Field lessons from operators and investors, distilled in our own words from the free public archive of Lenny's Podcast and Lenny's Newsletter (starter dataset). These are practitioner experience, not peer-reviewed research — for the evidence base on what actually works in learning, see data/research/.
Each lesson notes who it came from. Where a lesson links out, the link goes to that episode or post page on Lenny's site (lennysnewsletter.com). Most podcast episodes have no stable URL in the source data, so for those, search the guest's name on lennyspodcast.com to find the episode.
Count the workaround, don't survey the concept. Chesky sold Rabois on Airbnb with ~30 Craigslist posts where people typed out that they wanted to rent a stranger's room; DoorDash, with "93% of US restaurants don't deliver." A count of people already doing the painful workaround beats any volume of "would you use this." Look for who's already hacking a fix for your gap, or who has an open RFP for it. — Keith Rabois, Lenny's Podcast
Buyer interviews are trustworthy when one person decides. Customer development works where you can name a single utilitarian decision-maker — which is exactly the institutional buying setup. Treat your district CTO or VP of L&D conversations as real signal. "I surveyed eight teachers" is noise: free-teacher enthusiasm is not the budget-holder's decision. — Keith Rabois, Lenny's Podcast
Watch organic compounding, not the launch spike. Winning Product Hunt's day, week, and month told Gamma nothing. The signal that mattered was whether signups kept growing on their own after the spike faded; when they flattened, the team called it no fit despite the trophies. Friendly pilot users also quietly stop at day 30, 60, 90 unless the product is genuinely useful. — Grant Lee, Lenny's Podcast
A dead reply rate is a premise problem, not a tooling problem. Two companies ran identical outreach and got 2% versus 12% interest; the only difference was the sharpness of the founder's insight into the problem. If cold outreach to principals or L&D leads flops, suspect your premise before blaming the email tool. "Better than your current LMS" gets ignored; a felt pain like a compliance deadline or teacher time gets replies. — Jen Abel, Lenny's Podcast
Chase the buyer who pulls it out of your hands. Hold a strong thesis about how the world is changing, but stay loose on the exact wedge and go after the customer who is shockingly easy to sell to. If selling the next customer is a grind, the business won't scale — so trust the eager buyer over the district you spend nine months persuading that your pedagogically "correct" product is right. — Brendan Foody, Lenny's Podcast
You can't research your way to a zero-to-one product. People answer from what they've seen — ask about a touchscreen phone and they describe a better keyboard. Feedback loops are right for improving an existing category and actively mislead on category-defining bets. For an incremental tool, lean on educator feedback; for a new bet, expect teachers to ask for the familiar workflow and validate by putting a working prototype in front of them. — Caitlin Kalinowski, Lenny's Podcast
Your next product line is in the off-label requests. Canva began when one happy yearbook customer asked "can I also use this for newsletters?" Perkins checked that nothing on the market solved it and expanded there, rather than designing a wedge strategy from scratch. Mine the "can I also use it for…" asks from your most engaged school before inventing an expansion roadmap. — Melanie Perkins, Lenny's Podcast
Reconstruct the buyer's economics yourself. Restaurants wouldn't share real margins, so Uber Eats ordered food and weighed the ingredients against a supplier catalog to rebuild the cost structure — which gave them conviction to price boldly. Budget owners and procurement won't hand you real numbers either, so reconstruct district budget math independently before you price. — Jason Droege, Lenny's Podcast
Make the first 30 seconds magical, and treat onboarding as the product. Gamma's frame: what can you give a selfish, lazy, impatient new user in 30 seconds that earns you the next 30? That forces you to surface the single most valuable action instead of a feature tour. Teachers and students abandon tools that demand setup before value, so engineer one undeniable aha — a generated lesson, a graded draft — in the first session. — Grant Lee, Lenny's Podcast
