Why AI SaaS is Booming, and Why We're Building the Same in Kazakhstan
In January 2025, Cursor—a code editor for developers—reported $100 million in annual revenue. By June, that figure had reached $500 million. By November, it was over $1 billion. Twenty-four months from the first line of code to a billion dollars ARR. No SaaS in history has grown this fast.
Yet, Cursor is neither magic nor another "killer" of something. It's a code editor with large language models integrated, trained to understand the context of the entire project. A developer writes a comment—Cursor completes the function. Asks to refactor—Cursor edits dozens of files at once. Nothing fantastic. Just old work done in a new way.
This article is an analysis of why such companies today grow faster than any classic SaaS, what they have in common, and where the Kazakhstani market opens similar windows of opportunity. I look at this not as an outside observer: West Star Ltd is itself building two AI SaaS products—an AI accountant for Kazakhstan and OData Hub for integrations with 1C. The parallels between what Cursor or Harvey has done and what we're trying to do in Astana are much more concrete than I'd like to admit.
What Has Happened to the Market Over the Last Two Years
To appreciate the scale, it's useful to put the numbers side by side. OpenAI at the beginning of 2023 was generating about $200 million in annual revenue. By August 2025, it was $13 billion. Anthropic, over the same period, grew from $87 million to $7 billion—a growth of 80 times in less than two years. Cognition AI, which makes the autonomous programmer Devin, started with $1 million ARR in September 2024 and reached $73 million in nine months.
These are not isolated cases. Harvey—an AI platform for lawyers—went from zero to $190 million ARR in three years, with a valuation of $11 billion. Lovable, a tool for generating applications from descriptions, gained millions in weeks. Bolt.new from StackBlitz exploded to $4 million ARR in the first four weeks after release—after the parent company spent seven years grinding its way to a modest $700,000.
The overall picture is simple. Previously, a SaaS startup would reach $10 million ARR in an average of 5+ years. The best did it in 2 years and 9 months. Now, the best AI-SaaS achieve this in 6-12 months and double their revenue every two months at the peak stage. This is not a "new economy" in the style of the dot-com bubble. It's a one-time shift where technology capable of automating expensive intellectual work has appeared—and the companies that first packaged it into a ready product took the market.
What Unites Successful Cases
If you analyze the top 10 fastest-growing AI-SaaS, a stable structure is visible. I've identified five patterns that repeat almost everywhere.
First—they don't replace people, they remove routine. Harvey doesn't do a lawyer's job. It reduces the time to find case law from ten minutes to ten seconds, automates the first draft of a contract, checks documents for compliance with the client's internal requirements. The lawyer still makes decisions and communicates with the client because AI can't do that yet. Cursor doesn't write the entire application—it takes on the part of the work the developer hated: boilerplate, renaming variables, generating tests. This is a critically important positional difference: the product doesn't threaten the profession but makes it more enjoyable.
Second—they choose a narrow vertical and go deep into it. Harvey doesn't try to be "AI for all professionals." It specializes in corporate law, M&A, due diligence—and trains models on proprietary documents of specific law firms. Allen & Overy, its first major client, made 40,000 requests from 3,500 lawyers in a couple of weeks during the pilot before signing a contract. When a product is deeply embedded in a specific workflow, it's hard to copy with a general tool like ChatGPT, even if the model itself is smarter.
Third—they operate where there is money and a willingness to pay. Harvey charges from $1,200 per user per year with a minimum of 100 users—meaning the minimum contract starts at $120,000. No one is shy. The legal market in the US alone is estimated at $300+ billion, and large firms are willing to pay for a tool that really saves senior partners' hours. Similarly with Gong—a sales call analysis system at $300+ million ARR—is sold not to users but to sales departments of large companies that have a budget for increasing conversion.
Fourth—they have their own data moat that can't be replicated with an OpenAI subscription. When a large bank or law firm trusts Harvey with its documents and they become part of the client's training corpus, it creates a switching barrier. Tomorrow OpenAI will release a model three times smarter—but it won't have this corpus and these trusted integrations. Sacra, an analytical agency, writes directly: models are commoditized, those who build workflows and integrations around them win.
Fifth—they sell results, not functionality. Harvey recently began transitioning from a "pay per seat" model to revenue-share with law firms: "we automate this service, share the income with us." This only works when the product truly does the work, not just helps to do it. And this is perhaps the strongest signal of maturity—the client pays not for a subscription but for a specific economic effect.
Where Kazakhstan Fits In and Why It Matters Now
An obvious objection: "These companies are in San Francisco, they have $800 million rounds from Sequoia, we can't compete with that." The objection is only half true.
What happened with Cursor and Harvey is the application of universal technology (LLM) to a narrow professional niche. And every country has its own narrow professional niches where global players will never understand. Harvey won't enter the Kazakhstani market and won't learn to understand the rulings of the Constitutional Court of the Republic of Kazakhstan and the old norms of the Tax Code of the Republic of Kazakhstan—because it's not a priority for them. The same is true in Russia, Uzbekistan, any non-French-speaking European market. Local players with deep knowledge of specifics get a window of opportunity that wouldn't exist with ordinary (non-local) technology.
