"There is a chasm between the LLM and your enterprise business and context"

Hello and welcome to the latest edition of AI Gamechangers. It’s your regular newsletter where we chat with leaders from the games industry who are doing practical, innovative things with AI.

Since the last edition, the news has been full of the threat AI poses to humanity. If you’d like some historical and cultural context to this discourse, check out the book Myth-Made Machines, co-written by technologist and founder Kelly Vero and the AI Gamechangers editor Dave Bradley.

In the meantime, your newsletter Q&A this week is with Chris Han, co-founder of ThinkingAI (formerly ThinkingData), whose company has just launched its Agentic Engine: AI agents that don’t just flag an insight in your game’s data, but go on to rebuild the segment, prepare and launch the campaign, and A/B test the result, with human approval built into the process.

As always, read to the end for some useful news and links.

Chris Han, ThinkingAI

Meet Chris Han, co-founder of ThinkingAI. Han started out as an application engineer at Intel before founding ThinkingData from a Shanghai incubator in 2015 with three fellow engineers. Over the past decade, their platform has worked with games companies including Sega, Krafton, Habby and Century Games.

This April, the company rebranded from ThinkingData to ThinkingAI, and this month launched Agentic Engine: autonomous agents that turn player data into live campaigns and product changes, running on a studio’s own infrastructure rather than the open cloud. In August, Han and his team took to Gamescom in Cologne to explain more about the agentic way of thinking.

In this interview, he explains the opportunities of working with agents at an enterprise level, and how to trust AI to work with live revenue.

Top takeaways from this conversation:

1. Agentic Engine goes further than a typical analytics agent: rather than surfacing an insight for a person to act on, it updates the tracking plan, rebuilds the segment and launches the campaign (with human approval).

2. The whole system runs on the customer’s own infrastructure if required and is deliberately decoupled from any single model, so studios worried about data sovereignty or GDPR can plug in their preferred AI.

3. A decade of behavioural data from over 1,500 clients and 8,000 games has been turned into pre-built “skill sets” that Han says can bring an agent up to speed on retention or monetisation analysis in under an hour.

4. Han argues the models themselves are no longer the bottleneck, it’s trust; ThinkingAI’s answer is a “semantic layer” plus a beta Agent Monitor tool that traces agent behaviour and flags potentially dangerous signals.

AI Gamechangers: Before we get into Agentic Engine, can you tell us a little about your own background? How did you get to this point?

Chris Han: Before we founded this company, I was an application engineer at Intel. Basically, I was enabling our ISVs with the latest Intel drivers and hardware. But I found it a little bit boring in a big company, in a big tech company.

So in 2015, three friends, who were also engineers, and I founded this company in Shanghai in an incubator. We were engineers doing seven days a week, maybe 12 or 14 hours a day, always coding. It has already been 11 years, and it was always about the ups and downs.

“The games industry is a very open industry and a very cutting-edge industry. Whenever new technology comes out, games are always one of the first industries to integrate it”
Chris Han

In China, Japan and Korea, we did a pretty good job. A lot of game companies started to use us, and they saw pretty good and solid business value from that. So we decided to expand our business to Europe and the US. That’s how I got here - I’m leading the team as we expand our business across Western markets.

AI Gamechangers: ThinkingAI has just announced Agentic Engine, AI agents that run growth on a company’s own infrastructure. Please tell us more about this big news. And what’s the pitch for your launch this month?

Chris Han: Actually, in April we changed the name from ThinkingData to ThinkingAI. At that point, we introduced the product a little bit, but after several months and a lot of iterations, I think the product is mature now. We have case studies and several clients already on board. We are pretty confident about it, and the agent layer is running very smoothly.

What we help our clients do is enable a whole growth agent team. It can help you find insights and anomalies, and help you catch those atomic opportunities, as I have spoken about before.

Before agents, we needed to catch those opportunities manually using our software. Right now, the agents can help you do that. And it is not only about the insights. We can apply those insights to other agents, such as an engagement agent or an A/B testing agent, which can do the solid work, launch the campaign and help iterate the product autonomously.

