The project-management methods I learned in the game industry over the past ten years are failing as a set. Gantt charts, burndown charts, person-days, sprints, iterations, milestones, process rules — we used them to break down a version, constrain a team, and report progress to a producer. Agents are becoming the real unit of productivity. It is 2026.

1. Person-day estimates are turning into a collective illusion

The same programmer might need two weeks for a combat system’s basic framework without AI, and three days with AI coding; the same artist, a week for concept art without AI, and five options in a day with AI assistance. Output can swing by a factor of three to ten. Some people have already used Claude Code to rebuild their entire workflow; others still treat AI as a search engine. Same role and level, capacity can differ by an order of magnitude. Estimates are degenerating from a precise unit into a fuzzy range.

2. Linear process is being punched through by AI

The classic chain runs from the requirements document through review, task breakdown, development, handoff, QA, bug fixing, and acceptance. When a designer can use AI to generate a runnable prototype, “requirements document, engineering comprehension, technical review” can be compressed into one step. When AI can generate test cases and run regression automatically, “handoff, QA, regression” is being compressed too. A two-to-four-week sprint used to be reasonable; after AI compresses the middle, a week or even a few days can finish a former sprint’s work. Project management then shifts to coordinating several short chains in parallel.

3. The daily stand-up needs to be redefined

Fifteen minutes every morning, three sentences each: what I did yesterday, what I will do today, what is blocked. But when AI can track commits, task status, and document updates in real time, and can already auto-generate a daily progress summary, flag tasks drifting off plan, and identify likely blockers, do you still need a verbal report from everyone? If everything in a stand-up can be summarized by AI, that meeting should not be held.

4. Document-driven collaboration is being overturned

“Did you read the doc?” may appear almost as often as “When will this requirement be done?” When an agent can read the codebase, understand context, and trace earlier discussions, it needs a structured, machine-readable description of the requirement, not a carefully formatted design bible. In a fifty-page design document, perhaps only 20% is real decision information; the other 80% is background, explanation, and formatting, which AI does not need. A future design document may no longer be a Word file, but structured data, executable rules, and a definition an agent can consume directly.

5. Full-time QA is being rewritten

Much game testing is repetitive and mechanical: walking every corner of every map, trying every equipment combination, triggering every dialogue branch again and again. AI can run automated tests 7×24, simulate tens of thousands of player paths, and finish in minutes after each commit a regression that used to take a day. A version regression that once took ten QA people a week may be finished by an AI agent in a single night. Mechanical testing goes to AI; human QA moves to strategy, quality standards, and assessments of feel — work that requires an understanding of what “fun” is, which AI still cannot do well.

6. Management intuition is being overtaken by data agents

As a game project manager in my 10th year in the industry, what was I most proud of? Intuition: one look at the schedule and I knew where trouble was coming. But when an AI agent can analyze project data in real time — commit frequency, bug-convergence curves, requirement-change history, load across people — and then predict risk and offer advice, experience-only intuition starts to look crude and dangerous. One person can watch a dozen risk dimensions at most; AI can monitor hundreds of metrics at once. Experience still has value, but it has to be combined with AI’s analysis.

The old agreements are in pieces. Under the rubble, new rules are already growing.

Rule 1: Spending on AI is not a cost. It is an arms race

For a team that uses AI coding in depth, monthly token consumption can easily exceed ten thousand dollars — an investment, not a cost. Competitors are already producing code two to three times faster, iterating ten art options a day, and using AI for rapid prototype validation, while others still hand-paint a second draft or write a third requirements document. Saving money on AI use is conceding the strategy. The flywheel is more AI investment, faster iteration, higher quality, better market feedback, more revenue, then still more AI investment. Once it is spinning, teams that cannot catch up are thrown further behind.

Rule 2: From managing people to orchestrating agents

The old core — communication, allocating resources, controlling the schedule, managing risk — was, at bottom, “managing people.” When agents become de facto units of productivity, you need to know which tasks can go to AI and which must be done by a person, design people-and-AI workflows, and assess the quality of AI output, not only human deliverables. Just as a director schedules actors and a VFX crew, a PM has to schedule human teammates and AI agents so they meet at the right points. A PM who cannot orchestrate agents will be left behind as quickly as a PM who cannot use JIRA.

