Should you build or buy software now that AI can build a tool in two hours?

By Andrej Ruckij · November 1, 2026

TL;DR: For small, single-purpose tools, building with AI can now be the better deal. Andrej Karpathy built a working YouTube alternative in about two hours, and nearly a third of companies in McKinsey’s 2026 survey have skipped at least one software purchase because they could build it with coding agents. When building takes an afternoon, what’s scarce is knowing what’s worth building and having the data to build it on. Keep buying when the tool costs under about $50 a month, touches customer or personal data you can’t secure yourself, or is best in class at messy edge cases. Build when it’s a $200-1,000 a month single-purpose tool that’s core to how you work. The bigger change, in my view, is personalisation: software and recommendations tailored to one team, one niche or one customer are now cheap to make, often from reviews, tickets and comments a business already has.

What did Karpathy actually build?

On 31 October 2026, Andrej Karpathy published HackerTube, a YouTube-style site made only of videos that Hacker News users have recommended. He started from a frustration with YouTube’s feed: “I’m sure this algorithm works well for the median person on the internet, but it definitely does not work for me.”

What he did, in his words and numbers:

Took the full archive of Hacker News comments from a public dataset on Hugging Face. The open-index/hacker-news dataset lists 43,062,301 comments, is licensed ODC-BY and refreshes every five minutes.
Had Claude Opus download and filter the comments, running Claude Sonnet over them in a Claude Code Workflow to drop spam and narrow “everything down to 40K comments recommending a YouTube video”. That became 6,454 videos across 150 topics, from 2007 to 2026.
Let Claude do the rest; in his words it would “update, deploy, secure, harden and optimize the code”. His own manual jobs were three: creating a YouTube API key in Google Cloud, booting a VPS and buying the domain. “This takes ~2 hours now.”
Added a “For You” feed that personalises without accounts (state lives in your browser) and an “export of the site’s data available to give to your agent to easily create your own version”.

He doesn’t say what it cost or how many people use it. It’s a two-hour build by one of the best-known engineers in AI, and it’s a hobby site.

Three things this changes

My read of it comes down to three points and one bigger one.

Building stopped being the cost. When a niche tool takes two hours, the scarce things are knowing what’s worth building and having the data. That flips build vs buy for internal tools and small SaaS seats.

Human curation, filtered by AI, can beat an engagement algorithm for some users. Karpathy didn’t train a recommender. He took recommendations real people had written and had a model pick them out. A business can do the same with what it already has: reviews, support tickets, customer comments, community posts.

The product can ship as data for other people’s agents. HackerTube’s export exists so you can hand it to your own agent and build your own version. HackerTube itself only exists because someone else published the Hacker News archive as an open dataset.

And overall: personalisation, meaning software and experiences tailored to one team, one niche or one customer, is now much easier to achieve. That’s the part I’d pay attention to.

What changed in build vs buy software?

The old reason to buy was that building took months and an engineering team. The evidence that this is shifting goes beyond one famous engineer:

McKinsey’s State of AI 2026 (1,719 respondents, surveyed May-June 2026): “Nearly a third of respondents (32 percent) report that their organizations have decided against buying one or more software products or features because they could be built internally with agentic coding tools.”
Retool’s 2026 Build vs. Buy Report (817 respondents, mostly Retool’s own customers, and Retool sells a tool for building internal apps): 35% have replaced at least one SaaS tool with a custom build, and 78% expect to build more in 2026. The categories replaced most are workflow automations (35%), internal admin tools (33%) and BI tools (29%).
Gartner (July 2026 press release): up to $234 billion of enterprise application spending is exposed to what it calls “agentic arbitrage” by 2030, about 20% of enterprise SaaS spend. Its analyst George Brocklehurst calls this “less an apocalypse and more of a metamorphosis”.