For anything novel, the obstacle is comprehension, not friction. Removing clicks only helps when the buyer already knows what they want; for everything new, people arrive barely over the intent threshold and bounce because they don't understand what it is or what happens next. Educators are not power users and feel stupid when confused, so your homepage and onboarding have to build understanding — not just shorten the signup form. — Stewart Butterfield, Lenny's Podcast
Listen for the problem, then invent the solution. When users demanded a literal "send to all" button, Snap dug into why — social pressure, permanence, reverse-chron feeds — and shipped Stories, solving the real need without building the requested feature. Teachers and admins will ask for another report or another field; the win is mining the underlying job and shipping something they didn't know to request. — Evan Spiegel, Lenny's Podcast
Ship rough and learn from real users — you can't spec a non-deterministic product. You can't fully mock an AI product before shipping, so release early as a "research preview" and learn the real use cases and safety failures that lab evals miss. Pair it with fixing feedback within minutes, which makes people feel heard and drives a flood of more feedback. A fast pilot turns early schools into co-designers and shows where the tool produces wrong or unsafe output in front of students. — Boris Cherny, Lenny's Podcast
Demo, don't memo. Shopify hasn't let PMs pitch ideas as slides for two years; a four-hour prototype communicated more than a week of documents. With AI prototyping nearly free, build several versions and pick from the real thing. Walking into a district meeting or investor pitch with a working demo of the exact teacher or student workflow beats any roadmap slide — and forces honesty about whether the idea works. — Keith Rabois, Lenny's Podcast
Good-enough is no longer a differentiator — ship a minimum lovable product. AI lets everyone produce good-enough fast; the gap that pays now is the distance to world-class (taste, design, copy, emotional resonance). Make a designer an early hire. A generic AI-wrapper tool is trivially cloneable, so your edge with skeptical educators is craft: a tool that feels good gets evangelized in the staff room while a clunky-but-functional one gets abandoned. — Elena Verna, Lenny's Podcast
Match the technology to the job; don't force the LLM everywhere. Chess.com scores every move with a deterministic chess engine (LLMs are bad at chess) and uses the LLM only for the human-friendly explanation. For grading, adaptivity, or domain logic, a specialized or rules-based engine often beats an LLM — reserve the model for the tutoring and explanation layer, which also controls cost and hallucination risk. — Albert Cheng, Lenny's Podcast
"Thin wrapper on a model" is the wrong dismissal. Horowitz maps it to the 1980s "thin wrapper on a database" jab that Salesforce disproved. Cursor's moat is proprietary data and domain depth — internal models trained on real high-end developer interactions — not the base model. Because human behavior is fat-tailed, capturing the messy real-world edge cases of teaching, student responses, and how learners actually behave is where defensibility lives. — Ben Horowitz, Lenny's Podcast
As generation gets cheap, design the review surface. The bottleneck moves from making things to checking them — usually the least fun part. Show the human a rendered preview or an AI pre-review first, not raw output, so they verify in the fastest way and stay accelerated. If your tool drafts lesson plans, IEPs, or feedback at scale, the teacher's review burden is the real adoption gate — design it as carefully as the generation, or volume buries them and kills usage. — Alexander Embiricos, Lenny's Podcast
Before copying a feature that works elsewhere, ask why it works there. Duolingo borrowed a moves-counter from a match-3 game and it flopped, because that game's moves require strategy while answering a lesson question doesn't — so the counter was just an annoyance. Always ask why it works in the source product, whether that translates, and what you must adapt before cloning a streak, a chat UI, or a consumer growth loop into a classroom. — Jorge Mazal, Lenny's Newsletter
Start AI quality work with error analysis, not tests. Sample ~100 real production traces and write one free-form note per trace about the first thing that went wrong; stop when notes stop appearing, then have an LLM cluster and count them to find your most common failure. For LLM-as-judge, force a binary pass/fail per single failure mode (never a 1–5 score) and validate it against human labels with a confusion matrix — a judge that always says "pass" scores 90% when errors are only 10% of cases. Reading real student-AI transcripts surfaces failures (hallucinated content, missed handoffs) you'd never spec upfront. — Hamel Husain & Shreya Shankar, Lenny's Podcast