Rule 3: From large teams and long cycles to small teams and fast iteration

A mid-size game used to need a team of 50 to 100 people and a cycle of three to five years, because every step — art assets, feature development, level design, test and validation — demanded a large amount of human effort. When AI can help generate art assets, speed up coding, and automate testing, the same work needs fewer people: perhaps 10% fewer now, and 30% to 50% fewer in two years. The future winners may be elite squads of 10 to 20 people, each using AI in depth, with actual output equivalent to a team of 50 to 100 in the past.

Rule 4: The PM’s ultimate value returns to decision-making

Once AI can do progress tracking, risk warnings, meeting notes, data reports, and even a first-pass resource plan, what is left for a human PM? Judgment and decision-making. An agent has data, efficiency, and patience, but not an intuition for what is “fun,” a sense of the team’s mood, a nose for market timing, or the nerve to decide when information is incomplete. What genre to make, which feature to cut, whether to slip a test date, what to do when morale is low — these have no standard answer. AI can put every number in front of you; a person has to make the call. Like a general, the skill is deploying the formation: however many soldiers (agents) you are given, you can command them, and still win.

In the old world, people were the bottleneck, process was the safeguard, documents were the anchor, and experience was the moat. In the new world, agents are capacity, efficiency is an investment, data is the anchor, and decision-making is the moat. The old process is dead. Agents endure. Let us have the courage to throw away the old flowchart, and take up the new productive force.

This English version is adapted from the public Chinese original: https://pmnote.ai/articles/ai-change-in-game-pm-eyes/

More notes on game project management at https://pmnote.ai/, and in the WeChat mini program “Hao的游戏PM笔记”.

A tech lead's crash

The day before, I had a meal with our tech lead and we talked about AI coding. He said he had recently let AI write a fairly complex feature. It produced two crashes. He asked AI to fix them itself, and after several rounds it still couldn't.

So he had to read the code himself, and found something interesting: the AI-written code ran, but its logic paths were not quite how a person would think. Once something went wrong, it was hard for him to tell by intuition where the problem was. In the end he located the specific issue himself, gave AI a very precise description, and only then was the bug fixed.

His takeaway: for now, there is still no way to fully trust AI; you can only use it as a tool.

Where the trust boundary is

People on the team who code with AI share a very consistent feeling: handing simple features to AI is great. You describe the requirement clearly, AI writes it, tests it, and completes the delivery. The whole process is smooth and the efficiency gain is obvious.

Complex, architectural features are different. Even senior programmers don't really dare to let AI write the core architecture on its own. The reason: over time you risk losing control of your own project's code. If the code wasn't written by you, you don't know what the overall architecture looks like, and the day a bug appears that AI can't handle, you are in a very passive position.

Just like the opening case: you have to understand it from scratch, as if reading someone else's code — much slower than debugging code you wrote yourself.

The more mature way to collaborate right now:

• Let AI write a first version; a person reviews and modifies it.
• Use simple modules directly; be deeply involved in complex ones.

A person's core advantage is understanding the business — AI provides the ingredients, people provide the recipe. How to analyse requirements, how to break down a solution, and how to make decisions are still irreplaceable.

One underlying insight

To sum up the trust boundary for AI coding:

• Simple features: hand them to AI with confidence.
• Complex architecture: humans must be deeply involved.
• You can let AI do work for you, but you must not lose your judgment of it.
• Using AI is not just about knowing how; keep improving how you and AI work together.

More important is one underlying insight:

AI has lowered the cost of output, but it has not lowered the cost of decisions.

The bar for execution is dropping; the value of judgment is rising. Your competitiveness will increasingly show in judgment rather than execution speed.

As long as you keep improving your judgment and decision-making, AI's progress is not a threat but leverage — the stronger it gets, the more output you can move.

This English version is adapted from the public Chinese original: https://pmnote.ai/articles/bilibili-bv18jqhbee57/

More notes on game project management at https://pmnote.ai, and in the WeChat mini program “Hao的游戏PM笔记”.

Summary: AI is not making game work easier. It is raising the output demanded of the same role, widening the gap between people in the same title, and re-ranking three kinds of people: those who hold resources, those who combine a professional foundation with deep AI use, and pure executors whose work is mostly moving information.

