Grayhat Developers describes an internal trial of self-hosted AI in its case study. The company kept commercial coding agents for complex work and assigned local models simpler tasks, including finding organizational information and synchronizing project records. The preprint was submitted to arXiv on November 5, 2026.
Which work went to the local models?
The company started by trying local models as replacements for commercial coding agents. The team found their performance and inference speed insufficient for that role. It then split the work. Commercial agents remained responsible for tasks that demanded stronger coding or reasoning. Local models handled lower-complexity requests and coordination with self-hosted tools such as Outline, Plane and ERPNext.
The study treats that split as part of the architecture. The same model choice can have different costs when it drafts code, searches internal notes or updates a task. The authors describe using the task’s capability requirements and context to decide which part of the stack should handle it.
What did the company estimate?
The authors report about $325 per month in commercial AI spending at the start. They estimate that 30–40% of commercial workload involved activities such as transferring context, synchronizing documentation and managing tasks. If local inference can handle about 90% of that coordination, they estimate it would displace 27–36% of overall commercial AI workload or cost-equivalent usage.
Why deployment design mattered
The first internal deployment struggled to gain adoption because the workflow introduced friction. A centrally configured version improved feedback and adoption in the company’s account. That makes the case useful for teams comparing self-hosted options. The paper measures the trade-off as a change to everyday work as well as a choice of model. Its estimates come from the company’s own workflow, and a budget comparison needs to account for the cost of operating local inference.
Comments