Part II of the series on Implementing AI with Judgement.
In Part I, we explored why most AI strategies fail; because they're applied too broadly and too early, before organizations understand where AI can meaningfully improve operations.
Three structural dynamics make this worse: AI evolves faster than organizations can adapt, its true costs are obscured by subsidized economics, and applied indiscriminately it compresses the variation that makes companies distinctive.
But AI can and should make your business more capable. The established organizations getting real value from it are not the ones adopting fastest. They are the ones who identified specific operational bottlenecks, evaluated whether AI belongs there, and deployed it with discipline. The risk is not in using AI. It is in using it without evaluating what you are applying it to. Applied to the wrong bottleneck, AI does not just fail to help. It can cause harm like eroding differentiation, introducing errors into high-stakes workflows, and creating reputational exposure that is difficult to reverse.
This raises the obvious follow-up: once you've identified a bottleneck, how do you evaluate whether AI is the right intervention?
To answer this question, I built the Earned Automation Framework. It evaluates each bottleneck along three axes:
Strategic Differentiation — is the work core to how the business wins?
Variance Sensitivity — can the situation tolerate errors?
Cost of Reversal — if this goes wrong, what does it cost to undo?
Each produces a binary answer. Together they classify the bottleneck into one of eight distinct situations, each with a clear recommendation.
This framework rests on a few foundational beliefs. First, the evaluation should focus on the situation, not the technology. AI capabilities change quarterly; your bottleneck's characteristics don't. Second, organizational pressure to adopt AI is real, but pressure is not a business case. Third, automation is earned, not rationalized. If the situation isn't ready for AI, the right move is to invest in the process before investing in the tool. And fourth, narrowing where you apply AI is a feature, not a limitation. Less surface area, more leverage.
The Three Axes
The framework evaluates each bottleneck along three dimensions, in order. Each produces a binary classification: high or low. The sequence matters because each question builds on the last, narrowing the field and shaping the recommendation.
1
STRATEGIC DIFFERENTIATION
Does this bottleneck encode how the business wins?
This is the existential question, and it comes first for a reason. If the work at this bottleneck is a source of competitive advantage, replacing it with a generic AI tool risks regressing your output toward the industry mean. AI, by definition, draws from broad patterns. When applied to work that should be distinctively yours, it compresses variation rather than creating it.
Report formatting, document routing, data classification, internal knowledge retrieval
Principle
If the work is a source of competitive advantage, the cost of automating it poorly is not just operational inefficiency. It is the quiet erosion of what makes your organization distinctive.
2
VARIANCE SENSITIVITY
What is the cost of being wrong here?
AI is fundamentally probabilistic. It will produce errors and inconsistencies. That is not a flaw to be engineered away; it is a characteristic of the technology. The relevant question is not whether errors will occur, but whether the situation can tolerate them. This axis is not a measure of AI's accuracy. It is a measure of how forgiving the environment is when any output, from any source, is occasionally wrong.
HIGHErrors cascade
Compliance decisions, medical or legal judgment, financial reporting, safety-critical systems
If occasional errors are unacceptable in the situation itself, AI without careful and ongoing oversight is a risky substitution regardless of how good it currently is at the task.
3
COST OF REVERSAL
If this goes wrong, what does it cost to undo?
The first two axes evaluate whether AI should do the work. This one evaluates what happens when it shouldn't have. Every decision carries a cost to reverse. Sometimes that cost is operational: unwinding a pricing error, retraining a model, rebuilding a workflow. Sometimes it's reputational: how customers, employees, or the market interpret the fact that you used AI here at all. This axis is unique because it requires evaluation through two lenses.
Operational Lens
How hard is it to undo?
A hiring filter that screened out candidates for six months cannot be rolled back. A regulatory submission may have consequences that outlast the system. A pricing algorithm quietly setting wrong prices may be invisible externally but devastating to unwind.
Perception Lens
What does this signal about your company?
AI-generated cold outreach is trivially easy to stop, but every recipient has already judged how you value their time. AI replacing human expertise tells customers efficiency matters more than they do. The damage is silent: you rarely hear from the people who quietly wrote you off.
Internal documentation, workflow automation, internal analysis, draft pipelines with human review
Principle
If either lens is high cost, the axis is high cost. You don't average them. A decision that's easy to reverse but sends a terrible signal is still high risk. A decision that's invisible externally but operationally irreversible is also high risk. One is enough.
These classifications are intentionally binary and intentionally judgment-based. What constitutes high differentiation or high variance sensitivity varies across industries, companies, and even teams within the same organization. The binary choice forces a clear commitment: is this high, or is it low? If there's genuine disagreement about where a bottleneck lands, that disagreement is itself a signal worth examining before proceeding.
The questions don't need to be complex. For Strategic Differentiation: is this work something distinctive to your business, or something a competitor could buy or replicate with basic effort? For Variance Sensitivity: when errors happen here, can they be absorbed with a quick correction, or do they cascade into downstream problems? For Cost of Reversal: does unwinding this decision create meaningful operational setbacks or permanently change how affected audiences perceive you?
The goal is to evaluate these honestly, not optimistically. The organizations that get real value from AI are the ones willing to be truthful about what they're applying it to before they apply it.
Reading the Landscape
Each axis produces a binary classification. The combination defines one of eight distinct situations, each with its own risk profile and recommended approach.
This is not a scoring system. You are not averaging across dimensions or weighting one more heavily than another. The combination is a classification, not a calculation.
Some of these situations will feel obvious. A bottleneck with low differentiation, low variance sensitivity, and low cost of reversal is a straightforward automation candidate, and most leaders already treat it that way. The value of this framework is not in confirming the easy calls. It's in forcing clarity on the ones that feel ambiguous, the ones where organizational pressure to "do something with AI" runs ahead of a sober evaluation of whether you should.
