GAM.01
NPC & Dialogue Agents
LLM NPCs that talk freely inside writer-owned lore, change game state only through designer-exposed actions, and fall back to authored lines when off-script.
Open pageIndustry
Game AI,
inside the design.
Writers set the limits. The model plays inside them. NPC dialogue that stays in character, localisation that keeps the joke, and player support that knows when to hand over. Built into Unity, Unreal or your own engine, with writers and leads setting the limits.
Last reviewed 30 October 2026 · Engineering guidance, not legal or regulatory advice.
In context
AI in gaming has two meanings: classic game AI (behaviour trees, pathfinding, difficulty tuning) and generative AI used in development and live operations. We build the second kind, inside limits your designers and writers set.
Where it breaks today
01
An NPC off-script is a bug
A character that breaks lore, spoils a quest or says something unrateable is a defect players screenshot. Plugging a chat model into a character without limits trades a dialogue tree for a moderation problem.
02
Strings ship faster than translators
Live games push text every patch, across a dozen languages, with variables, gender and plural rules, and UI limits. Machine translation without context breaks the build or the joke.
03
Support queues spike on launch day
Refunds, account recovery, bug reports and harassment reports all land at once, and the urgent ones look exactly like the routine ones until somebody reads them.
OUR APPROACH
Generative AI for game development, with authored edges
Games are authored experiences, so the models we build into them are constrained by authorship. An NPC speaks from a character sheet and a lore base the writers own, and can only change game state through actions the designers expose. A localisation pass works from your glossary and style guide, and a translator approves every string. A support agent resolves what it is allowed to and hands the rest to a person with the history attached.
The constraints are practical as well as creative. Platform holders and storefronts ask about AI use: Steam’s content survey requires developers to disclose generated content. Performer agreements, including SAG-AFTRA’s Interactive Media Agreement, carry consent and disclosure terms for digital voice replicas. Age ratings from PEGI and the ESRB apply to what a character can say. If players can talk to each other, the EU Digital Services Act and the UK Online Safety Act may apply. We design for those limits; deciding how they apply to your title is for your producers and legal team.
HOW IT DEPLOYS
PHASE 01
Pick one system
One character, one language pair or one support queue. We agree what good looks like with the leads who own it before any model is chosen.
PHASE 02
Build in your pipeline
Integrated with your engine, string tables, TMS and help desk, behind a feature flag, with latency and cost budgets per call.
PHASE 03
Playtest and widen
Adversarial playtests and review samples gate each rollout. The scope widens only when the people who own the system sign off.
What we build
01
NPC and Dialogue Agents
LLM-driven characters that answer players in voice, inside a lore base and a set of allowed actions, with filters tuned to the title’s rating and a scripted fallback for every failure.
SEE NPC DIALOGUE →02
Game Localisation
Context-aware translation of game text, with glossary and placeholder checks, length limits and translator review in the tools your localisation team already uses.
SEE GAME LOCALISATION →03
Player Support
Agents that resolve routine tickets in the player’s language and route harassment, child-safety, payment and account-takeover cases straight to people.
SEE PLAYER SUPPORT →04
Studio Tooling Prototypes
Internal tools for designers and producers, such as quest-text checkers, bug-report triage and build-note summaries, prototyped in weeks and kept only if the team uses them.
SEE RAPID PROTOTYPING →Games & interactive use cases
-
-
GAM.02
Game Localisation
Context-aware translation drafts and automated LQA checks for game text, in your engine and TMS workflow, with a translator approving every string that ships.
Open page -
GAM.03
Player Support
Multilingual player support that resolves routine tickets and screens first for safety, payment and account-takeover cases, which go to trained people.
Open page
Questions
Questions about Games & Interactive.
How is AI used in gaming?
In two ways. Classic game AI controls non-player characters through behaviour trees, state machines and pathfinding, and adjusts difficulty to how a player is doing. Generative AI is newer and mostly used around development and live operations: coding help, localisation, prototype assets, player support and, increasingly, conversational NPCs. We build the generative layer, constrained by what designers and writers author.
What does AI mean in gaming?
Traditionally it means the logic that makes computer-controlled characters behave believably: moving around obstacles, reacting to the player, choosing tactics. It is designed, deterministic and tuned by hand. Generative AI, meaning language and image models, is a separate thing that studios now use for dialogue, text and tooling. Both can sit in the same game, but they are built and tested very differently.
Can generative AI NPCs stay in character?
Mostly, if they are built to. The character speaks from a writer-owned profile and lore base, can only take actions the game exposes, and passes every line through filters set to the title’s rating. When a player pushes it off-script, it falls back to authored lines. Adversarial playtesting before launch is what tells you how often that happens.
Do we have to disclose AI use in our game?
Often, yes. Steam asks developers to disclose AI-generated content in its content survey, and other storefronts have their own rules. Voice performer agreements can require consent and disclosure for digital replicas. In the EU, the AI Act requires telling people when they are talking to an AI system unless it is obvious from context. Your producers and legal team decide what applies.
What are the best AI tools for game development?
It depends on the job. Coding assistants help programmers, image models help concept work, and engine vendors are adding their own features. A custom build makes sense when the work depends on your own content, such as lore, string tables or player history, or has to run inside your engine and support stack. That is the part we take on.
