Ask Sage — Army AI Integration (Part 1)
From solo experiment to Division-wide AI program — pioneering generative AI adoption across the 78th Training Division and the U.S. Army Reserve.
May 24, 2026
Military, AI Strategy, Applied AI Design, Training & Change Management, Operational Integration
The Legacy Process
The Army Reserve's training evaluation process hadn't fundamentally changed in decades. During large-scale exercises, Observer/Controllers (OC/Ts) and Training Assessment Framework (TAF) Analysts serve as the senior mentor staff — observing unit performance in real time, documenting findings, conducting analysis against doctrinal standards, and producing formal assessment products that feed into After Action Reviews and unit readiness reporting.
Every step of that workflow was manual. Analysts observed, took notes by hand or in unstructured digital formats, synthesized observations against multiple doctrinal references, drafted narrative assessments, routed them through review chains, and ultimately deposited finished products into repositories. The time between observation and finished product was governed entirely by human throughput — and during peak exercise tempo, that created a bottleneck that degraded the quality and timeliness of feedback reaching the units being trained.
The Gap
Nobody in the training enterprise was asking whether AI could accelerate this process. Not because the technology wasn't mature enough — but because the institutional awareness simply wasn't there. There was no enterprise AI capability. No infrastructure. No policy framework. No SOPs governing use. No designated owner. The concept of integrating generative AI into Army Reserve training operations existed in exactly zero documents, at any echelon, anywhere in the command.
The Tasking
The 78th Training Division Command designated me as the lead for AI integration into the training enterprise. The scope was deliberately broad: assess the landscape, identify where generative AI could create measurable value in the training evaluation process, select a platform, build the capability, and operationalize it — from concept through adoption.
No team. No inherited program. No vendor already under contract. No budget line item. Just a mandate, a timeline, and the latitude to figure it out.
Platform Selection: Why Ask Sage
The first decision was foundational: what do we build on?
The requirements were non-negotiable. The platform had to operate within a government-authorized security environment. It had to meet DoD compliance standards for handling controlled unclassified information. It had to support multi-user access at scale across geographically dispersed training sites. And it had to offer sufficient configurability to support custom prompt architectures, persona frameworks, and structured data workflows — not just a general-purpose chat interface.
Ask Sage met every requirement. It's a secure, FedRAMP-authorized AI platform purpose-built for defense and government environments. It gave us the compliance posture that would have eliminated any commercial alternative from consideration, while still offering the architectural flexibility to build a custom operational program on top of it.
But I want to be clear: Ask Sage is infrastructure. Selecting the platform was step one of a much larger build. What the training enterprise needed wasn't a tool — it needed a methodology.
Program Architecture: Four Components
I designed the entire program architecture personally, built around four interdependent components:
1. Standard Operating Procedure (SOP)
Not a user manual. A battle rhythm. The SOP defined how evaluators would interact with the platform across every phase of a training exercise — pre-exercise preparation, live observation support, analytical synthesis, product development, and post-exercise reporting. It mapped AI touchpoints to existing operational workflows rather than asking users to adopt an entirely new process. The principle was integration, not replacement.
2. Prompt Library
A curated library of mission-specific, pre-built prompts designed for the exact analytical tasks the senior mentor staff performs. These weren't generic prompts pulled from a template — they were purpose-built for training assessment contexts: doctrinal analysis, observation synthesis, AAR narrative development, performance trend identification, and cross-functional pattern recognition.
Each prompt was structured with embedded context framing, output formatting instructions, and role-specific parameters. The goal was to eliminate the "blank page" problem — no user should ever need to figure out how to talk to the AI. The prompt library gave them a starting point calibrated to their specific function.
3. Persona Framework
Different evaluator roles require fundamentally different AI behaviors. An OC/T conducting a battalion-level tactical assessment needs the AI to operate differently than a TAF analyst compiling quantitative data for a brigade-level trends report. A senior mentor facilitating an AAR needs different output structures than an analyst building a repository product.
The persona framework configured distinct AI behavioral profiles for each role — adjusting tone, output format, analytical depth, doctrinal reference weighting, and response structure based on who was using it and what they were trying to accomplish. This wasn't cosmetic personalization. It was functional differentiation that made the platform immediately useful to users with very different jobs.
4. Data-Cloud Structure
The organizational backbone. This component defined how users, data inputs, AI outputs, and finished products connected and flowed within the platform — structured to mirror how Army training exercises actually operate, not how a platform vendor assumes they operate.
This included user access hierarchies mapped to exercise organizational structures, data segregation between training audiences and evaluator staff, output routing aligned to existing review and approval chains, and repository integration for finished product storage. The architecture had to accommodate the reality that training exercises involve constant personnel turnover, shifting organizational relationships, and multiple concurrent assessment lanes — all running simultaneously across dispersed geographic locations.
Timeline and Execution
Nine months from initial tasking to operational deployment. Solo build throughout the design phase. No development team, no contractor support, no external consultants. The advantage of that constraint: every architectural decision was internally consistent because one person held the entire system in their head. The disadvantage: the pace was relentless.
Training at Scale
A capability nobody knows how to use is a capability that doesn't exist. The training program was designed with the same rigor as the platform architecture itself.
