI experience, and employees' AI use will be evaluated in performance reviews. (WPXI)
• Gated headcount on automation. Business Today reports that new headcount is approved only if a team cannot automate more of its work. This is a paraphrase, not a direct quote.
• Stated its risk tolerance. “We'd rather move with urgency and take occasional small hits on quality than move slowly and miss the moment.” (Business Today)
• Explained why scale mattered. “Without AI, it would take us decades to scale our content to more learners.” (Business Today)
• Made a people commitment. “We will remain a company that cares deeply about its employees.” (LinkedIn)

None of this started from scratch. WPXI reports that Duolingo cut about 10% of its contract workforce the year before for similar reasons. The company has not disclosed how many contractors it employs.

On the content side, Jessie Becker, Senior Director of Learning Design, described the change to FOX: “It used to take a small team years to build a single new course from scratch.” And: “Now, by using generative AI to create and validate content, we're able to focus our expertise where it's most impactful.”

The Questions Duolingo Had to Answer

Going AI-first forced Duolingo to answer five operating questions in writing. Your team will face the same ones.

Which work should move from contractors to AI?

Duolingo's answer was work “AI can handle,” phased out gradually. That defines the scope by task, not by job title.

How do we scale content creation without scaling headcount?

Generative AI both creates and validates content. Experts then focus “where it's most impactful,” in Becker's words.

How do hiring and performance reviews change in an AI-first company?

Candidates are assessed for AI experience, and employees' AI use is evaluated in reviews. AI fluency becomes part of the job.

How much quality risk are we willing to take for speed?

Duolingo accepted “occasional small hits on quality” in exchange for urgency. It wrote that tolerance down instead of leaving it implied.

How do we reassure employees while changing how every function works?

The memo pairs function-level change initiatives with an explicit statement that the company “cares deeply about its employees.” Whether that pairing worked is a separate question, covered in the backlash section below.

What Did Duolingo's AI-First Memo Say?

Duolingo's April 2025 all-hands memo from CEO Luis von Ahn declared “we are going to be AI-first.” It said the company would:

• Gradually stop using contractors for work AI can handle.
• Assess candidates for AI experience.
• Evaluate employees' AI use in performance reviews.

How Duolingo handled it: it posted the internal email publicly on LinkedIn and framed AI as a platform shift. It also attached clear rules covering contractors, hiring, reviews, and headcount. The specificity is what makes the memo worth studying. Most corporate AI statements stop at aspiration. This one named the levers it would pull.

Do it this month

1. Write a one-page AI operating memo with 3–5 concrete rules. “We will use AI to draft every first-pass help article” is a rule. “We will embrace AI” is a slogan.
2. Name which work categories change first. Pick two or three, not everything.
3. State your quality-vs-speed tolerance in writing. Say whether small errors are acceptable, and where they are not, such as billing or security docs.
4. Include a specific people commitment. Training budgets, redeployment paths, or notice periods count. Sentiment alone does not.
5. Decide in advance whether the memo is internal-only or public. Publishing changes the audience. Customers, contractors, and journalists will read every line.

Caveat: the full memo text is not publicly retrievable. Work only from the published excerpt and press quotes, and be skeptical of “full text” versions passed around online.

How Did Duolingo Use Generative AI to Build 148 Courses?

Duolingo used generative AI to both create and validate course content, which let its learning designers focus their expertise where it mattered most. The result was 148 new courses in about a year, compared with roughly 12 years for its first 100, per von Ahn via FOX 29.

How Duolingo handled it: Becker described the before state as “a small team” spending “years to build a single new course from scratch.” The after state is AI-driven creation plus AI-driven validation, with human expertise moved to the highest-impact work. Notice that she names two steps, create and validate. Expert time was redirected to judgment, not removed.

Public sources do not describe Duolingo's internal pipeline. We don't know which models, prompts, review gates, or tooling it used, and we won't guess. The transferable idea is the structure, not the stack.

Do it this month

1. Pick one repeatable content type. Good candidates are help-center articles, onboarding email sequences, in-app tooltips, or template libraries.
2. Make “create” and “validate” two separate steps. Validation should never be a quick glance at the end of drafting.
3. Define what expert review means and where it is mandatory. For example, any content that touches pricing, compliance, or data handling gets a human sign-off every time.
4. Ship a small batch, such as 20 items, then measure time-to-publish against your previous baseline.
5. Expand only after the validation catch rate holds steady across at least two batches. If reviewers keep finding new error types, your spec isn't ready to scale.