The short version: higher technical efficiency does not automatically mean a lighter week. In the source’s account, it usually means more output is asked of the same person. The more important split is not job title. It is whether you hold resources, whether you can combine professional judgment with deep AI use, or whether most of your time is still spent moving information.

Efficiency does not buy you an easier day

A common hope is that AI will let people leave work earlier. The source’s reading of history points the other way. Medieval English farmers, as cited there, worked about 150 days a year, often six to eight hours a day. The length of a work year is not only a function of tools. It is technical efficiency multiplied by a power structure: who captures the gains from higher output. When a programmer can use AI to finish in a day what used to take a week, the usual result is not a four-day weekend. It is more features in the same calendar. When an artist can produce five concept versions in a day, the usual result is more iteration, not fewer reviews.

The staffing numbers in the source make the same point at team scale. In 2018, a mid-size mobile project often took 50 to 60 people and two to three years. In 2025, a project of similar scale might run with 30 to 40 people, at higher intensity. The work did not shrink. The demand on each person rose.

Two further observations follow. First, people in the same role and at the same level can now differ in output by an order of magnitude. One person uses a tool such as Claude Code to rebuild a workflow. Another uses AI as a slightly better search box. The title on the door can be identical. The output is not. Second, a “super individual” working with AI may outproduce a five-person team, not only because of raw speed, but because coordination cost is cut. Fewer handoffs can matter as much as faster execution.

Three ranks, not three job titles

The source groups people into three ranks, and it is careful not to treat them as official titles.

The first rank is resource holders: investors, founders, and core producers. The criteria named are not a fashionable stack of tools. They are deep industry knowledge, having learned from failure, and something like 10,000 hours in the craft. These people decide what gets made, who gets funded, and which bets are allowed to continue.

The second rank is people with a professional foundation who also use AI deeply. The examples are concrete. A project manager automates parts of planning, tracking, and checking. A designer generates a runnable prototype instead of only a document. A technical artist embeds AI in a pipeline, so the gain is repeated rather than one-off. An artist can go from concept toward a more finished piece without waiting for every intermediate handoff. This rank is not “people who have tried a chatbot.” It is people who can still do the professional work, and who have changed how that work is produced.

The third rank is shrinking: single-skill pure executors. The outsourcing-management example in the source is useful because the job can look like management. If more than 70% or 80% of a person’s time is spent moving information — collecting, forwarding, formatting, and chasing — the rank is falling, regardless of the title on the card. AI is especially good at that kind of work. A role defined by it is exposed.

The salary history in the source is a reminder that the last expansion was real, and that it does not automatically repeat. From 2015 to 2020, ordinary salaries in the industry could move from the low hundreds of thousands of RMB a year to three or four hundred thousand, or more. After the 2022 license winter, the more common company behavior described is to keep cash on the books. Even if a company earns, it may not share the profit in the old way. Meanwhile, a competitor’s project managers may already be using AI to run two projects. The comparison is no longer only to the person at the next desk. It is to a different production system.

The source also notes people leaving after seven or eight years — to a local studio, to education, or to a product-manager role in a more traditional industry. Leaving is not treated as failure. It is treated as a choice that should be made consciously, not as a drift that happens after a rank has already fallen.

The author mentions, in passing, having used AI to build a project-management community site. The point of that line in this English version is only that the same pressure applies to the person writing: using the tools is part of staying in the second rank, not a side hobby.

Three actions, if the ranking is already in motion

The source does not offer a guarantee. It offers three actions.

First, make an honest assessment of your rank. The test is not what you are called. It is what share of your week is judgment, decision, and coordination, and what share is moving information. If most of the week is the latter, the title will not protect the role.

Second, invest in the skills that are harder to automate: judgment, decision-making, coordination, and systems thinking. Tool fluency matters, but only on top of a professional foundation. A person who only learns shortcuts, without a way to tell whether the output is right, is still close to the third rank.

Third, choose actively. That may mean staying and changing how you work. It may mean leaving for a smaller studio, for teaching, or for another industry. The cost of not choosing is that the re-ranking happens anyway, and you notice it only after the work has already changed around you.