Design around the "lethal trifecta." Any agent that combines private-data access, exposure to untrusted input, and an outbound channel can be tricked into exfiltrating data — and prompt-injection filters top out around 97%, a failing grade. The fix isn't better guardrails; it's cutting one leg, usually the ability to send data out. An AI tutor touching FERPA records that also reads untrusted content and can email or post is a breach waiting to happen, so architect the exfiltration path away rather than trusting a prompt to behave. — Simon Willison, Lenny's Podcast
Treat "agent" as a dial, not a label. Build along an agency ladder: V1 only suggests to staff, V2 drafts a reply the human edits, V3 acts directly — and log what the human does at each stage as free training data, raising autonomy only once the data shows reliability. School buyers are nervous about AI acting unsupervised on student data or grades, so a deliberately less-agentic design is both safer and often the faster path to a signed pilot. — Aishwarya Naresh Reganti & Kiriti Badam, Lenny's Podcast
When the agent fails, it's usually context, not capability. Base-model intelligence is largely already there; the real product work is encoding the idiosyncrasies of a specific workflow — which logs to check, which steps to run, how to handle each failure — into structured context the model can see, rather than reaching for a bigger model. Your moat isn't a smarter LLM; it's encoding how a district's IEP process, your state's standards, and a course's rubrics actually work, which generic ChatGPT can't replicate and which improves with each deployment. — Scott Wu, Lenny's Podcast
Aim AI at processes that are currently bad, not ones that are already near-perfect. AI delivers huge value where the human process is only 10–20% accurate (getting to 60–80% is a celebrated win) and struggles closing the last 2% on a 98% process. And an automation that works 95% of the time isn't an automation — you can't rely on it until it's near-100%, so always design an escalation path to a human. Point it at feedback nobody has time to write or triaging which students need help, not tasks teachers already do well. — Jason Droege, Lenny's Podcast
The models eat your scaffolding for breakfast. Heavy infrastructure built to compensate for weak models — custom agent frameworks, mandatory vector-store RAG, forced to-do lists — becomes dead weight as models improve, and hand-built orchestration adds maybe 10–20% that the next release erases. Build for the capability that's ~80% there today, give the model tools and a goal rather than rigid step-1/step-2 workflows, and revisit your prompt scaffolding each model release. Architect a borderline-today district pilot for the accuracy you expect in 6–12 months. — Sherwin Wu, Lenny's Podcast
Own the data or be the model — the middle gets crushed. Point solutions lack the first-party data to do useful joins, and integrating via flat-file feeds is drinking through a straw. Defensible AI products either own the data ("the mine") or provide the model ("the shovels"); renting both crushes your unit economics. A single-feature tool pulling SIS/LMS/assessment data through brittle integrations won't be allowed good economics by the platforms it depends on — owning a data wedge and accumulating each teacher's students, materials, and corrections is the moat. — Matt MacInnis, Lenny's Podcast
Distrust one-click agents and multi-agent "swarms." Replacing a real workflow takes roughly four to six months even with good data, because enterprise data is messy — duplicate functions, broken taxonomies, undocumented rules. Current models can't reliably run a utopia of peer agents that self-coordinate; what works is one supervisor agent (or a human) orchestrating sub-agents. School SIS, LMS, and roster data are exactly this messy, so budget months, not weeks — and don't architect a tutoring product as a swarm of specialists talking to each other when a student or parent is on the other end. — Aishwarya Naresh Reganti & Kiriti Badam, Lenny's Podcast
For RAG, the wins are in data prep, not the vector database. Tune chunk size, rewrite source content into question-answer pairs, and add an annotation layer for things AI lacks common sense about. What improves AI apps is talking to users, preparing better data, fixing the end-to-end workflow, and better prompts; what wastes time is chasing AI news, swapping frameworks, agonizing over vector DBs, and fine-tuning. If you're building a tutor over curriculum or training docs, restructure that content into Q&A form before touching infrastructure — docs written for humans often fail for AI. — Chip Huyen, Lenny's Podcast