That is what Agentic Engine is about. It enables you to have a whole growth agent team, cooperating with humans, so you can do more, launch more and iterate more.

AI Gamechangers: In your Gamescom demo last month, the AI went from a question about retention all the way to drafting a live campaign and A/B tests, with an approval point before launch. Where does the line move from an agent reporting and recommending to an agent actually deciding and acting? How do studios get comfortable with that?

Chris Han: Great question. Everybody has agents right now, especially personal agents. For me, I cannot work without my personal agent.

But at the enterprise level, there are still a lot of challenges. You cannot just say, “Let’s do agents. Connect the data, get insights, approve the insights and apply those to campaigns right away!” We can’t do that.

There is a huge gap, a chasm, between the LLM and your enterprise business and context. The big models do not understand your business. Even for an individual, an agent cannot catch the full context. Sometimes you get hallucinations or answers that are not accurate enough.

“We have what you could call an AI governor, or what we call the Agent Monitor. When you engage more and more agents, more and more agent behaviour is generated. We need to monitor those agents and make sure they work within meaningful boundaries. You need to make sure your agents work together very closely and stay under your umbrella”
Chris Han

So before we enable the agents, we need what we call the semantic layer, and we help our clients build that. We have features such as the knowledge base and skill sets. Those functionalities help you build a semantic layer that can also be iterated.

If you have very good case studies from the campaigns you launch, you can write those back into the knowledge base. The agents remember that and build on it. So the first challenge is helping the big models understand your business better and more deeply.

Once we build that, we can enable the agentic technology and build agents on top of it. You get more and more confidence in your agents, and you become more comfortable letting your agents do the real work. That is how I would describe the architecture and the process, although there are still a lot of challenging points in the details.

AI Gamechangers: ThinkingData had already been around for a decade. What is fundamentally different about ThinkingAI and Agentic Engine, and what does the rebrand signal?

Chris Han: That is exactly why we did the rebranding. In the ThinkingData era, we believed in being data-driven, and we still believe in being data-driven. Data is still the fundamental layer for agents and AI. Without comprehensive and accurate data, and an appropriate tracking plan, your agents can do nothing.

Before ThinkingAI, our product was Thinking Engine. Now it is the Agentic Engine. Thinking Engine was an analytics and engagement platform that helped you drive more insights, analyse player behaviour in more detail and apply those insights in one platform, whether that meant launching campaigns or doing A/B tests.

We had a very solid software base. We have helped around 1,500 clients globally, with about 8,000 games integrated into our system. Among them are studios such as Sega, Krafton, Century Games and Habby, with a lot of very successful games integrated into the system.

But when the agentic layer and agentic technology came out, especially after OpenClaw appeared at the beginning of this year, we saw that the future of software is going to be software used by agents, not just by human beings. The way humans interact with software is going to change totally. That is what we believe.

So we added the agentic layer. The software I just talked about can now be used by agents rather than only by humans. Of course, you can still go directly into it and check the data yourself, but that is the fundamental change.

Over the last decade, we have collected a lot of knowledge from our clients and case studies, and built very mature and comprehensive skills that can be enabled for agents. That includes how to analyse Day 7 retention, how to work on monetisation, how to analyse detailed player behaviour patterns, how to tag users and players in appropriate ways, how to drive campaigns and how to do live ops.

An agent can be enabled with those skill sets very quickly, maybe in less than an hour. For a human to become familiar with everything can take a very long time. Software also becomes more and more complex. Sometimes clients would tell us, “The software is so complex, I don’t know how to get to the right function.”

With the agent layer and Agentic Engine, everything becomes more efficient. We can help clients grab more of those atomic opportunities to iterate the product and give players the best experience.

AI Gamechangers: What models sit behind Agentic Engine? You have a partnership with MiniMax, but can it also work with other major LLMs or a company’s own model?

Chris Han: At our rebranding event, we announced our strategic partnership with MiniMax. But from a technical perspective, we are totally decoupled from the LLMs. That means we can work with any kind of model, including your own model. At the enterprise level, some larger companies want to train or run their own models because of security concerns.

