Speaker:: Chase Hasbrouck Title:: Three Phases of AI Adoption Duration:: 23 min Video:: https://www.youtube.com/watch?v=UOtVmYR0mRg ## Key Thesis Enterprise AI adoption in large government organizations follows three sequential phases — access, cost, and culture — each requiring different solutions, and the hardest phase (culture) is the one currently underway. The US Army's multi-year procurement and accreditation cycles make it fundamentally slower at adopting AI than commercial counterparts, with bureaucratic constraints often defeating the technology itself. ## Synopsis Lt. Colonel Chase Hasbrouck, who runs forensics and malware analysis for US Army Cyber Command, drew on roughly two years of trying — largely unsuccessfully — to push government cybersecurity personnel toward AI tooling. His framing of "three phases" provides a candid operational autopsy of enterprise AI adoption at scale inside a one-million-person organization that includes part-time national guard and reserve personnel. **Phase 1 (2023) — Research Preview / Access Problem.** The initial military response was blanket prohibition (Space Force issued the first formal no-AI memo; other services followed informally). Power users like Hasbrouck worked around this by doing personal testing at home on commercial APIs, then translating insights back to work. The eventual approved tool, "Camo GPT" (the Army's internally accredited chatbot, built on Llama 2-70B), failed to gain traction for three reasons: model quality lagged far behind commercial frontier models; it lacked document parsing and web browsing; and availability was unreliable with query responses sometimes taking hours. The solution to the access problem was centralization — getting an approved, accredited platform in place at all. **Phase 2 (2024) — Enterprise Platform / Cost Problem.** A new platform arrived using commercial frontier model APIs with a reliable front-end. The problem shifted to token cost. Without differentiated power-user allocations, forensics analysts who needed to diff 20-page log files burned through their monthly token quota in one or two sessions. The only remediation process required seven high-ranking signatures from people who didn't understand what an LLM was, making it "effectively impossible." Shadow usage of personal AI Studio accounts continued. The solution came through enterprise agreements and DoD-wide access via genai.mil. **Phase 3 (Present) — Culture Problem.** Now that access and cost are largely solved through genai.mil connecting to major AI enterprise platforms, the remaining challenge is organizational culture. Hasbrouck describes the "silo reflex" — the Army's tendency to designate one expert per new technology and route all questions through that person, rather than distributing the capability broadly. He frames the core question as: will AI end up like VR (niche, stays in the box) or like PCs (pervasive, becomes everyone's tool)? As of his talk, it's trending toward the PC trajectory. A specific cultural artifact is that current requirements documents effectively exist as individual chat histories on LLM platforms, never formalized into institutional knowledge accessible to procurement. Hasbrouck also addressed questions about fine-tuning (his team sees little value given how fast procurement cycles lag model advances), reserve component access to genai.mil (a real gap being addressed via Azure Virtual Desktop and BYOD tools), and what he looks for in junior analysts — specifically the ability to translate technical findings into actionable "so what" statements for non-technical stakeholders. ## Key Takeaways - Enterprise AI adoption follows a predictable three-phase arc: access → cost → culture, each requiring a different solution approach - Government procurement and accreditation cycles (often 4+ years) fundamentally constrain AI adoption regardless of tool quality - Flat token allocation without power-user tiers actively penalizes the high-value use cases (forensics, log analysis) where AI provides the most ROI - The "silo reflex" — designating one expert rather than distributing capability — is a structural cultural barrier at large organizations - Fine-tuning models provides diminishing returns when general model advancement pace exceeds procurement cycle pace - The real knowledge gap is that operational requirements live in individual chat histories rather than institutional memory ## Notable Quotes / Data Points - Army's total workforce: over 1 million when counting active duty, National Guard, and civilian employees — roughly 1/3 of which are part-timers - Typical AI procurement path: 2-year procurement cycle + 2-year accreditation cycle = 4 years, by which time stakeholders have rotated and may reject the already-approved tool - Camo GPT used Llama 2-70B with a front-end that limited clipboard paste to ~400 characters, effectively blocking non-power-users immediately - "A lot of our requirements documents right now, they exist as people's chat histories on various LLM platforms" - Hasbrouck is retiring October 26 and exploring what's next #unprompted #claude