Trust

New technology usually outruns the operating discipline required to trust it. AI is no exception.

AI does not have a capability problem. It has three others.

A trust problem.

A performance problem.

A ROI problem.

The question we hear the most in the media is which model is best. But that question is not the important one. The real question is much harder.

I have been working on that question while building The Compass. Then I saw the pattern, and it felt familiar.

After grad school, my first job was as a programmer and systems analyst on Wall Street in the early 1980s. At that time we often wrote what we called spaghetti code. We optimized for speed. Get the application built. Get it running.

That worked until something broke. The faster we coded, the harder the software was to understand, maintain, and debug.

Structured programming became the thing.  It brought discipline to the process and things got under control.  Design and build took longer but debugging and fixing got much easier. Applications always have bugs.

Then computing started escaping the mainframe data center.

The problem was that building and maintaining mainframe applications was slow and expensive. Business units got frustrated with that, so they bought their own departmental machines, installed databases, hired programmers, and built their own applications. The Sun / Sybase configuration was all the rage.

We gained speed. Applications were built and deployed quickly. But the distributed computing architecture exposed a new set of problems. The enterprise was losing control over the quality of its systems.

It was no longer enough for an application to work. It had to be reliable, available, interoperable, scalable, and secure.

IBM helped establish part of that production mindset, particularly around reliability and availability. Standards bodies and enterprise architecture disciplines expanded it over time.

The precise history matters less to me than the pattern.

 AI is somewhere inside that transition now.

Executives do not need another AI hype cycle. They know AI is powerful. What they need is confidence that it can operate inside the business.

Three findings say what many executives already feel.

Trust has not followed usefulness. KPMG found that 61 percent of US workers reported positive effects from AI, while only 41 percent said they were willing to trust it.

Value is showing up as efficiency, not revenue. Deloitte surveyed 3,235 senior leaders across 24 countries. Sixty-six percent reported productivity or efficiency gains. Twenty percent reported increased revenue. Seventy-four percent expect revenue growth eventually. 

If you have followed my value stack essay series, you know that efficiency is the weakest of all the value types.

Control is behind deployment. In the same research, only one organization in five had a mature governance model for autonomous AI agents.

The conversation is moving from capability to performance, trust, and ROI.

Formal standards already exist.

NIST published an AI Risk Management Framework, and a generative AI profile alongside it. ISO and IEC published 42001, which specifies requirements for an AI management system covering governance, traceability, transparency, and continuous improvement.

Both operate at the organizational level. They govern how a company manages AI as a program. NIST is voluntary guidance. ISO 42001 certifies the management system around AI, not that any particular output is correct.

My question is narrower. Can I trust this specific system inside the operating flow of the business, on an ordinary Tuesday, when the person using it is busy and the stakes are real?

That gap is what the following list is for.

As I built The Compass, I kept returning to the disciplines I learned earlier in my career. Some are timeless. Others matter more now, because AI introduces new kinds of uncertainty.

Three of these deserve more than a line.

Auditable. Governments and regulators are going to become more involved. Companies will need to show that appropriate controls existed and were followed. AI audits are coming. They may not look like financial audits, but the underlying question will be familiar.

Transparent. Why doesn’t AI have a dashboard?

When I drive a car I can see how fast I am going, how much fuel I have, whether the engine has a problem, and how far I have left to travel. AI gives me remarkably little visibility into the system I am operating.

What does it know? What is it assuming? Is the context degrading? How much capacity remains? Which model is running? What is this costing? Is there evidence the answer may be unreliable?

Capital Efficient. The most powerful model is not automatically the right model. Sometimes the better business decision is a faster, cheaper model that does the required job well enough.

Cost is not named as a first-order requirement in any of the major frameworks. Neither is adaptability, which appears only indirectly through lifecycle management and continuous improvement. In the case of cost, that is less an oversight than a consequence of authorship. Technologists and regulators wrote those frameworks, and neither group carries a P&L. The CFO carries the P&L, and the CFO is the one being asked to approve the spend.

 Architecture

Working through these requirements, I came across a term that clarified the architecture: the agentic harness.