FAQ

Will AI make game-development jobs easier?

Not in this source. Higher technical efficiency is described as raising the output demanded of the same role, not as shortening the week. A faster coding day becomes more features. A faster art day becomes more iteration.

What does the source mean by people in the same role differing by an order of magnitude?

It means that two people with the same title can now produce at very different levels, depending on whether they use AI to rebuild a workflow or only as a search box. A single person plus AI may also outproduce a small team by cutting coordination cost.

How do you tell whether your rank is falling?

One test in the source is time: if more than 70% or 80% of the week is spent moving information, the rank is falling, even if the card says “manager.” Resource holders and deep professional users of AI sit in a different place.

Does the source say everyone should stay in games?

No. It describes people leaving after seven or eight years, and it treats leaving as a choice to make consciously. The warning is against drifting until the re-ranking has already happened.

This English version is adapted from the public Chinese original, without adding cases or figures that are not on the source page:

https://pmnote.ai/articles/ai-reshaping-game-industry/

The original article and other public notes are at https://pmnote.ai.

AI companies are treating games as a main arena for deploying technology, not only as lab demos, and that pressure is likely to keep rising. Of the five directions in the original article, the ones production teams should watch most closely are these: world models are accelerators in the near term, not replacements for a full level-production pipeline; AI rendering is changing how performance is optimized; and the repetitive "heavy lifting" in asset production is being taken over first. The practical move is to find the most painful step in your own pipeline and let AI try it now, rather than waiting to adapt under pressure.

Why these headlines are closer to production than they look

The original article starts from a Silicon Valley 101 episode about why leading AI companies are placing bets on games: NVIDIA open-sourced a foundation model, referred to there as Nitride, that can play more than 1,000 games; OpenAI was reported to be considering a $500 million acquisition of the game company Medal; xAI wants to have Grok compete against League of Legends professionals; and Google has released Genie 3 and SIMA 2 in succession.

Read only as headlines, "AI playing games" can sound like a tech demo, far from the version you have to deliver tomorrow. Behind the headlines is a more durable signal: AI companies are treating games as a main battlefield for real deployment. That can reshape toolchains, production pipelines, and even product form — not as a future possibility only, but as something already underway.

Why are AI companies so focused on games?

The reason matters because it tells you whether the investment is likely to last. The original article's answer is straightforward: games are currently one of the most suitable environments for training and validating AI.

DeepMind founder Demis Hassabis worked at a game company at 16, and a core idea behind DeepMind was to train intelligence in virtual worlds with clear rules and immediate feedback. From Atari to AlphaGo to AlphaStar, almost every major AI breakthrough in that lineage has had a relationship with games.

The article gives three reasons games are unusual. First, they offer a clearly ruled environment where an agent can try a decision millions of times; that is hard to do in the physical world, and it is what reinforcement learning and world models need. Second, games naturally produce large volumes of multimodal data — images, audio, text, player behavior, and physical interaction — that are unusually rich and structured for training multimodal models. Third, games are a vivid demonstration window. Whether a system can understand a complex environment, make reasonable decisions, and work with people is often more convincing in a match of Honor of Kings than in a paper. That is also why a public match between Grok and professional players has brand value for an AI company.

If that reading is right, investment in games is a long-term strategic bet, not a passing enthusiasm, and technology, tools, and resources will keep flowing into the industry. For production teams, the useful premise is that AI's presence in games is more likely to speed up than slow down. Ignoring it now makes later adaptation more expensive.

1. World models: still early, already close to pre-production

Most public discussion of world models stays at the technical layer: what the term means, whose model is stronger, and what kind of image it can generate. For game production, the more important change is how a "game world" gets built.

The traditional path is long and labor-intensive: design sets the rules, art makes assets, level artists build spaces, engineering writes logic, and QA tests the result. World models offer another possibility: generating an interactive 3D environment directly. The original article points to Google's Genie 3, which can generate dynamic, interactive game worlds from a text description in real time; Tencent's Hunyuan World, which can generate a 3D scene from a sentence or an image and supports real-time interaction and spatial memory; and Fei-Fei Li's World Labs, which is using spatial intelligence to reshape 3D content creation.