We onboarded 320+ OC/Ts, senior trainers, TAF analysts, and training-unit leads across both the 78th and 84th Training Commands in the first year. The training methodology was layered: conceptual overview sessions first (what is this, why does it matter, how does it fit into your existing workflow), followed by hands-on practical exercises with live platform access, followed by scenario-based application (use the tool to complete an actual analytical task under simulated exercise conditions), followed by troubleshooting and refinement based on real user feedback.
The standard was not "they can log in." The standard was "they reach for it instinctively during an exercise because it makes their job measurably better." That's a fundamentally different training objective, and it requires fundamentally more time, repetition, and user-centered design in the training approach.
Quantitative Results
The first exercise to use the capability delivered a measured 47% improvement in analyst product throughput — defined as both the volume of products completed and the time-to-repository for each finished product. That metric was collected through structured time trials comparing baseline manual process performance against AI-assisted performance under identical task conditions.
47% is not a marginal gain. In an environment where every hour of analytical delay degrades the quality and timeliness of feedback reaching training audiences, that improvement translated directly into better training outcomes for the units being evaluated.
Institutional Recognition
The Commanding Brigadier General awarded me the Army Commendation Medal for the development and fielding of the capability. The Commanding General of the training enterprise publicly designated me the command's subject-matter expert on enterprise generative AI adoption — a distinction that carried institutional weight in subsequent conversations about scaling the program.
Colonel-level division and group staff across both training commands provided sustained recognition, and the program became a recurring reference point in command-level discussions about modernization and technology integration.
Strategic Elevation
The result that mattered most for the program's future: selection as an AI integration advisor to the 84th Training Command's AI Integrations Working Group. This was a small, deliberately convened group operating under a Colonel-level charter to shape command-wide AI strategy and policy.
That selection represented a shift. The program had graduated from a single-division initiative to a command-level strategic consideration. It moved from "Trey's project" to institutional priority — which was always the intended end state.
Foundation for What Followed
Everything that came after — Operation Sentinel Justice, the 97% efficiency results, the partnership with the 75th Innovation Command, the conversations about USARC-wide training reform — traces directly back to this nine-month build. Without the SOP, the training architecture, the prompt library, and the field-proven results from this initial phase, none of the subsequent scale was possible.
The Procurement Problem
The single biggest hurdle had nothing to do with AI, technology, or user adoption. It was license acquisition.
Figuring out the correct echelon to purchase the Ask Sage enterprise account — who owns the contract, which funding mechanism applies, what approval authorities are required, which staff section processes the request — consumed more time and organizational energy than any technical challenge in the entire build. The Army's procurement infrastructure wasn't designed for rapid acquisition of a platform nobody had heard of six months earlier. Every conversation required re-explaining what Ask Sage was, why it mattered, and why existing tools couldn't accomplish the same objective.
Navigating that bureaucracy while simultaneously designing and building the program tested patience in ways that the technical work never did. The lesson: in large institutions, the administrative friction of acquiring the tool will frequently outpace the technical difficulty of implementing it. Plan for that. Budget time for it. And start the procurement conversation as early as humanly possible — ideally before you need the answer.
The Scalability Gap
If I started this build over tomorrow, I'd begin with a scalability framework already in place before writing a single prompt or configuring a single persona.
The reality is that I built this program iteratively — piece by piece, problem by problem, off the cuff — without a comprehensive strategic document underneath it. It worked. The results prove that it worked. But there were moments where I was clearly building the plane while flying it, and the bolts that should have been tightened before the first user logged in had to be tightened mid-flight instead.
The strategic charter, tactical rollout plans, and operational documentation that eventually came out of this program should have existed before deployment — not after. The lesson isn't that iteration is wrong. The lesson is that iteration without a strategic scaffolding creates technical debt that compounds faster than you expect, especially when adoption accelerates beyond your original planning assumptions.
Designing for a Moving Market
The third realization hit mid-build and reshaped the entire design philosophy: the AI market was moving faster than the program.
What was the best available platform and approach in March 2023 might not be the best option in March 2024. Model capabilities were advancing quarterly. New platforms were entering the government-authorized space. Pricing structures were shifting. Feature sets were expanding. If I built a rigid, platform-dependent solution — one that could only function on Ask Sage, structured in ways that only Ask Sage could support — the program would have an expiration date.
That forced a fundamental design principle into the architecture: transferability. The SOP, prompt library, persona framework, and data structure all had to be designed in a way that could migrate to a different platform if the landscape demanded it. The program's value had to live in the methodology, not the tool. That wasn't in the original plan. It became a non-negotiable requirement once I understood the pace of change in the market.
Reflection
Building something from nothing inside a large institution requires equal parts vision and stubbornness. The vision identifies what's possible. The stubbornness gets you through procurement delays, bureaucratic friction, and the hundred small moments where it would have been easier to build the chatbot everyone expected and call it done.
What I built in nine months wasn't just a tool implementation — it was an operational program with doctrine, training, architecture, and measured results. The 47% throughput improvement validated the capability. But what I'm most proud of is the infrastructure beneath the metric — the transferable design, the training methodology, the documentation rigor — because that's what gave the program institutional legs.
The Principle
Build it right. Train it thoroughly. Design it to outlast the tool it runs on. And start the procurement paperwork before you think you need to.
This project sits at the center of AI Strategy & Roadmapping, Applied AI Design, Operational Integration & SOP Development, and Training & Change Management — the same combination I bring to every problem worth solving.