Company World Model
Train AI on your proprietary data so your model becomes a durable, owned company asset
The Company World Model treats an AI model as a strategic business asset rather than an API dependency. Instead of routing every query through expensive frontier models that retrieve context at runtime, companies post-train a custom small language model on their proprietary data—embedding institutional knowledge directly into the model's weights. The trained model answers questions about company history, processes, and relationships instantly at 3–10x lower cost per token. A continuous reinforcement learning loop updates the weights overnight, keeping the model current with new business activity. Companies deploy multiple specialized sub-models by department, resulting in faster agents, dramatically lower inference costs, and an AI asset the company fully owns.
- Company knowledge embedded in model weights is up to 100x cheaper and faster than retrieved at runtime
- AI should be an owned asset on the balance sheet, not a rented API dependency
- Proprietary data is a competitive moat—bake it into your model rather than expose it to third-party APIs
- Continuous RL keeps models current without expensive full retraining cycles
- Specialized sub-models by domain outperform a single generalist company model
- Audit all proprietary data sourcesMap every internal data repository: Slack channels, email archives, CRM records, support tickets, documentation, and quarterly reports. Estimate total data volume and historical depth to scope the training effort.Pro tipPrioritize data sources that reflect how decisions are made and how customers are served—these will drive the most useful reward signals during post-training.
- Benchmark current frontier model inference costsRun your top five most frequent AI workflows and log token usage and cost per run. This baseline is your primary ROI anchor and determines which workflows justify custom training first.WarningWithout a clear cost baseline you cannot make the business case internally or validate improvements after deployment.
- Select an open-source base modelChoose a small language model (e.g., Kimi, DeepSeek, Llama) that fits your compute budget and domain requirements. The base model's quality sets the ceiling for your post-trained model's performance.Pro tipEvaluate candidate models on a representative sample of your proprietary queries before committing to a base model for full post-training.
- Post-train on proprietary data using reinforcement learningFine-tune the base model on your company's historical data using RL. Define reward functions tied to business outcomes—accuracy on internal queries, correct entity resolution, task completion rates.Pro tipStart post-training on a single high-value domain such as customer support history before training on all company data at once.WarningPoorly designed reward functions will train the model to optimize for the wrong outcomes. Invest time defining what 'correct' means for your specific business context.
- Build a continuous learning pipelineSet up an automated nightly job that ingests new interactions, decisions, and documents from the day and updates the model weights, keeping the model current without manual retraining.WarningSkipping continuous learning means your model reflects your company as it was, not as it is—knowledge goes stale within weeks without nightly weight updates.
- Deploy specialized sub-models by domainTrain separate or branched models for distinct business units—customer support, sales, operations, engineering—so each sub-model is highly tuned to its domain vocabulary and task patterns.Pro tipA dedicated sub-model for a specific domain can use far fewer parameters and still outperform a large generalist model on in-domain tasks.
- Measure and iterate on cost and performance ROICompare cost-per-query, response latency, and task accuracy against the frontier model baseline. Use findings to prioritize which workflows to migrate into model weights next.Pro tipTrack peak-cost workflows specifically—these yield the largest savings and the strongest ROI case for expanding the program.
Corgi, an AI-native insurance carrier, was spending heavily on frontier model inference for repetitive workflows. Aragon deployed a custom SLM trained on Corgi's coverage logic and client profile data. Because the model already knew Corgi's policies and startup client patterns in its weights, it no longer needed to re-read documentation on every query. Inference costs dropped to roughly one-third of prior frontier model spend, with dramatically faster response times for the quote-generation workflow.
During a live demo, a partner sent a request to draft an MSA and 'send it to Josh' without specifying which Josh. The Aragon model correctly resolved the ambiguous name by cross-referencing recent calendar events and Slack conversations embedded in its nightly-updated weights. A generic frontier model would have asked for clarification or queried all contacts named Josh. The trained company context eliminated the disambiguation step entirely, illustrating the practical lift of institutional knowledge baked into weights.
Extracted from This Week in Startups (E2278) featuring Josh Cerot, CEO of Aragon, who coined the term 'company world model' to describe training company-specific SLMs on proprietary data via reinforcement learning.