These systems are still early. The generated worlds are far from commercial-game standards in precision, stability, and controllability. The direction, though, is clear. They cannot yet replace a complete level-production pipeline, but they can already help in a few places:

• Faster prototype validation. Early on, design often needs to test whether a gameplay idea works in a particular space. That used to mean waiting for a blockout, from a few days to one or two weeks. If a world model can quickly generate a "roughly like this" interactive space, the validation cycle can shrink a lot. That is especially useful in greenlight and pre-production, when you want to try more directions at lower cost.
• A communication tool for scene concepts. Like concept art, a generated 3D scene can help a team align on atmosphere, spatial relationships, and lighting. Text descriptions are easy to misread; a 3D preview can reduce that gap.
• Training data and test environments. If a project needs to train AI NPCs or bots, a world model can generate many differently configured spaces. Automated testing also needs diverse scenes for boundary cases.

What they still cannot do is drop straight into a production pipeline. Game engines are strict about asset specs — polygon counts, UVs, LOD, collision, navigation meshes — and world models do not handle those. They are closer to an upstream aid than a replacement for the current pipeline. Hao's judgment in the original article is that over the next two to three years, world models in game production will be more "accelerator" than "replacement." Early validation, communication, and test coverage can get faster. Core asset production and engine integration will still need people.

2. AI inside the engine: a quieter change in the rendering pipeline

If world models change how a world is built from the outside, AI inside the engine changes how that world is presented from the inside. The original article argues that this direction may be the most underestimated.

For two or three decades, the rendering pipeline did not change in a fundamental way. NVIDIA's hardware ray-tracing units in 2018 were the first major upgrade. 2025, in the article's account, was the second: AI began to take part in rendering computation itself. The best-known example is NVIDIA's DLSS (Deep Learning Super Sampling). Instead of asking the GPU to compute every pixel the hard way, the frame is rendered at a lower resolution and AI fills in high-resolution detail. The result is higher frame rates and better image quality, with a lower hardware bar.

That is not only "the game looks better." It changes the logic of performance work. Technical artists and engineers used to optimize mainly by reducing load: fewer polygons, more compressed textures, fewer lights, tighter draw-call budgets. Those methods trade visual quality for frame rate. AI rendering offers another approach: you do not have to cut as aggressively at the source, because AI can recover some quality at the last step. Art still cannot ignore performance, but the trade-off space between image quality and performance gets wider.

Cross-platform work may also get lighter. Shipping smoothly on PC, consoles, and mobile used to require a lot of asset tiering and per-platform rendering strategy. If AI rendering can run on more hardware than high-end GPUs, some of that adaptation work may shrink. The article notes that Tencent's game-engine team is working in this direction: making AI rendering features that now run only on high-end GPUs light enough for more platforms and devices. For multi-platform projects, that is a concrete benefit.

The technical artist's skill stack also widens. When the pipeline includes AI modules, a TA may need to understand not only traditional shader work and engine rendering, but also how an AI rendering model works and how it is tuned. That is not a call for TAs to become AI researchers. The tools changed, so the people who use them have to keep up.

The practical limit is hardware coverage. At the time of the original article, DLSS 4 still depended heavily on AI acceleration units in NVIDIA's latest high-end GPUs. For mass-market games, especially mobile games, real availability is still limited by how widely that hardware is installed. Not every player can use an RTX 5090. In the near term, AI rendering is more a bonus for high-end experience than a default for every project. The trend is still clear enough that technical teams should start building the related knowledge early.

3. AI outside the engine: the heavy lifting is being taken over first

Compared with the rendering change inside the engine, AI has moved faster, more broadly, and more visibly in asset production outside the engine.

People outside games often underestimate how many steps sit between a character design and a character that can actually run in the engine: concept art, orthographic views, high poly, low poly, UV unwrapping, texture painting, rigging, skin weights, animation, then engine import and debug. Only a small part of that chain is truly creative judgment — concept, key poses, final review. A large middle stretch is repetitive, rule-bound, and very time-consuming.

Skinning is a typical example. The job is to tell the engine how each vertex should follow a bone. A simple character may take three to five days of manual parameter work; a complex character takes longer. The step creates almost no creative value, but without it the character cannot move.