It describes the management and control layer around an AI model. The terminology is still settling, and it overlaps with scaffold, runtime, orchestration, and control plane. The concept is more useful than the label.

It helped me see the architecture more clearly. Four layers, a human present across all of them, and outcomes feeding back into the system.

The Compass AI Architecture. Four layers, one governance boundary, the human across the whole loop.

The AI model provides intelligence. Compass can use multiple models rather than depend on one. Different models are better at different jobs, and they carry different costs, limitations, and rates of change. That makes the model increasingly interchangeable. The intelligence matters. The brand name on it matters less.

The agentic harness provides control. It manages the environment around the intelligence: context, memory, tools, permissions, routing, execution, and monitoring. The model may be capable of doing something. The harness determines whether, when, and how it is allowed to.

The evidence layer provides ground truth. Strategy built on bad information is still bad strategy. The system has to distinguish between what is known, what is assumed, and what is unknown. Sources matter. Dates matter. Assumptions matter. Unknowns matter most.

The Compass provides strategy. The AI model is not the Sales Strategy. Compass supplies the business logic: Time, Velocity, Position, the eight strategies, the Value Stack, Persona, bridging, and the pattern library. The model helps apply intelligence. Compass determines what it is being applied to.

The human provides judgment. AI can observe, analyze, recommend, and increasingly act. The human determines what matters, applies experience and context, weighs risk, and owns the decision.

Results provide learning. Then reality gets a vote. Did the recommendation work? Did the assumptions hold? Was the ground truth actually true? Did I pick the right model? The outcome comes back into the system and improves the next cycle.

Y2K is the precedent that stays with me.

Through the 1980s, to conserve expensive computing resources, we routinely stored years as two digits instead of four. Dates were everywhere. 1985 became 85.

It seemed perfectly reasonable. We were building systems expected to last three to five years, not the fifteen or twenty many of them stayed in production.

Then someone asked what all those systems would do when the year became 00.

Y2K became a multi-billion-dollar remediation effort because a tiny design decision, embedded across an enormous number of systems, had to be found and corrected one instance at a time.

I wrote two-digit year fields myself. I was part of the problem.

The lesson stuck.

 Now imagine the small AI agent you are building this month is still running in fifteen years.

We have an opportunity to get ahead of that cycle rather than clean up after it.

These ten are not the Enterprise AI Standard. They are the requirements I am using as I build The Compass. Three places I am not confident.

Serviceability. The original mainframe discipline treated it as its own requirement. I folded it into Reliable. That may be a compression that loses something.

Fairness and bias. NIST names it as a pillar. I treat it as belonging to the model provider and the legal function. That may be the right boundary or a convenient one.

Thresholds. A batch job either produces the right number or it does not. A model produces a distribution. I do not know how to set a meaningful threshold for reliability in a system that is not deterministic, and I am not sure the field has settled it either.

I am not claiming authority here.  I’m just trying to figure this out like you.  I’m doing this because I’ve been through 5-6 technology transitions and the same movie plays over and over.  Its pattern recognition.

Your organization may need ten different requirements. Maybe eight. Maybe fifteen. Take what is useful here and leave the rest.

Develop your own. But establish the discipline while AI is still being built into the business, rather than after it becomes too expensive to unwind.

Music. This essay was written while listening to “Sunrise” by Ryan Bingham.

Photograph. New capability arrives first.  Discipline follows.

About these requirements. They reflect my own experience building enterprise systems, current research, existing industry frameworks, and what I am learning while building The Compass. They are not an official or authoritative industry standard.

Sources. KPMG, Trust in Artificial Intelligence, April 2025. Deloitte, State of AI in the Enterprise, survey conducted August and September 2025, 3,235 senior leaders across 24 countries.

My use of AI in writing this essay. I use AI to research, test, structure, and refine this work. The judgment, experience, and point of view behind it were developed over decades in the market.

What I do. I help B2B and enterprise sales organizations create the power required to close business and make their number.

The Compass. The Compass is the Sales Strategy Operating System designed to help Sellers make their number navigating uncertainty by managing Time, Velocity, and Position.

Strategy. Sales Strategy is the discipline of creating the power to achieve your revenue objectives in competitive, rapidly shifting and AI-powered markets.