This window, by the way, is not eternal. When GPT-6 or Claude 5 will understand Kazakhstani legal acts "out of the box," the local advantage will shrink. But "out of the box" is still not "deeply integrated into the workflow of a specific accountant or lawyer." Therefore, the window exists and is wide enough to build a business.
Parallels in Specific Products
The five patterns listed are not theory. I can honestly say how we try to apply each of them in our two products.
AI Accountant for Kazakhstan. This is a Telegram bot that answers questions about taxes, personnel, and accounting with references to specific articles of the Tax Code of the Republic of Kazakhstan and regulatory acts. Under the hood—Django, PostgreSQL with pgvector, and a RAG pipeline on OpenAI embeddings over the corpus of adilet.zan.kz, kgd.gov.kz, and dialog.egov.kz.
A direct parallel with Harvey: the product doesn't replace the accountant but reduces the time for typical questions. A typical request—"How to calculate OPV for a foreigner on a labor contract in 2026?"—takes a live consultant 15-30 minutes of searching forums and calls. The bot gives an answer in 10 seconds with a link to the current edition of the article. It's the same logic of "removing routine, leaving judgment" that worked for Harvey.
Narrow vertical—yes, we deliberately don't build "AI for all Kazakhstani business," only accounting and taxes. Market readiness to pay is another story: a Kazakhstani accountant pays for reference systems like "Uchet.kz" or "Paragraph" from 30 to 100 thousand tenge a year, and if it's about a corporate subscription for an LLP accounting department—orders are higher. The willingness to pay for a truly working AI assistant is being tested now, in pilot sales. Data moat—it's a corpus indexed for the Kazakhstani context; the global ChatGPT hasn't seen this corpus and hallucinates on 30-50% of tax questions.
OData Hub. Here the story is different, but the logic is the same. It's a SaaS platform for integrating 1C with external systems via OData. The problem the product stands on: Kazakhstani companies have data about everything in 1C—sales, balances, contracts, payroll—but extracting it into a modern stack (Python, BI tools, AI models) is painful. Each company solves this anew, hiring a 1C programmer for $30-50 an hour.
Parallels with global cases: removing routine (writing integrations for each request), narrow vertical (specifically 1C + modern systems), data moat (accumulated connectors and business logic of processing 1C objects that ChatGPT will never "learn" itself). Monetization model—implementation packs from 1 to 5 million tenge plus support 50 thousand tenge per user per month. This is structurally closer to Harvey ($1200/seat/year minimum 100 seats) than to B2C applications.
I'm not claiming we'll create a Kazakhstani Harvey. I'm asserting that the structure of successful AI SaaS is reproducible in local markets—if you choose the vertical correctly, focus on routine, and don't try to replace people.
What Doesn't Work (and What Is Rarely Talked About)
An analytical breakdown would be dishonest without the other side. ChartMogul, a SaaS metrics aggregator, warned in 2025: AI-native companies show lower customer retention than classic B2B SaaS. The explanation is simple: an AI product is easy to buy, easy to cancel. If the client hasn't deeply integrated the tool into the workflow—it will drop off in 3-6 months. Therefore, rapid ARR growth doesn't equal a sustainable business: half of the AI startups of 2024 will be in trouble in 2026-2027 when the hype subsides and investors start asking about NRR (net revenue retention).
The second limitation is competition from the models themselves. Anthropic has already released a Legal Plugin that partially overlaps with Harvey. OpenAI is doing the same in a dozen verticals. The thinner the "wrapper" around the LLM, the higher the risk that the base model will eat it. The defense is deep integrations and industry data that general models don't have.
Third—B2B sales cycles are long. Harvey took about a year to reach the first million ARR, another year to ten. Allen & Overy needed several weeks of piloting with 3,500 lawyers to decide on a contract. In Kazakhstan, the rhythm is similar: a serious B2B client doesn't sign a contract in a week based on a Telegram presentation. Therefore, the strategy of "launching a product and waiting for clients" doesn't work here—you need to invest in sales.
What Follows From This
To simplify to one paragraph: the main AI SaaS of the last two years won because they found a narrow profession with expensive routine and shifted this routine to the machine, leaving judgment to the human. The technology—large language models—is publicly available. The advantage is in local knowledge, data, and how deeply the product is embedded in the workflow of a specific specialist.
This means that the window for local AI SaaS is open right now, and it's quite wide. Kazakhstani accounting, law, tax consulting, document management, warehouse accounting, 1C integrations—all these areas have the same profile as law in the US: expensive professionals, repetitive tasks, specific context that global models don't understand. And in each of these areas, someone will build a "Kazakhstani Harvey" in the next two to three years.
Whether we will succeed is a matter of execution. Whether it will be a Kazakhstani player at all or whether Uzbek, Russian, or Chinese teams will overtake us is another question. But the window is there, and the logic of global cases provides a fairly clear template of what works in this window: narrow vertical, routine instead of judgment, deep data, monetization through results.
In 2018, no one seriously believed that a real IT business could be built in Astana. In 2026, no one seriously believes that a real AI SaaS can be built in Astana. History shows that usually the next five years change everything.