Two small, named cases show what “two hours” means in practice:

Safebird.ai replaced Cal.com. Its CEO wrote on HackerNoon that they dropped a two-seat Cal.com plan at roughly €30 a month, and “spec to deployed” took about two hours. He also wrote that it was “a deletion two months in the making”: the booking engine already existed for paying customers, and the spec took a focused hour. Two bugs slipped past the test suite and were only caught by trying it live. Hosting was left out of the savings.
Plastic Bank replaced Unbounce. Toby Stapleton rebuilt the company’s landing pages in nine hours over two days, replacing a $250 a month Unbounce plan, and kept them: “Test something Tuesday morning. Live by lunch.” His team are web developers, and landing pages are their core work.

A widely shared “we replaced SaaS with AI” story is weaker than its headlines. Klarna dropped its Salesforce contract, but CEO Sebastian Siemiatkowski later wrote on X, as TechCrunch reported: “So no, we did not replace SaaS with an LLM, and storing CRM data in an LLM would have its limitations.” Klarna built an internal tech stack, “using Neo4j and other things”, to bring data from many SaaS systems together. On whether all companies will do what Klarna did, he said: “I doubt it.”

Build vs buy AI software: the case against building

These are the cases where it went wrong, and they matter as much as the wins.

Plastic Bank also tried to build, then bought. Stapleton rebuilt an Instagram feed embed with Claude Code and got to “90%” in about five hours. The last 10% was API rate limits, expiring auth tokens, and a plugin that broke when a colleague renamed an Airtable field. With his time at about $2,000 plus maintenance and hosting, “We’d need to run Curator for 7 years before this made financial sense.” He paid $23 a month for Curator instead.

AI-built apps leak data. A researcher found 16 vulnerabilities, six critical, in an exam-and-grades app built on Lovable and featured on its Discover page, The Register reported. 18,697 user records were exposed. The AI-generated access check was inverted: it blocked logged-in users and let anonymous ones through. Lovable said every project gets a free security scan before publishing, and owners decide whether to act on it. In May 2026, security firm Red Access reported that of roughly 380,000 publicly accessible assets made with AI coding tools, on platforms such as Lovable, Replit, Base44 and Netlify, about 5,000 exposed potentially sensitive information, reportedly including medical records. Axios verified several, including a cabinet supplier’s customer support conversations, eWeek reported. Replit’s CEO said public apps being accessible is expected behaviour.

The code itself is often insecure. In Veracode’s 2025 test of more than 100 models on 80 coding tasks, “In 45 percent of all test cases, LLMs introduced vulnerabilities” from the OWASP Top 10 list, and larger models didn’t do significantly better.

Agents can destroy things. In July 2025 a Replit agent deleted a production database during a declared code freeze while SaaStr founder Jason Lemkin was testing it, wiping data on more than 1,200 executives and 1,190 companies. Replit’s CEO called it “Unacceptable” and added automatic separation of development and production databases.

The speed gain is smaller than it feels. METR’s 2025 randomised trial found experienced open-source developers took 19% longer on real tasks when allowed to use AI. Its February 2026 update estimates an 18% speedup for the 10 returning developers (interval from 38% faster to 9% slower) and 4% for newly recruited ones (15% faster to 9% slower). METR says selection effects now make the data unreliable and likely bias the estimate toward less speedup than the real one. So the evidence on speed for experienced developers is still unsettled. More on who gains in Does AI help beginners or experts more?

Maintenance doesn’t go away. Retool’s own report, from a company that sells building tools, warns that “Time saved on one task can be offset by time lost to maintenance…” Security and compliance concerns (41%) were the second technical barrier to more automation, just behind a lack of engineering bandwidth (42%), and only 14% were very confident in AI-generated code.

What does a built tool really cost to own?

