What Building AI Tools for MCEN Taught Us About Constraints
Every vendor demo assumes a network you don't have. Here's what actually happens when an AI tool meets MCEN — and how it changed the way we build.
The demo network doesn't exist
When we built the CGIP AI Assistant — a tool that generates inspection checklists with direct citations from MCO and NAVMC directives — the AI part worked quickly. Source-grounded generation, a React front end, an API-driven language model: on a commercial network, that architecture is a weekend pattern.
Then it met the real environment. Restricted cloud access. Serverless runtimes blocked inside AFPIMS. No administrator rights on the workstations that needed the tool most. None of this is a complaint — those controls exist for good reasons. But it means the standard commercial AI architecture simply does not deploy where Marines actually work.
Constraints are requirements, not obstacles
The instinct is to treat network restrictions as friction to argue with: request exceptions, wait on approvals, escalate. We took the opposite lesson. If the environment forbids a backend, the tool shouldn't have one. Our current generation of DoD-facing tools is built on three rules:
- No install, no backend, no admin rights. A static browser file that runs from a workstation as-is. If it needs an approval chain to execute, it will die in that chain.
- Data stays where it lives. No raw uploads to external services. Documents are processed in the browser; nothing leaves the machine without the user deciding it should.
- Meet the approved AI where it is. Instead of bundling a model, tools generate structured, source-grounded prompts for GenAI.mil — the workflow that's already sanctioned — with a manual export fallback for the most restricted environments.
What this buys the unit
A tool built this way needs no ATO negotiation for a hosted service, no sustainment contract for a server, and no waiting on a cloud environment decision. It works the day it's copied onto the machine. Our Local Turnover Generator and Mission Checklist Assistant are both built on this pattern — direct descendants of what the CGIP build taught us.
The broader point for anyone buying AI for constrained environments: ask the vendor to demo on your network, not theirs. The architecture that survives that demo is the one worth funding.