The original article points to Tencent Games' VisWise tool family as one attempt to take on this kind of work. For skinning, AI can calculate weights and coefficients automatically, with efficiency gains of up to 8x — work that used to take three to five days can now be finished in around two hours. Similar shifts are already visible in other steps:

• Animation assist. Basic locomotion cycles (walk, run, jump), facial performance, and motion blending can already get a decent first pass from AI, which is much faster for an animator than starting from zero.
• Batch materials and textures. When a scene needs dozens of ground, wall, and foliage variants, AI can generate style-consistent variations from a few samples and cut repeated painting.
• Automated asset checks. Polygon budgets, naming rules, overlapping UVs, and texture-resolution specs used to be checked by hand, one asset at a time. AI can do that in batches.

The production-side effects follow from that. When heavy lifting shrinks, the art bottleneck moves from "hands are not fast enough" to "judgment is not sufficient": taste, style control, and creative direction become the new constraint. An art lead whose main value is assigning work, watching the schedule, and protecting delivery volume will see more of that value diluted. Taste and direction will weigh more.

Outsourcing logic may change with it. Games currently rely heavily on vendors for asset volume. If AI takes over some repetitive production, demand for purely labor-intensive outsourcing may fall, while demand for creative work and high-precision tuning may rise. For PMs who manage vendors, the evaluation standard may need to shift from "more people, faster, cheaper" toward "stronger taste, able to use AI tools well, and stable quality."

If a project decides to bring AI into production, the asset pipeline has to plan for it in pre-production: input and output formats, quality-check standards, and human review gates. This is not something to bolt on after mass production has already started.

4. Player experience: from scripted NPCs to AI-native interaction

The first three directions change how games are made. This one changes what a game can be.

Traditional NPC conversation is a fixed dialogue tree. Players pick A or B, and they only get as many branches as a designer wrote. Even a carefully written NPC usually shows its "bottom" after a few turns. A large-model-driven NPC is different. It can hold an open conversation based on character setup, world rules, and the player's history, and the reply is generated in the moment.

This is already in commercial products, not only in lab demos. NetEase's 逆水寒 has used intelligent NPCs at scale, with each NPC having its own "life" and behavior logic. Enemy AI in the shooter ARC Raiders can react dynamically to the fight. Honor of Kings has a Lingbao (灵宝) interaction feature that supports voice control across scenes.

AI is also moving into core gameplay, not only conversation. In Game for Peace (Peace Elite / 和平精英), the AI war dog Bruce is not just a command-following helper. It is a new tactical unit inside a battle-royale frame: it can understand natural-language orders, perform distraction, scouting, and rescue tasks, and affect overall strategy. The design point in the original article is that AI is no longer only an add-on feature. It is part of the core mechanic, and it changes the player's decision space. There is also an AI-coach direction: the Honor of Kings commander already gives players real-time tactical advice, and Delta Force and Wild Rift have similar AI voice-assistant systems.

For production teams, several roles change. Level and narrative design have long been "deterministic": the designer pre-authorizes every path and outcome. AI-driven interaction is closer to "emergent" design: the designer defines rules and character thinking, and what actually happens is only known at runtime. The important skill shifts from writing every line to defining how a character thinks and how the world runs. Designers need systems thinking, not only storytelling.

QA faces a new problem: how do you test a system that is different every time? Traditional cases assume that input A should produce output B. AI-driven output is not fixed. What you need to test is whether the system says things it should not say, whether its behavior stays in a reasonable range, and whether the player experience still hits the design goal. That calls for new methods: rule-based automated boundary checks, large-scale player-behavior simulation, and subjective experience evaluation. The original article says the industry does not yet have a mature solution.

Operations and safety pressure also rise. Generated content can be unpredictable: an NPC may say something that violates policy, create inappropriate content, or behave in an unintended way. That needs matching safety mechanisms — real-time filtering, behavior monitoring, and anomaly alerts — planned early by operations, not treated as a purely technical afterthought.