What Does an AI-First Company Strategy Look Like? Examples from Duolingo

In Duolingo's version, an AI-first strategy means redesigning how functions work around AI rather than adding AI to existing processes. The memo's own line sums it up: “Making minor tweaks to systems designed for humans won't get us there.”

How Duolingo handled it: four levers, each tied to the AI-first operating model.

1. Function-level initiatives. “Most functions will have specific initiatives to fundamentally change how they work.”
2. Hiring criteria. Candidates are assessed for AI experience.
3. Performance-review criteria. Employees' AI use is evaluated.
4. A headcount gate. Per Business Today's paraphrase, new roles are approved only when a team can't automate more of its work.

The levers reinforce one another. A function initiative without a review criterion is optional. A review criterion without training is a trap. A headcount gate without function initiatives just slows hiring. Duolingo's memo connects all four, which is why it reads as an operating model and not a press statement.

Do it this month

1. Audit three workflows per function. Label each one automate, augment, or keep human.
2. Assign one named owner per function initiative. A committee is not an owner.
3. Add one question to your headcount request form: “What part of this role's work could be automated, and what have you tried?”
4. Review results in 30 days. Kill labels that turned out wrong and promote the ones that worked.

Should You Replace Contractors with AI for Content Creation?

Duolingo chose to “gradually stop using contractors to do work that AI can handle,” and the operative words are “gradually” and “can handle.” Together they describe a phased, task-by-task contractor-to-AI transition, not a single cutover. Teams that copy the policy often drop those qualifiers and lose the nuance.

How Duolingo handled it: the shift was already under way. Duolingo cut about 10% of its contract workforce the year before for similar reasons (WPXI). It has not disclosed its total contractor count, so the real scale of the change can't be measured from public sources. Anyone claiming to know how many contractors Duolingo replaced is guessing.

Do it this month

1. List contractor tasks by volume and repeatability. High-volume, low-variation work such as reformatting, translation first drafts, or tagging is the usual starting point.
2. Pilot AI on the single most repeatable task while keeping contractors on review. Their expertise is your validation layer during the pilot.
3. Track error rates and rework hours for AI output against the contractor baseline.
4. Decide based on evidence, and communicate timelines early. Contractors who hear about changes through a public post have every reason to talk about it publicly.

Ethics and reputation note: a contractor transition is a people decision as well as a cost decision. Give notice periods, offer reference letters, and consider paid review work during the handover. These choices shape how your company is talked about long after the savings show up in the budget.

How Can You Scale Content Production with Generative AI Without Scaling Headcount?

Use generative AI for both creation and validation, and move human experts to the highest-impact review and design work. This is the model Duolingo's learning design lead described, and it is the most transferable part of the story.

How Duolingo handled it: scale was the stated goal. Von Ahn: “Without AI, it would take us decades to scale our content to more learners.” The 148 courses are the result of treating content production as a throughput problem with a quality constraint, not a staffing problem.

Do it this month

1. Build a prompt and style spec from your best existing content. Pull your ten highest-rated articles or lessons and write down what makes them good: structure, length, tone, terminology, examples.
2. Add an automated validation checklist covering factual accuracy against source docs, tone, completeness, and broken references.
3. Route only flagged items to experts. This is how human-in-the-loop review scales. Experts see the 15% that needs judgment, not the 85% that passes.
4. Set a throughput target and a quality floor together. For example, “40 articles per week, with under 2% needing post-publish correction.” A throughput target alone guarantees quality drift.
5. Collect user signals on new content from day one. Internal validation catches what your checklist anticipates. Users catch everything else. More on this below.

The quality floor is the part most teams skip. Without it, “occasional small hits on quality” turns into a steady pile of them.

What Are the Lessons from the AI-First Memo Backlash?

The memo drew widely reported public backlash, and the lesson for teams is that customers and the public read an AI-first announcement too, not only employees. Framing matters as much as policy.

Caveat: we don't cover the backlash here with sourced specifics, so we won't cite numbers, platforms, or quotes about it. What we can do is point to tensions visible in the memo's own sourced language:

• “Take occasional small hits on quality” sits next to a product that learners pay for and use daily.
• “Stop using contractors” sits next to “cares deeply about its employees.”

