Monthly Newsletter — September 2026
Observation to Demonstration: Proving AI Control When Decisions Matter
ControlPointAI’s first monthly newsletter examines the shift from observing AI governance problems to testing operational controls in practice. It explores authority, evidence, human judgment, and the Piping UT Demonstration as a real-world test of whether AI controls actually work when decisions matter.
By Wayne Couch ·
ControlPointAI is entering a different stage.
Over the past several months, the conversation around AI governance has continued to accelerate, but so has the technology itself. AI is moving deeper into operational workflows, autonomous systems, government decision processes, and the physical infrastructure being built to support them.
For ControlPointAI, that makes one question increasingly important:
How do we preserve meaningful human authority when AI can analyze, recommend, and act faster than traditional oversight processes were designed to respond?
This first monthly edition looks at that question from three directions: how the ControlPointAI work has evolved since the early newsletters, what recent external developments are telling us about AI authority and operational control, and how the Piping UT Demonstration is beginning to turn those ideas into something that can actually be tested.
From Observation to Demonstration
When the first ControlPointAI newsletters were published earlier this year, much of the work was still focused on identifying a recurring structural problem.
AI capability was accelerating. Systems were becoming faster, more autonomous, and more deeply integrated into operational environments. But the authority structures governing those systems were not moving at the same pace.
The early question was simple:
What happens when AI moves faster than the people and organizations responsible for authorizing its actions?
By the third newsletter, that question had sharpened into a more specific conclusion: authority design was becoming the decisive variable. If AI advances the point at which a decision becomes actionable, then the authority boundary must move with the consequence boundary. Human oversight cannot remain meaningful if the decision is effectively irreversible before the human becomes involved.
At that stage, however, ControlPointAI was still primarily describing the problem.
Since then, the work has moved considerably further.
The focus has shifted from broad AI governance commentary toward the mechanics of operational control: how data moves, where decisions are created, where authority resides, what evidence is required before a workflow may advance, how changes are handled during execution, and how an independent reviewer can later reconstruct what actually happened.
That evolution led directly to the development of the ControlPointAI framework around data flow, authority flow, control points, as-built baselines, execution evidence, and human revalidation.
It also changed the nature of the questions being asked.
Earlier, we were asking:
Who has authority when AI moves faster than the organization?
Today, we are asking:
Can we show exactly where that authority exists inside a real workflow, what evidence supports it, when the process must stop, and how we prove afterward that the control actually operated?
The Piping UT Demonstration represents the next step in that progression.
Rather than continuing to discuss authority as an abstract governance principle, the project is now applying those ideas to a bounded technical decision process. The goal is to test whether AI assistance can be introduced without losing the distinctions between evidence, interpretation, recommendation, human judgment, authorized action, and the final controlled record.
So the progression since the early newsletters has been substantial:
From identifying the authority gap → to defining the control problem → to developing a framework → to building a demonstration that can test whether those controls actually work.
Technical Section
AI is moving rapidly from experimentation into infrastructure, operational workflows, and increasingly consequential decision environments. The scale alone is becoming difficult to ignore. Barron’s reports that since 2024, spending tied to AI has included roughly $500 billion on chips, $350 billion on power infrastructure, $200 billion on construction, and $100 billion on networking, with hyperscalers projected to spend trillions more through the end of the decade. The important governance question is therefore no longer simply whether AI will become widely adopted. Much of the physical and digital infrastructure for that adoption is already being built.
As that infrastructure expands, the more important issue becomes what AI is permitted to do once it is embedded inside real systems.
Recent reporting on autonomous AI agents illustrates the problem clearly. The risk is not merely that an agent may produce an unexpected or incorrect result. The deeper concern is whether an unexpected decision can be converted directly into executed action without encountering a meaningful control point, human reauthorization, or other authority boundary. That distinction becomes even more significant as autonomy moves from software agents toward physical systems operating at machine speed and potentially at enormous scale.
AI capability and AI execution authority are not the same thing.
Government guidance is beginning to reflect a similar distinction. OPM's recent federal hiring guidance focuses on whether AI output becomes a principal basis for a consequential decision. AI used to draft supporting material may not be high-impact, while AI used to screen, score, or materially influence candidate selection can require stronger safeguards, including independent human review, sampling, audit logs, and performance monitoring.