Onboard an agent like an employee, and build the undo. An agent can make hundreds of changes in seconds, so the UX flips: you need approval queues, a summarized inbox of what happened, change logs, and one-click rollback — not a real-time collaboration view. Give it its own scoped credentials (never your master password), widen permissions as trust is earned, and hard-code that it takes instructions only from the user on one channel, treating email and the open web as data, not commands. A least-privilege, stated-trust-boundary design is also far easier to get through district IT and security questionnaires. — Dan Shipper, Lenny's Podcast
Sell a paid service before you sell the technology. 40–50% of B2B startups must sell a time-boxed (90-day) paid service first, because the buyer has no process or human-in-the-loop plan to adopt something new yet. The service gets you the logo, revenue that signals real intent, and the position of educating the buyer. If a district has no workflow for adopting AI tutoring, sell a paid 90-day implementation first rather than waiting 18 months — whoever educates the buyer wins the eventual contract. — Jen Abel, Lenny's Podcast
Shrink the change to de-risk the yes. When a commitment feels scary, repackage it as a one-week proof of concept with explicit success criteria and a pre-set check-in date. It's the certainty of a next checkpoint, not certainty of the outcome, that calms a cautious decision-maker. Risk-averse school and university buyers say yes faster to a tightly scoped, time-boxed pilot with defined metrics and a scheduled review than to anything open-ended. — Jessica Fain, Lenny's Podcast
Pick at least three metrics across different dimensions. Teams default to easy-to-instrument activity counts that look productive but say nothing about whether the work is good or sustainable — which is why SPACE forces metrics across satisfaction, performance, activity, collaboration, and flow. A pilot reported purely on usage will get gamed and won't convince a skeptical district; pair a usage metric with a learning-outcome metric and a teacher-satisfaction survey. — Nicole Forsgren, Lenny's Podcast
A demo is not a deployable system — budget for the nines. Getting an AI from a 60–70% demo to production reliability follows a data-center-uptime curve: each additional "nine" is an order of magnitude more work, and serious automation takes 6–12 months of engineering, legal, and change management. Cheap proofs-of-concept inflate the failure-rate headlines. Set district expectations accordingly so you don't get labeled a failed pilot when the demo doesn't survive contact with real classrooms. — Jason Droege, Lenny's Podcast
Distrust flashy leaderboard "wins." Quick crowd-vote leaderboards reward flashy, emoji-laden, longer outputs even when the answer is wrong, because people skim for two seconds. Real evaluation needs domain experts working through the actual task — checking the code, the equations, the reasoning. An AI tutor that sounds confident and looks polished can score well while being pedagogically wrong, so prove learning outcomes with educators rigorously checking accuracy, not vibe-based demo reactions. — Edwin Chen, Lenny's Podcast
Give a pilot one accountable owner. Anything cross-functional dies when it's split across sales, implementation, and product — "if you want to kill a plant, have two people water it." Name one Directly Responsible Individual with authority to direct other teams. A district pilot touches sales, onboarding, support, and product, so a single accountable owner keeps a make-or-break reference deployment from falling through the cracks. — Brian Halligan, Lenny's Podcast
Keep word of mouth above half, or you're on a CAC treadmill. Gamma keeps over 50% of signups coming from word of mouth and treats paid acquisition as a hard ceiling; if more than half your growth is ads, the core engine is broken. Mercor reached a nine-figure run rate with nobody in sales or marketing. Prove that teachers refer colleagues and schools refer schools before pouring money into ads or hiring SDRs. — Grant Lee, Lenny's Podcast