Each line makes sense on its own and was candid internally. Read together by a contractor, a paying learner, or a reporter, they can land differently than the author intended. A memo that is internal in tone but publicly posted has several audiences at once, and each will focus on a different sentence.

Do it this month

1. Red-team your memo. Have three people outside leadership read it as a customer, a contractor, and a journalist. Ask each to underline the sentence they'd quote.
2. Lead with user benefit and quality protection. Explain what customers get and how you'll prevent bad outputs before you describe internal efficiency.
3. Say how you'll measure quality impact, and commit to a review date.
4. Prepare follow-up FAQs before publishing. The questions are predictable: jobs, quality, data use, and timelines.

How Should AI Factor into Hiring and Performance Reviews?

Duolingo said it would assess candidates for AI experience and evaluate employees' AI use in performance reviews, which makes AI fluency an explicit job expectation.

How Duolingo handled it: these criteria sit in the memo alongside the headcount gate (Business Today's paraphrase: headcount is approved only if a team can't automate further). The three policies point the same way. You hire for AI skill, you review on it, and you add people only after automation runs out.

Do it this month

1. Define good AI use per role, based on outcomes. For a PM, that might be “cut PRD drafting time in half without losing customer evidence.” For a support lead, “raised first-response accuracy.”
2. Add one practical AI exercise to interviews. Give candidates a messy input, like a call transcript or ticket dump, and ask for a usable output. Watch how they check the AI's work.
3. Add an AI-workflow reflection to review templates: “Which workflow did you change with AI this cycle, and what was the measurable result?”
4. Offer training before you evaluate. Reviewing people on a skill you never helped them build damages trust.
5. Don't measure raw usage volume. Prompt counts and login streaks reward activity over results, and people will game them within a quarter.

What to Copy and What Not to Copy from Duolingo

The most reusable parts of Duolingo's AI-first strategy are its written rules and its separate create-and-validate structure. The least reusable parts depend on Duolingo's scale and brand.

Copy

• Explicit written rules instead of aspirational slogans.
• Separate create and validate steps.
• Moving experts to high-impact review and design work.
• Stating your speed-vs-quality tolerance openly, at least internally.

Copy carefully

• Headcount gating and contractor transitions. Both depend on trust and clear communication. Without those, they read as cost-cutting dressed up as strategy.

Don't copy blindly

• Scale. Duolingo serves a very large consumer audience. Your B2B product may have 400 accounts, each worth a lot.
• Content type. Structured language lessons are well suited to templated generation. Enterprise configuration guides or clinical workflows are not.
• Brand position. A consumer brand with a playful voice can absorb some public criticism. A B2B vendor selling into regulated industries may not be able to.

What's self-reported and what's unknown

• The 12-years vs. about-one-year comparison and the 148-course count come from Duolingo leadership via FOX coverage.
• Duolingo's investor release was not accessible to us (it returned a 403 error).
• Contractor totals are undisclosed.
• There is no public data on learner outcomes or quality metrics for the 148 courses, so the “small hits on quality” can't be independently assessed from these sources.

Where Customer Evidence Fits: Checking the “Small Hits on Quality”

When you can generate content and features at scale, customer feedback is the only reliable check on whether “occasional small hits on quality” are actually reaching users. That feedback lives in support tickets, reviews, sales and success calls, and survey comments.

The faster you ship, the faster quality issues pile up. Internal validation catches the problems you anticipated. Users find the rest: a help article that confidently describes a deprecated setting, an onboarding step that skips a permission, a template with the wrong field name. Each one is small. Spread across hundreds of AI-generated items, they add up to churn risk.

To be clear, this section is our recommendation for your team. It was not part of Duolingo's public story.

A practical feedback loop

1. Tag feedback by the content or feature it references. “The SSO guide is wrong” should link to the SSO guide, not sit in a generic “docs” bucket.
2. Compare complaint rates on AI-generated vs. human-built items. If AI-generated content draws twice the confusion tickets, your validation checklist has a gap.
3. Route recurring issues back into the validation checklist. Every repeated user complaint should become a new automated check.
4. Close the loop with the customers who reported it. Fixing quietly misses the chance to rebuild trust.

BuildBetter is built for this loop. It brings calls, Slack threads, support tickets, and surveys into one place, so product teams can see whether AI-shipped content is causing friction. Then it turns those findings into Jira or Linear tickets and PRDs instead of a chart nobody opens.

