Framing Freedom: A Governance Model for Agentic AI
Analysts predict that a large share of agentic AI projects will be canceled within the next couple of years. Not because the technology failed, but because organizations deployed agents without orientation: no clear outcomes, no cost controls, no guardrails. At the same time, the loudest cautionary tales of full autonomy have produced security vulnerabilities and a frenzy where companies forced adoption before they understood it. These look like opposite problems, one too much control and one too little. They share the same root failure: neither side understood that constraint is what gives agency its shape.
Four quadrants of AI agency
Every organization deploying AI lands somewhere on two axes: constraint and orientation. Low constraint with low orientation is the let-it-rip failure, agents doing whatever they are prompted to with no guardrails, producing vulnerabilities, wasted spend, and the cancellation wave. High constraint with low orientation is the lockdown failure, where everything is prohibited by default, which drives innovation underground. Low constraint with high orientation is rare and unstable, strong vision with no structure to execute it.
High constraint with high orientation is the goal. Call it formative freedom: constraints that shape rather than restrict, and direction that gives autonomy its meaning. Here constraints are not no, they are how. How does this tool serve our business objective? How does this workflow change align with where we are going? This is where companies actually ship AI to production. The quadrant you are in is not about how many policies you have. It is whether your constraints express something real.
Technology in search of a problem
We see plenty of companies running both failure modes at once: locked down on an approved tool that is clunky, so people quietly run everything by a different tool on the side. We also work with a company whose agentic AI project has been in pilot for more than a year. The problem is not the technology. They do not really understand what they are doing. This happens constantly: technology in search of a problem, rather than a problem in search of technology. You need orientation toward a realistic business goal that aligns with what the company is actually trying to accomplish. A year or two ago everyone had to do something with AI, and it was rarely about value. It was about not being left out.
Why constraint is formative, not restrictive
Modern culture has fetishized freedom from constraint, but constraint is what shapes things. Solving a problem without constraints is not really solving it, because there is no frame. What is the budget, the timeline, the regulatory considerations? There is no perfect solution, only trade-offs, and choosing the right ones requires the correct framing. Freedom can be constrained by orientation rather than by rules alone. A clearly defined business outcome dictates the necessary permissions, the required human checkpoints, and the boundaries of agent autonomy. Rules can express orientation, but they cannot be the end in themselves. You cannot orient toward north if you do not know where you are standing.
A computer must never make a management decision
A 1979 IBM training manual put it plainly: a computer can never be held accountable, therefore a computer must never make a management decision. Nothing about that has changed. AI should help humans make decisions, not make them for us. What has changed is that the causal chain has gotten longer and more obscure. When an agent hallucinates a price and triggers a contract breach, someone bears the legal and moral weight. The AI does not. The human who oriented the agent or clicked run is where accountability lands. Deploy an autonomous agent and you have not escaped the burden of decision. You have only made it harder to see who is deciding.
Human in the loop as a feature, not a bottleneck
A leader recently wanted a tool that would pull from timekeeping software and automatically email people whose entries were missing or poor. The answer was no. Nothing frustrates people more than an automated email telling them their time is wrong when it is fine. Instead, have the AI draft the emails, have a person verify the flags before anything sends, and train the system on which drafts were valid. The AI assists. The human decides. Human in the loop is not a bottleneck and not a security blanket. It is what makes agentic systems cohere.
Two companies, one difference
Consider two companies in the same technology landscape with radically different outcomes. The first built a proof-of-concept chatbot over an existing data interface and kept going. Their governance is lean and responsive, and the lead developer and the decision-maker talk daily. Time to production: months. They are shipping, iterating, and capturing value. The second has been in pilot for over a year. Nobody is clearly in charge, and they are not engaging the people they have with the relevant expertise. This is not a personality clash. It is what poor orientation does to capital. The technology was never the problem. The absence of orientation was.
Freedom needs a frame
Effective leadership means understanding the tools your people want to use. You cannot just lay AI over a workflow. You have to understand the workflow first, then find where AI fits. AI is genuinely better than people at some things, such as summarizing large amounts of information quickly. People are still far better at deciding from incomplete information. The target is high constraint with high orientation: constraints that express your actual business objectives, and orientation everyone from the C-suite to the developer can point to. Constraint is formative. Orientation gives freedom its shape. That is what freedom actually is.