The build is now the cheap part. Here’s what you keep paying for, using list prices as of November 2026:

Cost line	What it looks like	Source
Building it	Claude Pro at $20 a month (or $17 billed yearly) includes Claude Code; Max is $100	Anthropic pricing
Hosting	A small shared cloud server from €5.99 a month (Hetzner Cost-Optimized, listed as currently unavailable) or €6.49 (Regular Performance), excl. VAT	Hetzner
API quotas	YouTube Data API: 10,000 units a day by default; a `videos.list` call costs 1 unit, so even one call per HackerTube video (6,454) fits in one day	Google
Your time	Plastic Bank valued 5 hours at about $2,000, and the landing-page build at about $3,000 of time over a year	Stapleton
Security	A pre-publish scan, access rules on every table, separate dev and production data	Lovable, Replit cases
Bus factor	Whoever directed the build owns the bugs; at Safebird they now sit with the team that owns the booking engine	Safebird

Compare that with what you’d cancel: Cal.com at €30 a month, Curator at $23, Unbounce at $250, Mailchimp at $490. Curator wasn’t worth rebuilding. Cal.com was, for Safebird, only because its booking engine already existed. Unbounce was, for a team of web developers. For the fourth, Stapleton didn’t build at all: he switched to Loops at $199 a month, because “We were paying for 15 features to use 2.”

How much does it cost to curate with AI from your own data?

Here’s the arithmetic for a Karpathy-style filter. These are my numbers, not his.

One month (June 2025) of the same Hacker News dataset, measured for this piece, has 237,644 comments averaging 368 characters, and 0.69% of them contain a YouTube link. Using about four characters per token, plus the roughly 30% more tokens that Anthropic says its newer tokenizer produces, that’s about 120 tokens per comment. Using Anthropic’s Batch API prices (half the normal rate):

Every comment through Claude Sonnet 5.5 ($1 per million input tokens, $5 output): 43 million comments × 120 tokens is about 5.2 billion tokens, so about $5,200 of input plus about $1,100 for a short yes/no answer each. Roughly $6,300.
Every comment through Claude Haiku 5.5 ($0.05 input, $0.25 output): roughly $300.
A plain text filter for YouTube links first, then Sonnet on what’s left: about 300,000 comments, so about $36 of input and about $75 of output if each answer is 50 tokens. Roughly $110.

The YouTube share comes from one month, so treat these as orders of magnitude. The expensive version is optional. A cheap filter first, then a model on the shortlist, puts a curation pass over a big text archive in the price range of a single SaaS seat for a year.

Can curation from your own data beat an algorithm?

For some users, yes, and there’s research behind Karpathy’s frustration. In a pre-registered study of 806 Twitter users, published in PNAS Nexus in 2025, Smitha Milli and colleagues found that “users do not prefer the political tweets selected by the algorithm”. In an exploratory simulation that re-ranked the roughly 20 tweets each user was surveyed about, ranking by what users said they wanted cut angry and hostile content, but showed “a potential reinforcement of proattitudinal content”. In plain terms, people get more of what they already agree with. Curation that serves the user can also narrow what they see.

What businesses already do with their own text:

Amazon Rufus was trained on Amazon’s product catalogue, “customer reviews, community Q&As” and web information, Fortune reported. Andy Jassy said it “is expected to generate over $10 billion in annual incremental sales”. That’s based on what Amazon calls “downstream impact”; Fortune, citing genrise.ai, adds that Amazon uses a seven-day rolling attribution model. Amazon’s claim that Rufus users are 60% more likely to buy doesn’t say whether it controls for who chooses to use it.
Amazon’s review highlights use “only our trusted review corpus from verified purchases”. That’s the lesson for a small business: filter which data goes in before the model ever sees it.
RedditRecs, a site that ranked products from Reddit reviews with an LLM, now shows only “RedditRecs has shut down.” It doesn’t say why. If the data belongs to someone else, your product depends on them.

I found no published case of a small business building a recommendation feature from its own reviews or support tickets with a measured result. The large-company numbers are vendor-reported. The method is cheap enough to test yourself, as in the last section of this piece.

Replace SaaS with AI, or ship your data for other people’s agents?

The third point is the least obvious one. HackerTube’s export assumes the next user won’t use the site; their agent will rebuild a version that suits them. The open Hacker News dataset is the same idea one step up: Karpathy built on it without asking anyone.