5. A structural mismatch: AI companies have technology, game companies have contexts

The original article isolates one structural problem. On one side, leading AI companies — OpenAI, Google, NVIDIA, Anthropic — have strong models and enough compute, but not much live experience of putting that technology into a real commercial game. Their demos can be impressive, and there is still a large engineering gap before the same technology sits inside a shipping project. On the other side, traditional game companies have users, mature IPs, and complete production pipelines, but often a weaker grip on frontier AI. Many are still at the stage of "let employees use ChatGPT to go faster," far from the deeper uses of current models.

The result is that a game form that truly matches this jump in AI capability has not really appeared yet. Companies that can stand in the middle have an advantage. The article uses Tencent as an example: it has the open-source Hunyuan model series and game-engine research on one side, and very large products such as Honor of Kings and Game for Peace as deployment contexts on the other. Most pure AI companies or pure game companies do not have both.

That mismatch is not only a giant-company issue. For mid-size and small studios, it can also be an opening. A small team that is good at using AI tools may ship more product than its headcount suggests. The article's illustration is that if a 10-person team can use AI in depth, its effective capacity may approach what used to take 30–50 people. The key is not whether you can train your own model. It is whether you can fold existing AI capabilities into your production process efficiently.

What production teams should think about now

These five directions are not independent. They can speed one another up: faster scene generation, more efficient presentation, cheaper asset volume, and richer NPC and content experiences. Taken together, the path from idea to a finished build can get much shorter.

The original article closes with three things teams can start thinking about now:

1. Look for the insertion point on your own pipeline. You do not need to follow every direction. Your project almost certainly has one step that is the most painful, the most time-consuming, and the best fit for an AI trial. Find it, and try it seriously.
2. Assess how unevenly the team already uses AI. Inside the same team, the gap between people can be very large. If that gap is not noticed and brought closer together, it becomes a hidden efficiency loss. A PM is no longer only managing human capacity, but the combined capacity of people plus AI.
3. Think early about product design, especially in greenlight or pre-production. Is there a natural place for AI in this product — NPC interaction, content generation, player assistance? You do not have to build it. You should have thought about it.

In a recent interview cited in the original article, Demis Hassabis said that one thing he still intends to do is apply general AI technology back to games and create the "ultimate game." That ultimate game is still far away. The way AI is changing game production has already started.

FAQ

Can world models replace level production today?
No. They can already help with early playable validation, team communication, and generating environments for training or testing. They still cannot meet engine requirements such as polygon counts, UVs, LODs, collision, and navigation meshes, so they cannot go straight into a mass-production pipeline. Over the next two to three years, they are more likely to be accelerators than replacements.

Should every project treat AI rendering as a default?
Not in the near term. It widens the trade-off between image quality and performance, and it may eventually reduce some cross-platform work, but features such as DLSS 4 still depend on high-end NVIDIA hardware. For mass-market and mobile games, coverage is still a limit. It is currently more a high-end bonus than a standard feature for every project.

Which production work is AI taking over first?
The repetitive, rule-bound middle of the asset chain: skinning, basic animation cycles, batch materials, and automated spec checks. In the VisWise example from the original article, skinning work that used to take three to five days can be finished in around two hours, with efficiency gains of up to 8x. Creative judgment, style control, and final review remain human work, and those become the new bottleneck as the heavy lifting shrinks.

If NPCs are model-driven, what changes for design and QA?
Design shifts from writing every line and pre-authoring every path to defining how a character thinks and how the world runs. QA can no longer rely only on "input A should produce output B." It has to test whether the system stays in bounds, avoids disallowed content, and still delivers the intended experience. The original article says a mature method for that does not yet exist in the industry, and operations need safety filters, monitoring, and alerts planned in advance.

Do small teams need to train their own models to benefit?
No. The original article's point is integration, not training. The structural gap is that AI companies have models but few live game contexts, while many game companies have contexts but only shallow AI use. A 10-person team that uses existing AI capabilities deeply may approach the effective capacity that used to require 30–50 people.

Author and source

Hao is a game project manager with about 10 years of experience, and the author of PMNote (Hao's Game PM Notes).

This English version is adapted from the public Chinese article, without adding later news, products, or figures that are not on the source page:

https://pmnote.ai/articles/five-ai-directions-game-industry/

The original article and other public notes are at https://pmnote.ai. If you want the WeChat mini-program entrance, search in WeChat for 「Hao的游戏PM笔记」.