Lead with a use case the buyer already measures. Companies buy AI use cases with measurable outcomes far more readily than fuzzy productivity ones — a sales chatbot sells because you can compare conversion before and after, while internal-knowledge tools stall. Lead with a metric districts and L&D already track (completion, support-ticket deflection, time-to-onboard), not vague "engagement" or "productivity," because clear outcomes unlock budget. — Chip Huyen, Lenny's Podcast
Earn "permission to play" before you build. Before a new feature or product line, ask whether buyers will find it credible that your company offers it and whether you already have a route to reach them; Jeetu Patel kills ~99% of new-idea proposals on this test. A literacy-platform founder has permission to launch an adjacent assessment tool to the same districts — but an unrelated product to those schools wastes scarce calories, because the relationship and category credibility aren't there. — Jeetu Patel, Lenny's Podcast
Answer Engine Optimization is a channel you can win this week. Unlike SEO, which needs years of domain authority, you can show up in a ChatGPT or Perplexity answer tomorrow via a Reddit thread, a YouTube video, or a single blog mention — and being cited most often across sources beats being the #1 link. Teachers, instructional coaches, and L&D buyers increasingly ask AI "what's the best tool for X," so a young product can get into those answers now by seeding authentic mentions. — Ethan Smith, Lenny's Podcast
Tell pilots to throw their hardest real problem at it. A professional-grade tool earns trust by solving the gnarly problem nobody else could; dumbing the trial down both undersells it and hides where it breaks. In a district pilot, ask the skeptical department head for their messiest real case — the impossible-to-grade essays, the tangled scheduling conflict — because succeeding there converts a doubter faster than ten polished demos. — Alexander Embiricos, Lenny's Podcast
Channels sag before they die — assume decay. Channels don't plateau on a clean S-curve; they're an "elephant curve" where audiences saturate and decline while vendors report healthy numbers right up to failure. Constant Contact restarted growth with in-person workshops; HubSpot built an agency channel that became ~50% of revenue. Durable edtech distribution often comes from ecosystem partners — district PD workshops, university teaching centers, reseller networks — that start you with an existing audience. — Jason Cohen, Lenny's Podcast
Your funnel breaks at qualification, not closing. Abel has never seen a genuine bottom-of-funnel problem; deals stall because you reached the wrong person, used the wrong message, or pitched a problem you can't solve. Book the next call during the current call, and treat "I'll email you" as a soft no. When school deals fizzle after warm first meetings, the cause is almost always upstream — you're talking to a teacher who can't buy. — Jen Abel, Lenny's Podcast
Do all the procurement work for them, and pre-arm the signer. Fill out their forms, and state precisely what you do and don't do — vague claims get you classified high-risk and routed to the kitchen-sink MSA. Abel lost a month because a CFO got a contract he didn't understand and kicked it to the back of the queue. District and university procurement, security review, and privacy vetting stall unless you project-manage every form and hand the actual signer — CFO, CISO, department head — a clear one-liner. — Jen Abel, Lenny's Podcast
Sell against the risk of doing nothing. Roughly four in five buyers purchase to avoid pain or reduce risk, not to chase upside, so the founder-vision pitch mostly lands only with other founders. A district CIO or VP of L&D is protecting a budget and a career — lead with the cost of falling behind, audit or compliance exposure, or a renewal that won't show outcomes, not the visionary future state. — Jeanne DeWitt Grosser, Lenny's Podcast
Hand-build 30 prospects before you buy sales tooling. Manually find 30 ideal prospects and spend 15–20 minutes writing each a real note across email, LinkedIn, and a call. The exercise reveals whether your buyers are even discoverable and what they share — only then do enrichment tools work. If you can't find 30 curriculum directors or L&D managers, or you get zero replies, you've learned something cheap before automating outreach and torching your domain. — Jen Abel, Lenny's Podcast
First-rep rules: close ~10 deals yourself, hire two, pick who you'd buy from. Personally close around 10 deals before hiring; hire two reps, not one, so you can compare; pick the rep you'd actually buy your own product from over the impressive logo; and favor reps whose last product was slightly harder to sell than yours. Wait on a VP of sales until two reps hit quota. A rep who fought procurement-heavy, security-paranoid sectors finds edtech easier than one from a frictionless SMB product. — Jason Lemkin, Lenny's Podcast