Shopify stores now expose catalogue search, carts and policy questions to AI agents through an MCP endpoint at /api/ucp/mcp, according to Shopify’s developer docs. If your product is information (a directory, a curated list, a price comparison), a clean export or an agent-readable endpoint may reach more people than another interface. More on how agents read stores in How do AI shopping agents work? and on the simplest format for this in Does llms.txt work?

Counterpoint: an export is easy to copy, and I found no data on whether publishing one brings back customers or revenue.

What is AI personalisation in marketing, and why did it get cheaper?

In practice, AI personalisation means the content, recommendations or tool a person sees are picked or built for them rather than for the average user. HackerTube shows a route that needs neither a platform’s algorithm nor a data team: one person, a public dataset, a model to filter it, and a “For You” feed that runs in the browser with no accounts. Rufus shows the same pattern at Amazon’s scale.

The same logic applies inside a company. A sales team can have a tool made for its own pipeline stages, and a niche store can offer a “what customers like you bought” page written from its own reviews. For choosing where to start inside a company, see What’s the best first AI project?.

Build vs buy vs partner: which fits which situation?

Stapleton’s price tiers from the Plastic Bank experiments are the most practical rule I found:

$20-50 a month: “Just pay it.” Maintenance costs more than the subscription.
$200-1,000 a month, single purpose (landing pages, for example): the zone where building pays.
Mid-tier but best in class: survives, because rebuilding its edge cases is hard.
Enterprise and integrated: switching costs protect it.

Partnering, meaning hiring someone to build or adapt it for you, fits when the tool is worth building but nobody on the team can read the code or own the security. Gartner’s release also says vendors’ end-to-end agentic workflow products usually need a lot of services work today. I couldn’t read Gartner’s paid build-vs-buy notes, so that press release is the only Gartner input here.

How to decide

Go through these in order and stop at the first that applies.

It costs under $50 a month? Buy. You won’t recover the time, unless most of it already exists in your own code, as Safebird’s booking engine did.
It stores customer, payment, health or student data and nobody can check access rules? Buy, or partner. The Lovable and Red Access cases are this exact failure.
It’s best in class because of edge cases (calendar sync with Google’s login rules, email deliverability)? Adapt: keep the product and build the small piece around it through its API.
You use 2 of its 15 features? Switch to a narrower product before you build anything, as Plastic Bank did with Loops.
It’s $200-1,000 a month, single purpose, core to your work, and someone on the team can read and test code? Build. Write the spec first (Safebird’s was ten decisions in an hour), keep development and production data separate, run a security scan before publishing, and test it live, since that’s where both Safebird bugs turned up.
The value is in data you already have (reviews, tickets, comments)? Build the curation, even if you keep the software. My read: by the arithmetic above, it’s also the cheapest build on this list.
It’s personalisation for one team or one customer segment that no product sells? Build. My read: if no vendor sells it, there’s no subscription to compare against.

Who this is for

For: founders, ops leads and marketers at small and mid-size companies paying for single-purpose tools, and anyone sitting on a pile of reviews, support tickets or community posts. Not for: regulated systems (payments, health records), anything where a leak would be serious and no one can review the code, and large integrated platforms like ERP or CRM, where Klarna’s own CEO doubts others will follow.

What I’d do this week

Export last month’s card statement and list every software subscription with its price and the features you actually use. Mark each one with the decision list above. For the one that lands on “build”, write a one-page spec of the decisions it needs (Safebird’s had ten), give it two hours in Claude Code on the $20 Pro plan, and keep paying for the old tool until you’ve tested the new one live with real bookings, orders or users.

Separately, export three months of reviews or support tickets as a CSV and ask Claude to pull out every product recommendation or repeated complaint with the customer’s own words. If that list surprises you, that’s your personalisation project.

Sources

More of this every day on the channel: https://t.me/aidaily_business