August 12, 2026

Barbed Wire Defines the Edge, Not the Direction.

Assiduity AI

AI reliability is often treated as an engineering problem: constrain the machine. But the most valuable forms of intelligent work create a different problem entirely.

The Economist ran a good analogy this week. The American frontier did not become economically useful simply because explorers pushed west. It became useful because settlers arrived, and settlers needed infrastructure. Alan Greenspan had a particular fascination with one such technology: barbed wire. AI now has plenty of frontiersmen, but what it still lacks are settlers.

Businesses remain reluctant to hand consequential work to autonomous systems because those systems can be unpredictable. In response, an industry has emerged to supply AI’s equivalent of barbed wire: guardrails, evaluations, monitoring, identity systems, kill switches, permission layers, and increasingly sophisticated mechanisms for making agents behave. The analogy is apt, but it is incomplete in one important way: barbed wire works when you already know where the boundary is. Some of the most valuable enterprise work begins where you don’t.

The strongest version of the fence

The Economist highlights Scaled Cognition, one of the more interesting companies taking the reliability problem seriously. Its approach is fundamentally architectural: make important behaviors predictable by designing reliability into the system rather than relying on a model to comply with instructions after the fact. For many tasks, that makes enormous sense.

Consider the kinds of actions that dominate transactional workflows: block the card, transfer only the authorized amount, apply the correct fare rule, protect sensitive information, and refuse to execute an action until required conditions have been satisfied. Where the critical behaviors and boundaries can largely be specified in advance, reliability is fundamentally a problem of specification and engineering. Reduce unwanted variability, define the permissible operating space, and make violations structurally difficult. That is an important problem, and companies solving it are doing valuable work.

But there is another class of work where the problem looks very different.

When the path cannot be specified

Consider evaluating an acquisition, investigating the cause of a product failure, developing a restructuring plan, identifying an overlooked regulatory exposure, reconciling conflicting evidence into a defensible recommendation, or designing a strategy nobody has tried before. These tasks do not have a single predetermined sequence of correct actions. The objective can be specified, as can constraints, evidence standards, authorities, and escalation requirements. What cannot be fully specified is the solution path itself.

That distinction matters because, if we already knew every correct step, we would need a procedure rather than intelligence. The first reliability problem asks how we make specified behavior reliably executable. The second asks how we govern intelligent behavior when the useful path cannot be specified beforehand. The first is primarily about engineering predictable execution. The second is about governance under discretion.

This is where the familiar language of AI reliability begins to break down. An intelligent system is valuable precisely because it can select among possible paths, combine information in novel ways, and discover solutions that were not written into a rulebook beforehand. If reliability requires eliminating that freedom entirely, then reliability can eventually consume the very capability we were trying to deploy.

The failure is not always at the boundary

Complex systems can produce globally undesirable outcomes even when every local action appears reasonable. A sentence can be reasonable, the next sentence can also be reasonable, every tool call can be authorized, and every individual step can comply with policy. Yet twenty steps later, the system may be solving a subtly different problem from the one it was originally given. Nothing obvious was violated; the work is simply no longer the work you asked for.

That kind of failure is especially dangerous because it can remain invisible while it compounds. A boundary control can tell you that something prohibited occurred. An execution rule can tell you whether a particular action was allowed. Observability can tell you what the system did. But open-ended intelligent work raises another question: is the work, as it develops, still faithful to the purpose for which authority was delegated?

That is a trajectory problem. It becomes more consequential, not less, as models become more capable because a capable system can move away from the original purpose while remaining persuasive, coherent, and useful-looking. A weaker system may fail noisily. A stronger system can fail elegantly.

This is also why the record has to be independent of any particular point of intervention. Some systems expose moments where a path can still be redirected; others increasingly perform more of their reasoning inside a single step, leaving less for an outside system to act on directly. A control approach that only works when it can steer the outcome is only as durable as that access. A measure of fidelity to the mandate has a broader role: it can be maintained across the execution that is observable to the organization even when the available steering surface changes, which makes the record itself valuable as underlying systems evolve.

Discretion is not the defect

Organizations already know how to govern capable actors whose decisions cannot be completely specified in advance. A board does not give a chief executive a decision tree covering every circumstance the company might encounter. It establishes objectives, authorities, constraints, risk limits, evidence expectations, reporting requirements, and escalation conditions, and then it grants discretion. That discretion is not an unfortunate side effect of executive work; it is part of what the organization is paying for.