The eight situations range from full automation to full stop. Most of them land somewhere in between, which is where the real decisions live.
Download the Earned Automation Framework
A two-sided field manual covering the three axes and all eight situations. Designed to be used in the room where the decision is being made. Get the PDF →
The Terrain
1. THE GREEN LIGHTAUTOMATE AGGRESSIVELY
Differentiation: LOWVariance Sensitivity: LOWCost of Reversal: LOW
Classic AI-friendly work. Boring, bounded, and forgiving. Nobody outside the building sees it, errors are cheap, and none of it is why anyone buys from you. This is where AI should feel obvious and boring.
Recommendation
Automate aggressively. No special guardrails needed.
Examples
Report formatting, expense categorization, scheduling, data cleanup, document routing
2. THE QUIET RISKPILOT QUIETLY
Differentiation: LOWVariance Sensitivity: LOWCost of Reversal: HIGH
Mechanically safe but culturally visible. The work itself isn't core and errors are tolerable, but the decision to use AI here sends a signal or is hard to walk back. People notice who (or what) is doing it, and that perception is hard to undo once formed.
Recommendation
Pilot quietly. Measure perception as carefully as efficiency.
Differentiation: LOWVariance Sensitivity: HIGHCost of Reversal: LOW
AI can help, but the cost of a bad output is real. The saving grace is that mistakes are easy to catch and cheap to fix. The output needs to be good because downstream decisions depend on it, but a bad draft gets caught and rewritten before it matters.
Recommendation
Augment, don't replace. Humans stay in the loop on every decision that matters.
Differentiation: LOWVariance Sensitivity: HIGHCost of Reversal: HIGH
Errors here are expensive, visible, and become screenshots. The work isn't what differentiates the business, but getting it wrong carries real consequences, both operationally and reputationally. The danger is that the low differentiation makes it feel safe to automate, while the high variance and high reversal cost make it anything but.
Recommendation
Extreme caution. If AI is involved at all, it stays backstage and a human signs off on everything.
Differentiation: HIGHVariance Sensitivity: LOWCost of Reversal: LOW
This is where the business wins, but errors are tolerable and easy to reverse. The danger is subtler: AI doesn't fail dramatically here, it slowly regresses your output toward the industry mean. Each individual output looks fine. But over months, AI pulls everything toward generic. Your messaging starts sounding like everyone else's and you can't pinpoint when it happened.
Recommendation
Constrain heavily. Use AI only with tight guardrails and active monitoring for quality drift.
Differentiation: HIGHVariance Sensitivity: LOWCost of Reversal: HIGH
Core to differentiation, hard to undo, and high signal risk. But variance is low, which means leaders often convince themselves the risk is manageable. It isn't. Occasional errors won't break anything. But the decision to automate signals what you value, and once stakeholders know a human was removed from the process, that perception is nearly impossible to reverse.
Recommendation
Do not automate. This is not an AI problem. Fix the underlying process instead.
Differentiation: HIGHVariance Sensitivity: HIGHCost of Reversal: LOW
The most dangerous archetype because it looks like the most attractive one. The work is strategically important, errors are costly, but the low cost of reversal makes leaders feel like they can always pull back if it doesn't work. That confidence is misplaced. By the time you realize AI has been making subtly wrong calls on strategically important, high-variance work, the damage is already embedded in decisions downstream.
Recommendation
Slow down. Test in parallel with human judgment. Do not scale until the parallel results converge.
Differentiation: HIGHVariance Sensitivity: HIGHCost of Reversal: HIGH
Everything is high. The work is core to how the business wins. The situation cannot tolerate errors. Reversal is expensive or impossible, and the signal damage is severe. This is where organizational pressure to "be innovative with AI" runs hardest into the reality that some things should not be automated.
Recommendation
Don't. If pressure exists, name the risk explicitly and make the cost of proceeding visible to decision-makers.
The archetypes describe the situation as it exists today, not as it has to be. If a bottleneck lands in a restrictive archetype but the business case for AI is strong, the question isn't "how do we justify automating anyway.", it's "what would we need to change about the process, the oversight, or the recovery path to move this into a different archetype?"
Adding a human-in-the-loop review step can move ‘Variance Sensitivity’ from high to low. Building a fast, transparent correction process can move ‘Cost of Reversal’ from high to low. Neither of those changes the AI. They change the cost structure of the situation, which is what the framework fundamentally evaluates.
Consider algorithmic loan decisions. Today, it sits in the Red Zone: the work is core to a financial institution's identity, errors directly harm people, and wrongful denials create legal and reputational exposure that's nearly impossible to reverse. But what if every AI-generated denial triggered an automatic human review within 24 hours? What if applicants had a clear, fast appeals path? What if the system was designed to flag edge cases rather than make final calls? You haven't changed the AI. You've changed the situation around it. You've moved ‘Cost of Reversal’ from high to low, and potentially ‘Variance Sensitivity’ as well, depending on how the oversight is structured. The bottleneck might now classify as The Blind Spot or even The Wingman.
This is the most actionable part of the framework. The eight situations are not permanent labels, they're a diagnosis of current conditions. And like any diagnosis, they point toward interventions. The question is whether leaders are willing to invest in the process changes that make automation responsible, rather than simply arguing that the technology is good enough to skip that step.
This is the difference between rationalizing automation and earning it.
In Part III, we'll get into the harder problem: once you've identified the right bottleneck, how do you implement without breaking the organization around it. The framework tells you where, but how you bring people through it determines whether.
A downloadable field manual summarizing this framework is available here. It includes the three axes on one side and all eight situations on the other, designed as a reference for evaluation conversations.
Peter Kruger
