meh.. not bad... | Ryan Connell

meh.. not bad...

I saw this online from Ryan Connell. Not perfect but just shows how much AI progressed in our field over the last few months.

Out of curiosity, I just asked the same question for me

That’s a great and very timely question — buying AI in the U.S. Department of Defense (DoD) is tricky because traditional acquisition processes (like the FAR/DFARS) weren’t designed for rapidly evolving technologies like AI. The best approach depends on what kind of AI capability you’re acquiring (e.g., COTS tools, R&D, or operational systems), where you are in the lifecycle, and how fast you need it.

Here’s how I’d recommend approaching it — broken into five key strategies used successfully across DoD programs

1. Start with the problem, not the tech

Before buying “AI,” clarify:

• What decision or task are you trying to improve or automate?

• What data do you have, and is it usable?

• What level of autonomy or human oversight is acceptable?

This shapes not only the contract but also the risk posture, data rights, and testing approach.

2. Use flexible acquisition pathways

The DoD has several tools built for emerging tech like AI:

✅ Other Transaction Authority (OTA):

• Great for prototypes and rapid iteration.

• Lets you skip many FAR constraints.

• Use a follow-on Production OTA if the prototype succeeds.

• Commonly used by DIU, AFWERX, Army AI Integration Center, etc.

  
  

✅ Software Acquisition Pathway (per DoDI 5000.87):

• Designed for continuously delivered software like AI models.

• Allows iterative delivery, frequent updates, and ongoing user feedback.

✅ Middle Tier of Acquisition (MTA):

• For fielding mature prototypes within 2–5 years.

• Works well for scaling an AI model after successful demonstration.

✅ Commercial Solutions Opening (CSO):

• For acquiring commercial AI products with limited competition.

• Used by organizations like AFWERX and the Defense Innovation Unit (DIU).

3. Prioritize data access and lifecycle management

AI’s success depends more on data than on algorithms. When contracting:

• Specify data ownership, usage rights, and labeling requirements early.

• Require access to training, validation, and operational data.

• Include Model Lifecycle language — retraining, monitoring, and explainability.

• Consider AI Assurance and responsible AI requirements (DoD AI Ethics Principles).

4. Partner early with the right stakeholders

• Chief Digital and AI Office (CDAO) – offers AI contracting and MLOps guidance.

• Service-level AI offices (e.g., Army AFC AI2C, Navy N9I, AF Chief Data & AI Office).

• DIU or AFWERX for rapid prototype and transition pathways.

• Defense Innovation Marketplace and SAM.gov for ongoing opportunities.

5. Build contracts for adaptability

AI evolves fast. Contracts should:

• Allow for data and model updates over time.

• Include performance-based metrics, not fixed specifications.

• Support agile development cycles (e.g., monthly sprints).

• Encourage vendor collaboration rather than one-off delivery.

Example Acquisition Path

If you’re, say, buying an AI-enabled predictive maintenance tool for aircraft:

1. Prototype: Use a DIU or AFWERX OTA with clear data-sharing clauses.

2. Transition: Evaluate and then issue a follow-on Production OTA or MTA effort.

3. Scale: Move into the Software Acquisition Pathway for sustainment and updates.

If you’d like, I can show you example language from successful AI-related solicitations (e.g., DIU’s Predictive Maintenance CSO, or CDAO’s Tradewinds AI procurement templates) or outline a step-by-step acquisition plan for your specific use case (e.g., AI analytics, computer vision, LLMs, etc.).

Here's my concern with applying AI to government acquisitions - it's drawing from what we've already done, it is not creating anything new/novel. And what we've done/been doing over the past two decades isn't exactly great. The result is that AI is going to spit out the same or similar products and we're going to get the same or similar results. Is that what we need right now? To get to mediocrity (or worse) faster? I don't think so, but I'm just speaking for myself.