- ThinkingAI began in Shanghai in 2015 as ThinkingData, with four engineers and the wrong first product.
- Customer conversations produced a 2017 pivot to granular, self-service game analytics.
- Agentic Engine now joins detection, analysis, experiments and campaigns in one governed workflow.
- The company sells to enterprises by demo, offers private deployment and does not publish list pricing.
The revealing object in ThinkingAI's product demo is not an artificial-intelligence model. It is a button. A company wakes to find seven-day retention down 12 percent. An agent examines behavioral events, support notes and community chatter; it traces the drop to a new onboarding flow; it proposes an A/B test. Then it stops. On the screen sits the button: approve.
That pause explains the company better than the vocabulary around it. ThinkingAI wants software to notice, reason and operate continuously, but it also knows the fastest way to frighten an enterprise buyer is to let a probabilistic system rearrange the business while everyone sleeps.
So the pitch is a loop with a gate. The agent does the tedious work. A person makes the consequential call.
The useful mistake
ThinkingAI is not a startup conjured by the agent boom. It is the second name of a business founded in 2015. Back then, four engineers left large technology companies and built a social-listening product for app stores, forums and online communities. There was one small problem: they had not first asked who would pay.
The company now tells this story on itself. “We shipped our first product before we had asked a single customer who would pay for it,” its history reads. That sentence is more instructive than most founding myths. The first thing to fail was not the engineering. It was the premise.
For two years, the team listened. Game operators kept describing a different headache. Their products generated huge streams of events, but a simple question could spend a day or two waiting in a data team's queue. Where did players quit? Which acquisition channel brought people who stayed? Did an event improve revenue or merely move it around?
In 2017, ThinkingData followed the questions. Thinking Game Analytics let operators explore granular behavioral data themselves. By early 2021, investor Blue Lake Capital said more than 300 game companies and 2,000 games were using the platform. Public references included Lilith Games, FunPlus and Kunlun Games. The narrow niche was the point: games produce enough events, fast enough, to punish vague definitions and slow handoffs.
“We shipped our first product before we had asked a single customer who would pay for it.”
A dashboard with hands
Traditional product analytics answers what happened. Amplitude, Mixpanel, Sensors Data and a long tail of gaming tools all compete somewhere in that territory. Engagement systems such as Braze help teams act on audiences. Enterprise agent platforms help assemble automations. ThinkingAI's wager is that the joins between these categories - analysis to explanation, explanation to experiment, experiment to campaign - are where time disappears.
Agentic Engine tries to own those joins. It connects structured events with less tidy material such as support feedback, community reviews and internal documents. A knowledge layer teaches agents what a particular company means by “revenue,” “new user” or “last week.” More than 100 packaged skills cover jobs such as retention, payments, advertising and operations analysis. Specialist agents can be chained in series or parallel: acquire data, run a multidimensional analysis, prepare a report, send a notification.
The system can run inside the customer's infrastructure and supports multiple language models. MiniMax is the default partner, not the only option. There are isolated sandboxes, permission management, token budgets, action logs, hallucination checks and staged A/B releases. In other words, much of the product is devoted to containing the product.
That is not a contradiction. It is what enterprise autonomy looks like when it leaves a keynote slide. A useful agent needs tools; tools create consequences; consequences require boundaries.
Who pays, and what it costs
The customers are data, product, growth and LiveOps teams. Gaming remains the sharpest use case - public customer names include Sega, Krafton, Xiaomi, Yostar and Habby - while ThinkingAI also points to social apps, e-commerce, live streaming and short-form drama. The company says its software has reached more than 1,500 businesses and 8,000 products in 20 countries.
This is B2B software sold by conversation, not credit card. Prospects request a demo; public list prices do not exist. That makes the honest answer to “what did it cost?” twofold. The sticker price is undisclosed. The organizational cost is easier to see: instrument events, reconcile metric definitions, connect operational tools, choose permissions, fund model usage and decide which actions require approval.
The company itself was financed heavily enough to make that second act possible. A CNY 100 million Series B arrived in March 2021, led by Blue Lake Capital with GSR Ventures and Linear Capital. A reported CNY 376 million Series C followed later that year. In August 2022, GGV Capital invested roughly US$15 million in an extension intended for product work, hiring and international expansion. Revenue is not public.
Four engineers launch social listening in Shanghai.
Customer interviews pull the company toward game analytics.
Thinking Analytics 1.0 turns the pivot into a platform.
Three reported rounds finance product and overseas expansion.
ThinkingData becomes ThinkingAI and launches Agentic Engine.
The thing anyone can copy
The transferable lesson is not “add agents.” It is to find a repeated decision with an observable outcome. Preserve the raw events. Agree on the meaning of the metrics. Let software draft the next move. Require approval while the system earns trust. Measure what happens, and feed that result back into the playbook.
This is a less glamorous recipe than autonomy, and more likely to survive contact with a finance department. ThinkingAI learned it in games, where a retention decline can be traced to a progression wall and a win-back message can be tested against a control. The loop is visible. The result arrives quickly. There is enough volume to distinguish signal from noise.
The catch is the plumbing
The model will not rescue missing events, disputed definitions or a workflow nobody is authorized to change. The approach is a poor fit when outcomes arrive slowly, the data is too sparse to test, actions cannot be safely reversed, or a company will not connect its systems. Private deployment helps with sovereignty; it does not make messy data clean.
That condition may also be ThinkingAI's best defense. A generic chat box can be copied in an afternoon. Ten years of event schemas, operational edge cases and practiced questions are harder to imitate. The company calls this history a knowledge architecture. A skeptic might call it accumulated scar tissue. Both descriptions are useful.
A company built around the handoff
ThinkingAI now operates from Sunnyvale with teams across cities including Shanghai, Singapore, London, Istanbul, Tokyo and Seoul. It reports more than 200 employees. Its market position is peculiar but coherent: too operational to be only analytics, too data-heavy to be merely an agent builder, too broad to remain a gaming tool.
The risk is equally coherent. Horizontal platforms from Microsoft, Salesforce and Google can bring distribution; analytics specialists already own dashboards; engagement vendors already own channels. ThinkingAI has to prove that combining the pieces produces more than a crowded control panel. It also has to show that lessons learned from games travel into businesses with slower feedback and less immaculate telemetry.
Still, there is intelligence in building around the handoff. Companies rarely lack charts. They lack agreement about what a chart means, someone with time to investigate it, and a safe path from conclusion to action. ThinkingAI's product is an attempt to make those gaps smaller.
The first product failed because the founders built before they listened. The present product is designed to listen forever. Then, crucially, it waits at the button.