Tools to Monitor Quality When You Ship AI-Generated Content at Scale

The right quality-monitoring tool depends on where your customer evidence lives and what you need to do with it. Here are four options, ranked for teams shipping AI-generated content at volume.

1. BuildBetter — Best for tying customer evidence to AI-shipped content

BuildBetter combines internal voice (calls, Slack) with external feedback (Zendesk, Intercom, HubSpot, surveys) across 100+ integrations. It analyzes each piece of feedback individually for severity and business impact, applies your product taxonomy, and generates PRDs, tickets, and loop-closure emails. For a team testing whether AI-generated help content is increasing confusion, it can show which accounts hit the problem and what that means for revenue. Honest caveat: it isn't built for enterprise-scale survey distribution or dedicated research-repository workflows like highlight reels.

2. Looppanel

A research-call analysis tool for smaller UX teams. It's useful when your quality check is a handful of moderated sessions on new content.

3. Usersnap

Captures in-app feedback and bugs on specific screens or pieces of content. It's a good fit when you want users to flag a broken tooltip right where it appears.

4. Zonka Feedback

Lightweight in-product and post-interaction surveys. Use it to add a “Was this helpful?” prompt to newly generated content.

When another tool is the better choice: if you only need a thumbs-up/down widget on docs, a survey tool is enough. If your team runs a formal research practice with clip libraries, choose a research-focused tool. A unified platform pays off when feedback is scattered across calls, tickets, and Slack and you need it turned into shipped fixes.

FAQ: Duolingo's AI-First Strategy

What did Duolingo's AI-first memo say?

In late April 2025, CEO Luis von Ahn told staff “we are going to be AI-first” and framed AI as the next platform shift. The memo said Duolingo would gradually stop using contractors for work AI can handle, assess candidates for AI experience, evaluate employees' AI use in performance reviews, and launch function-level initiatives to change how work is done. Business Today also reported that new headcount would be approved only if a team could not automate further.

Where can I read the full memo?

Duolingo posted the email on its LinkedIn page, but we could not reliably retrieve the full text. Quotes in this guide come from the LinkedIn excerpt and from coverage by Business Today, WPXI / Pittsburgh Business Times, and FOX Business via FOX 29.

How many courses did Duolingo build with generative AI?

148 new courses in about a year, compared with about 12 years for its first 100, according to von Ahn via FOX coverage. Duolingo reported these figures itself, and independent quality data for the new courses is not public.

How did Duolingo use AI to build courses so quickly?

According to Senior Director of Learning Design Jessie Becker, Duolingo used generative AI to both create and validate content, which let its experts focus “where it's most impactful.” Public sources do not describe the specific models, tools, or workflow.

Did Duolingo replace contractors with AI?

The memo committed to “gradually” stop using contractors for work AI can handle. Duolingo had already cut about 10% of its contract workforce the year before for similar reasons. It has not disclosed how many contractors it employs, so the scale of the change can't be measured publicly.

Does Duolingo evaluate employees on AI use?

Yes. The memo says employees' AI use will be evaluated in performance reviews and candidates will be assessed for AI experience.

What quality tradeoff did Duolingo accept?

In the memo's words: “We'd rather move with urgency and take occasional small hits on quality than move slowly and miss the moment.”

How can a smaller team monitor AI content quality?

Pair internal validation with customer feedback signals. BuildBetter brings tickets, calls, Slack, and surveys together to surface issues tied to specific content, then turns them into tickets or PRDs your team can act on.

Sources and Further Reading

• Duolingo LinkedIn: all-hands email from Luis von Ahn
• WPXI / Pittsburgh Business Times: Duolingo CEO says company will replace contract roles with AI
• Business Today: Duolingo goes all in on AI, phases out contractors
• FOX 29 / FOX Business: Duolingo launches 148 courses built with AI
• Background: Duolingo (Wikipedia)
• Background: Luis von Ahn (Wikipedia)

Ship Faster Without Losing Track of Quality

Going AI-first raises your output. It also raises the number of places quality can slip without anyone noticing. BuildBetter gives product teams one view of every call, ticket, Slack thread, and survey, connects that evidence to what you shipped, and turns problems into PRDs and tickets. When the fix ships, it tells the customers who asked.

Make churn optional. Book a demo.