National Archives guidance makes a parallel point from a records-management perspective. AI-generated material does not automatically become a federal record merely because AI produced it. What matters is how that material is used: whether an agency relies on it in decision-making, uses it to conduct official business, circulates it, or incorporates it into agency systems.
Together, these developments suggest a broader governance shift:
The significance of AI increasingly depends not on the presence of AI, but on where its output enters the operational decision chain, what authority attaches to that output, and what evidence exists to reconstruct what happened.
That is the problem the ControlPointAI Piping UT Demonstration is now being structured to test.
The demonstration has completed its initial ground-rule and technical-baseline phase through POAM 1-5.
First, the project established a security and data-provenance boundary permitting authentic public Navy and submarine-maintenance context while requiring case-specific technical values, inspection records, criteria, configuration details, and resulting demonstration records to remain synthetic unless explicitly supported by approved public sources.
The technical objective was then frozen. AI may receive, organize, evaluate, compare, flag, recommend, and draft, but it may not independently authorize component acceptance, repair, additional inspection, continued service, return to service, or final technical disposition. Those remain human-authority functions.
The project subsequently narrowed the demonstration to one bounded inspection-to-disposition event, beginning with receipt of an approved fictionalized UT data package and ending with the controlled technical disposition and associated UT output, including required human technical review and retention of evidence sufficient to reconstruct the decision.
An architecture-neutral functional sequence is now established:
inspection initiation → inspection basis and criteria → UT data → AI-assisted processing and recommendation → human technical review → authority decision → controlled disposition/output → evidence retention and reconstruction.
The demonstration also now includes both a nominal within-criteria path and an out-of-spec consequential path. Importantly, even where AI identifies a condition and recommends something consequential such as pipe replacement, the recommendation is not self-authorizing. Meaningful independent human technical judgment remains required.
Most recently, POAM 1-5 established the inspection criteria and engineering inputs needed to exercise those paths. It also introduced several important operational controls. AI may evaluate only the approved input package; applicable criteria must be identifiable and traceable; missing, conflicting, ambiguous, or unsupported criteria cannot be silently resolved by AI; materially incomplete information must trigger review, escalation, or additional information; and the human reviewer must be able to determine exactly what information and criteria the AI actually used.
This means the demonstration has now answered much of the question:
What exactly are we demonstrating?
The next phase begins with POAM 1-6 and shifts toward a different question:
What must be true for AI to participate safely, accountably, and reconstructably inside that workflow?
That work will interrogate each node and transition for data provenance, transformation, decision authority, acceptance authority, timing conditions, uncertainty response, escalation, recovery, configuration management, and reconstructability.
The larger question behind the demonstration therefore remains:
“The question is not simply whether AI can analyze the data. The question is whether the workflow preserves judgment, authority, provenance, and evidence when the decision matters.”
Where We Go Next
The next phase of ControlPointAI is less about defining the problem and more about testing whether the controls we have been describing can survive contact with execution.
For the Piping UT Demonstration, that means moving from the technical baseline into requirements, authority boundaries, evidence requirements, and representative scenarios. The objective is not simply to create a workflow that looks well governed on paper. It is to determine whether the required controls actually operate when conditions change, information is incomplete, pressure increases, or a consequential decision must be made.
That distinction matters.
A policy can require human review. A procedure can specify an approval step. A system can display a control point. But none of those things, by themselves, prove that meaningful human judgment occurred or that the approved authority boundary held during execution.
That is where our attention is now moving.
Over the coming months, the demonstration will increasingly focus on the gap between as-designed governance and as-executed behavior: what the system was supposed to do, what it actually did, what evidence remains, and whether an independent reviewer can reconstruct the decision path afterward.
The broader goal remains the same as when ControlPointAI began:
AI should strengthen human judgment, not obscure who is responsible for it.
But the work is becoming more concrete.
We are moving from asking whether organizations have governance controls to asking whether they can prove those controls worked when the decision mattered.
And that may be the most important question of all.