Make the first call a working session that leaves an asset. Stripe's first sales call was a whiteboard of the customer's own payments architecture, so the buyer walked away with a diagram they'd never drawn — helped, not quizzed. Run your first school meeting as a co-mapping of their student journey, data flow, or current tool stack, so they leave with something useful even if they don't buy. That's the trust that wins a slow committee sale. — Jeanne DeWitt Grosser, Lenny's Podcast
Ask spicier discovery questions. "What's top of mind?" has gone generic and yields rehearsed answers. Fain asks emotion-surfacing questions instead: "What's the most urgent thing you're scared of messing up?" and "What is your board pushing you on?" With a district or L&D buyer, probe what their board, superintendent, or state accountability metrics are pressuring them on this year, then tie your tool to relieving that specific fear. — Jessica Fain, Lenny's Podcast
Raising the price often raises signups in B2B. Price signals quality and selects which buyers even consider you. One founder selling to enterprise and government at $300/year saw zero change after switching to $300/month — 12x — proof he was nowhere near the ceiling. A cheap per-teacher price screens you out of district procurement, where buyers equate low price with weak security, support, and governance. — Jason Cohen, Lenny's Podcast
Reposition from savings to growth and charge far more. Pitched as "cut your AdWords cost in half," a tool is worth ~$5K/month because the buyer keeps most of the savings; pitched as "double your leads at the same ROI," the identical tool is worth ~$40K, because growth budget dwarfs cost-cutting budget. Sell "raise completion, retention, or enrollment," not "save teachers time" — outcomes that leadership is measured on carry an order-of-magnitude bigger budget than efficiency claims. — Jason Cohen, Lenny's Podcast
Make the free tier a live sample of the full product. Grammarly's free users only saw spelling and grammar fixes, so they assumed that was all it did; interspersing a capped real-time taste of paid suggestions nearly doubled upgrades — despite fears that giving more away would hurt conversion. If your free or pilot tier shows only the boring basics, schools and learners will price you as a basic tool. — Albert Cheng, Lenny's Podcast
Heavy spenders have more intense needs, not deeper pockets. Tinder's biggest à-la-carte spenders weren't wealthy flaunters but people with urgent needs — military, frequent movers, new to a city — who priced the app against the cost of dating, not other subscriptions. Find the buyers with the most acute pain (a district under a compliance deadline, a school with a specific outcome gap) and anchor pricing to the cost of their alternative, not to competitor subscription prices. — Ravi Mehta, Lenny's Podcast
Don't set a price until users are begging, then verify margins. Gamma waited until users wanted to pay, then used a Van Westendorp willingness-to-pay survey plus conjoint to land on one simple plan around $20/month, anchored to ChatGPT — and checked it produced positive margins on inference rather than assuming they'd figure out economics later. For a prosumer teacher tier, run a quick willingness-to-pay survey, keep the plan dead simple, and confirm the price covers AI cost from day one, since education budgets won't tolerate later hikes. — Grant Lee, Lenny's Podcast
Know your growth ceiling: new customers per month divided by monthly churn. At 100 new logos a month and 5% churn you will never pass ~2,000 customers, because cancellations scale with your base while marketing doesn't. Run this number before chasing more acquisition. Districts and universities churn on a budget cycle, so a bad churn rate quietly caps you no matter how hard your reps push. — Jason Cohen, Lenny's Podcast
Model your users as states, and find the one lever that compounds. Duolingo bucketed every user into mutually exclusive states (new, current, reactivated, resurrected, at-risk, dormant) and modeled the transition rates between them. A sensitivity analysis showed current-user retention drove 5x more daily-active growth than the next-best lever, because retained users loop back and compound. It's a concrete way to find the one retention lever that actually moves the active-usage number districts judge renewals on. — Jorge Mazal, Lenny's Newsletter
