Back office ops · Production

Ubisoft uses AI21 Labs to automate NPC dialogue variation generation for game scriptwriters

The problem

Ubisoft scriptwriters faced extremely tedious manual work generating large numbers of NPC dialogue variations (bark trees), and lacked the fine-tuning data needed to automate the process internally — creating it by hand was wholly impractical.

First attempt

Other LLM providers introduced legal compliance friction and charged for both input and output tokens, making large-scale data augmentation experiments prohibitively expensive.

Workflow diagram · grounded in source
1
Writer needs bark tree variations
trigger
“writing bark trees which are standalone NPC lines that require as many variations as possible (and allowed by the voiceover budget)”
2
AI21 generates dialogue suggestions
ai_action
“Instead of having to painstakingly brainstorm 5-10 different ways to express motivation (hunger, in this example), Ubisoft's internal models could use AI21's output as a launch pad”
3
Writers review and curate output
human_review
“It's easier for writers to thumbs up and edit or thumbs down than to come up with novel and diverse training data themselves”
4
Pairwise comparison refines models
feedback_loop
“The writers are shown pairwise comparisons of outputs so the models can get further trained through human feedback”
5
NPC generator embedded in tools
output
“Ubisoft embeds the NPC generator and other similar models they've trained into the writing tools for their scripters”
Reported outcome

Ubisoft achieved rapid scaling of NPC content production, unlocking an unlimited fountain of training data while optimizing costs with a 30:1 input-to-output token ratio under AI21's output-only pricing.

Reported metrics
Input to output token ratio30:1
Content production scalingrapid scaling
Training data availabilityunlimited fountain of training data
Reported stack
AI21Wordtune
Source
https://www.ai21.com/blog/ubisoft-case-study
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Ubisoft achieved rapid scaling of NPC content production, unlocking an unlimited fountain of training data while optimizing costs with a 30:1 input-to-output token ratio under AI21's output-only pricing.

What tools did this team use?

AI21, Wordtune.

What results were reported?

Input to output token ratio: 30:1; Content production scaling: rapid scaling; Training data availability: unlimited fountain of training data (source-reported, not independently verified).

What failed first in this deployment?

Other LLM providers introduced legal compliance friction and charged for both input and output tokens, making large-scale data augmentation experiments prohibitively expensive.

How is this back office ops AI workflow structured?

Writer needs bark tree variations → AI21 generates dialogue suggestions → Writers review and curate output → Pairwise comparison refines models → NPC generator embedded in tools.

Related back office ops cases