“The lesson is that you have to take a very deep look into the players, because we believe that every player deserves the best experience if they want to spend time on your game”
Chris Han

Basically, we can connect to whatever model you want to use. It is the same principle as in the Thinking Engine era, when we were decoupled from the cloud service provider and could deploy our software on different kinds of servers.

We want to keep it decoupled. Whatever our clients choose, we can work with that.

AI Gamechangers: Can you share a specific example of how a game company is using the platform and what it has learned from the data?

Chris Han: Maybe I shouldn’t speak about any specific company, but I can give you an example. There is one case study I remember very clearly because we worked with them for years, and now they use our agents.

When they launched the game, it was a simple casual game with pretty simple gameplay and not that many users. But they had a very long-term strategic vision. They took a very deep look at the players and, gradually, they changed the gameplay. The whole casual game became more of a mid-core game.

It has been operating for about nine years now. We have enabled them with Agentic Engine, which helps them find more detail about the players and helps them iterate the product in more detail.

The lesson from that example is that you have to take a very deep look into the players, because we believe that every player deserves the best experience if they want to spend time on your game.

Players are different. There will be different cohorts, and you should feed them different content based on the balance of the game and what they need. If you give them the best experience, they will stay. They will retain and grow with your game.

One reason that team changed the gameplay was that they found the players wanted to spend more time in the game, so they added more content and social systems as the game grew.

That example really impressed me. Sometimes you are amazed by your users and your clients, and by how they engage with your software. That is probably the most touching point for me.

AI Gamechangers: Does ThinkingAI work for studios of all sizes? Can a small team without a dedicated data department get started with it?

Chris Han: We can work with teams of different sizes. For enterprise-level companies, they may already have a mature database, data collection and big data infrastructure, and we can work on top of that.

For small teams and startups, they may not have that infrastructure at all. One big advantage of our platform is that we support self-deployment. From data collection, data storage, data calculation and data visualisation all the way up to the agent application, we can help them and deploy everything to their servers. That means they can have mature data infrastructure like an enterprise-level company.

“I think the way humans interact with software is going to change totally. How you operate your game and how you analyse your game will change. You will have more and more mature agent teams helping you analyse your game and do A/B testing”
Chris Han

Sometimes a small team does not want to buy or manage servers. We also support the traditional SaaS approach. We can open an account for them, and they can import their data.

With agents, I think the first one or two steps of the integration also become much easier. Before, you had to create the tracking plan and then integrate the data collection code. Now the tracking plan can be generated very quickly based on the skills I talked about, and SDK integration becomes much easier.

AI Gamechangers: You said at Gamescom that ThinkingAI itself has changed the way it works, as agents have become more important. How are you using these tools internally, and how has the company evolved?

Chris Han: We are going to release a book about the AI-native organisation as well. But to answer the question, yes, we need to embrace this era totally, and you need to catch up.

There are a lot of different perspectives. The first is development and software releases. Before, we might release one version every two months. Right now, it can be every two weeks, or sometimes every week. That is very quick and very flexible, and everything happens so fast.

That generates new questions. Because everything happens so fast, cooperation and communication can become a problem. How do we catch up with each other? Right now, everyone has a personal agent, and the personal agent will update your work in one place. Agents can then sync development information and logs with other agents.

“[AI] can help you find insights and anomalies. Before agents, we needed to catch those opportunities manually. Now, the agents can help you do that. And then we can apply those insights to other agents, which can do the solid work”
Chris Han

The most obvious change is probably the speed of our software releases, but it applies elsewhere too. Take customer success. We already provided a very comprehensive and friendly onboarding service with our customer success team. With agents, we can collect all that professional knowledge and those solutions in one place, and enable an agent to answer questions very quickly, with the approval of the customer success team. I think the whole onboarding process becomes even friendlier.

It is the same in marketing. When you are building the website or trying to build the inbound funnel, every detail changes.

Across the whole ThinkingAI team, I would say everyone embraces agents. And yes, of course we use our own platform to understand how users interact with our software and how they interact with the agents. We generate patterns from that and iterate our product based on those suggestions and the data.

AI Gamechangers: Is there anything else in the current product that you are particularly excited about, or that you think people should know about?