"Too expensive" is almost never the real reason they canceled. A customer who got through your whole funnel and paid had already accepted the price. Ask "what made you cancel?" as an open question (it roughly doubles usable responses), then dig past the excuse to the real failure. School buyers will blame budget when the truth is your tool didn't sync with their SIS or a teacher couldn't onboard students, and fixing the stated reason wastes a renewal cycle. — Jason Cohen, Lenny's Podcast
Leads from AI answers convert far better, and you probably can't see them. Webflow saw a 6x conversion gap between ChatGPT-referred traffic and Google, because the buyer already had a long qualifying conversation before clicking. Most of it is mis-attributed as "direct" or "branded search" because people open a new tab and type your name, so add a post-signup "how did you hear about us?" If your dashboard shows direct or branded traffic rising, AI may already be sending you high-intent district and L&D buyers you can't see. — Ethan Smith, Lenny's Podcast
Reframe failure instead of rubbing it in. Chess.com found 80% of users review a game after a win, not a loss, so they flipped the post-loss screen to surface best moves and encouragement instead of blunders, and grew the feature 25% and subscriptions 20%. Learners abandon tools that punish their mistakes, so reframing errors positively can lift retention and paid conversion across any K-12 or L&D product. — Albert Cheng, Lenny's Podcast
Keep the friction that helps a user see the product is for them. Don't reflexively strip onboarding steps. Anthropic, MasterClass, Mercury, and Calm all keep deliberately long onboarding quizzes, because asking who the user is lets you route them to the right feature and personalize later. Cut friction that adds nothing; keep friction that earns a better first experience, and reuse the profile data for re-engagement. Activation in edtech hinges on a teacher or learner seeing "this is for my subject, grade, or role." — Amol Avasare, Lenny's Podcast
Protect your notification and email channels as a hard constraint. Groupon escalated to five emails a day, briefly won on metrics, then permanently lost the channel, because opted-out users never come back. Duolingo let the team optimize timing, copy, and images freely but required CEO approval to raise frequency. Edtech leans on reminder emails and push notifications to drive student and teacher engagement, so over-testing frequency can burn the exact channel that keeps a district's usage numbers up at renewal time. — Jorge Mazal, Lenny's Newsletter
Treat every "no" as a missing slide. When Canva heard "market's too small" or "you're like X," Perkins added a market-size slide and a competitive-gap slide to pre-answer those exact objections — so later investors grasped in 10 minutes what the first took six hours to get. The vision stayed constant; only the articulation sharpened through rejection. Edtech founders face investors who don't understand the sector, so log each objection and add the slide that pre-empts it. — Melanie Perkins, Lenny's Podcast
Name your accumulating advantage against the model labs. Investors increasingly ask whether foundation-model labs will leave "no oxygen" for you over an 8–20 year horizon. You don't need a proven moat at seed, but you must articulate, in sequence, where the compounding advantage — data, network, switching costs — will come from and when you'd start measuring it. A thin AI wrapper on a teacher tool reads as model-fodder; show how district data, integrations, or workflow lock-in compound. — Keith Rabois, Lenny's Podcast
Lock in mission-protective structure before you raise. A Public Benefit Corporation filing is a roughly two-page Delaware filing with essentially no downside, and the only time to do it is before priced rounds, because leverage only shrinks afterward — only ~20% of founders are still CEO three years after IPO. "It's always too early until it's too late." A founder who wants control over what happens to student data and learning outcomes should lock this in at the SAFE stage, not discover at acquisition that the board is bound to take the highest bid. — Eric Ries, Lenny's Podcast
"Never quit" serves VC incentives, not yours. Seed investors model every bet to zero and want you trying against all odds, because that's the only way they get their money back. If you're in year four or five, still pivoting without rip-roaring growth, quitting and resetting the cap table is often the rational move — product-market fit is unmistakable when it's real, and if you're unsure, you don't have it. Edtech's long sales cycles disguise the absence of demand, so founders rationalize "almost there" for years. — Matt MacInnis, Lenny's Podcast