The same principle increasingly applies to AI. If we want machines only to execute behavior that we have completely specified beforehand, deterministic systems are an excellent answer. But much of the economic promise of AI lies somewhere else: in finding useful answers we did not already know how to specify. That means unpredictability cannot simply be engineered out without qualification. Sometimes variation is where the value is.

The challenge, therefore, is not to eliminate discretion. It is to make discretion governable.

The missing control point is generation time

This is where the current reliability stack still has a gap. Guardrails are good at enforcing boundaries. Observability is good at recording what happened. Evaluations are good at testing systems before or after deployment. Delegated intelligent work, however, creates another requirement: maintaining fidelity to the mandate while the work is still developing and intervention can still matter.

That is the generation-time control problem. Once a consequential deviation is embedded in a completed analysis, logged perfectly, and sent to a reviewer, the organization may have excellent evidence of a failure that has already occurred. For consequential work, that can be too late. The control point has to move closer to execution itself.

That is the problem we are working on at Assiduity. The organization defines the mandate governing the work: what is being attempted, what must remain true, what evidence matters, where authority ends, and when human judgment must return.

The core of the system is a continuous, quantitative record: a measure of how far the developing work has moved from the mandate, tracked over the life of the task and available for audit at any point along the way. That record exists whether or not anyone intervenes. Where the model exposes a choice among possible continuations, the system can also act on that measure directly, steering the work back toward the mandate before a deviation is committed to the output.

The machine retains freedom over how it solves the problem, while the control system focuses on whether the developing work remains faithful to that mandate. The objective is not to predetermine the answer. It is to keep intelligent execution attached to institutional intent while the path is still forming.

The model proposes. Assiduity disposes.

Control should be proportional to risk

Governance also does not mean placing the same level of control around every task. A routine internal analysis and a billion-dollar acquisition should not operate under identical tolerances, just as drafting a marketing note and preparing a regulatory filing should not require identical oversight. Organizations already understand this principle because authority, supervision, and escalation vary according to consequence.

AI should be governed the same way. The relevant question is not simply whether a system is controlled. It is how much discretion the system should have for this task, under this mandate, at this level of risk. That makes governance a deliberate organizational decision rather than an accidental property of the model.

Low-consequence work can permit greater machine discretion. Higher-consequence work can justify tighter control and earlier human involvement. The goal is not maximum control but appropriate control. Control, like discretion itself, should be proportional to risk.

In practice this means setting tolerances for a given task before the work begins, and treating a breach of that tolerance as a trigger for review rather than as a surprise discovered afterward — the same logic organizations already apply to financial and operational risk limits.

Control need not mean capability loss

There is an obvious objection to this argument. If we govern AI more tightly, do we simply make it less useful? Reliability and capability are often treated as opposing forces, with stronger control purchased at the cost of creativity, flexibility, or performance.

Our research suggests that trade-off does not have to be inevitable. In one 200-document evaluation, controlled generation improved mandate adherence on 199 of 200 documents, with the pattern subsequently replicating across other models and task domains. More importantly for enterprise use, our more recent work with frontier models and realistic professional tasks found no visible degradation in the quality of the resulting work product.

That is the result that matters to us. The point is not that a control layer magically makes a frontier model smarter. It is that governability need not require making the model less capable. Enterprises do not want reliability achieved by reducing capability. They want the benefit of increasingly capable models while retaining authority over the work those models perform.

Two control frontiers

Enterprise AI therefore appears to have at least two different reliability problems. The first is specified execution: how do we make known behavior reliably executable? Deterministic architectures, permissions, schemas, validation, and hard constraints are extraordinarily powerful here. The second is governed discretion: how do we preserve intelligent freedom while keeping the resulting work faithful to institutional intent? That requires a different frame built around mandate, trajectory, bounded discretion, proportional control, and human escalation.

Enterprise AI will need both because they solve different problems. Confusing them risks pushing AI toward one of two bad extremes. At one extreme, we allow systems so much discretion that organizations cannot reliably say what happened between instruction and outcome. At the other, we make systems reliable by specifying their behavior so completely that we remove much of the intelligence we wanted to deploy.

The interesting territory lies between those extremes: machines capable enough to discover, and organizations capable of remaining in control.

Barbed wire helped settle the American frontier because settlers knew which land needed fencing. But intelligence is valuable precisely because it can take us somewhere we did not already know to map. When behavior can be specified, engineer it. When intelligent discretion is necessary, govern it. And when AI is doing the work, that governance has to reach the work while it is happening.

The next control frontier may be less about building ever higher fences than about learning how to measure and govern what happens inside them.