Chris Han: There is one thing. You mentioned the dangers and risky behaviour from agents. We have what you could call an AI governor, or what we call the Agent Monitor, and it is currently in beta.

When you engage more and more agents, more and more agent behaviour is generated. We need to monitor those agents and make sure they work within meaningful boundaries.

In that module, you can see agent traces, the agent stack and agent behaviours, and it can help you find potentially dangerous signals from your agents. You need to make sure your agents work together very closely and stay under your umbrella.

I think that is pretty important for our clients.

AI Gamechangers: If agentic tools such as ThinkingAI become standard across the games industry, how do you think the industry will look in two or three years?

Chris Han: In terms of operations and analytics, I think the way humans interact with software is going to change totally. How you operate your game and how you analyse your game will change.

Nowadays I still see a lot of teams using very legacy software stacks. They are still using [Business Intelligence] tools and looking at metrics that tell you nothing. I think that is going to change totally. You will have more and more mature agent teams helping you analyse your game, do A/B testing on your game and operate your game.

The games industry is a very open industry and a very cutting-edge industry. Whenever new technology comes out, games are always one of the first industries to integrate it.

From the development perspective, a lot of great things are already happening. We have so many 3D generation tools that can inspire and spark new ideas. We can see NPCs becoming more and more human-like, and more and more agents being used inside games.

From developing and testing your game, to launching it and then operating it post-launch, the whole lifecycle is going to change. I think the other phases of the business will change totally as well.

Why not subscribe? It’s free, and you’ll receive interviews with industry insiders and studio leaders in your inbox. We have many conversations in the pipeline, including Spencer Wade of MAXX Apps, Robyn Pacetti of Protagonists and more.

Further down the rabbit hole

Some useful news, views and links to keep you going until next time:

• Jack Buser, Google Cloud’s Global Director for Games (and a past AI Gamechangers interviewee), wrote a Gamescom-timed blog calling AI an “Iron Man suit” for developers, arguing it can make creative risks affordable again. His examples include Capcom running 30,000 hours of automated playtesting a month and the work done by 10Six Games.

• The EU has designated Roblox as a Very Large Online Platform under the Digital Services Act, alongside Reddit. ChatGPT gets a designation too, as a Very Large Online Search Engine. That puts the UGC giant and the LLM in the same regulatory bracket, with tougher obligations on risk assessment and content moderation to follow.

• TypeSafe AI, the $40 million startup founded by ChatGPT co-inventor Diogo Almeida, has launched Jev, a model that doesn’t write sentences at all. It returns typed, probabilistic decisions in milliseconds, which is fast enough to play Doom in real time. Games are the showpiece, but its real day job is likely to be sorting customer service tickets.

• Steam’s “AI Gaming Festival” has been a lesson in how not to win hearts and minds: players are blocking the games involved en masse, and AI and Games’ Tommy Thompson says the event is run by an AI-games platform rather than Valve, with a paid conference attached. The festival is on until the end of the month.

• Ramen has launched Aura 1.0, its AI agent for Unity and Unreal Engine developers, after seven months in beta. The headline feature is a Verification Agent that playtests games by itself, now eight times faster. There's also project memory, Mac support and a listing on the Fab store. It's the tool that CEO Andy Tsen told us about earlier this year.

• Solo players are getting AI teammates. Artificial Agency's Behavior Engine will power a new Single Player Mode in Owlchemy Labs' Dimensional Double Shift, where voice-controlled AI robot colleagues run a diner shift with you. It's due in the first half of 2027, and the company says a small, purpose-trained model keeps costs down.

• Round one in the war against the machines: influencer Frankie Lapenna has taken on a 6ft android in an MMA cage in San Francisco. Rek, the robot’s maker, admits it was remote-controlled by a human, so it’s less Skynet, more very expensive puppet.

• Studio Atelico has opened an iOS closed beta for Bobium Brawlers, a card creature battler where you describe a monster in 140 characters and an on-device AI engine turns it into a creature with its own deck. The models are trained on the studio’s in-house art team’s work. The full launch is due later this year.