Count your "barrels," not your headcount. Barrels are the rare people who own an initiative end-to-end and get it over the hill; everyone else is ammunition that amplifies a barrel. PayPal had only 12–17 barrels among 254 people. Adding ammunition behind the same barrels adds coordination tax without throughput, so staff against your number of true owners rather than raising and over-hiring against a thin pilot or sales motion. — Keith Rabois, Lenny's Podcast
AI adoption needs top-down budget plus a bottom-up tiger team. Pure mandates fail; what works pairs exec buy-in with a small team of your most excited people who find real workflows, run hackathons, and evangelize — often technically adjacent non-engineers like the ops lead who's an Excel wizard, not coders. This is exactly how your district and university customers will or won't succeed, so build a customer-side champion team into your implementation plan rather than relying on an administrator's mandate that teachers quietly ignore. — Sherwin Wu, Lenny's Podcast
Use blind references and sharper hiring questions. Ask "On a scale of 1–10, how likely are you to rehire this person?" and "Were they in the top 1% or top 10% of your reports?" Founders overrate their own gut, and roughly half of senior hires are gone within 18 months. Shrink the interview panel (HubSpot went from 8 to 4) and pick spiky candidates with real strengths over safe 3-out-of-4 generalists. A mis-hire on a small team is slow and expensive to correct. — Brian Halligan, Lenny's Podcast
Agree on what the goal means before anyone builds. Forsgren's most common failure — 80% of teams — is starting work before defining the goal. "Improve developer experience" splinters into culture versus tooling versus friction, and teams spend months building the wrong thing; the fix costs one week of writing it down. When a district says it wants to "improve engagement" or "use AI," pin the exact definition in writing before scoping the pilot, or you'll ship something they didn't mean and lose the renewal. — Nicole Forsgren, Lenny's Podcast
Every useful AI agent needs a named human owner. Agents quietly stop being useful unless someone owns them and keeps adding context; individuals won't do the upkeep, so the working model is a single company "super agent" maintained by a forward-deployed-engineer type. If you sell an AI agent into a school, name the staff owner who maintains it during the pilot — an unowned agent degrades and your renewal dies. Build the owner role into your deployment plan. — Dan Shipper, Lenny's Podcast
Don't make the impressive big-company hire too early. A VP from Microsoft, Google, or Salesforce expects you to "have your act together," which a 50–500 person startup doesn't — Halligan saw near-100% attrition on those hires. The underrated alternative is promoting homegrown people who've already proven themselves and whose weaknesses you can see. A founder dazzled by a Pearson or Google-for-Education resume often gets a leader who can't operate without big-company infrastructure. — Brian Halligan, Lenny's Podcast
Separate identity from behavior before a hard conversation. Open by affirming the person is capable and on your team, then address the specific behavior. Skip this and people hear a verdict on their worth and get defensive, and you end up silently arguing about whether they're a good person instead of fixing the issue. It's the most reliable way to give a struggling teammate, co-founder, or a misfiring implementation lead at a customer site feedback that changes behavior. — Dr. Becky Kennedy, Lenny's Podcast
Hesitation is the most destructive thing a founder does. When both options look bad, avoiding the call freezes the company and pushes nervous senior people to fill the void, which turns political fast. The fix is rarely more analysis — often you already see the answer but don't know how to have the hard conversation, so make it specific and behavioral. Edtech's long sales and budget cycles punish a frozen team severely, so choosing beats stalling on whether to re-architect, cut a pilot, or replace a hire. — Ben Horowitz, Lenny's Podcast
Distilled from the free public starter dataset of Lenny's Podcast and Lenny's Newsletter (github.com/LennysNewsletter/lennys-newsletterpodcastdata), used under its personal/non-commercial terms. Takeaways are paraphrased, not quoted. Last updated 2026-05-29.