The Lightest-Mechanism Discipline

Definition

The Lightest-Mechanism Discipline is the single operating rule that the Mechanism Axis produces in Greg Kihlström‘s AI Capability Framework, from the book Stop Saying “AI”:

Use the lightest mechanism that clears the task’s accuracy bar.

The whole spectrum of mechanisms — rules, predictive, generative, agentic — collapses to that one instruction. Don’t pay generative prices, and don’t swallow generative variance, for a job a cached predictive score or a five-line rule could handle for a fraction of the cost. Reach rightward along the axis only when the task’s openness genuinely requires it, and when you do, expect to pay for it twice: once in Forces (compute) and once in Trust (verification).

The discipline is the framework’s lean-over-comprehensive ethic pushed down to the engineering layer. It also mirrors the book’s central thesis of uneven disappearance: lighter, more deterministic mechanisms suit the low-stakes, ambient uses that fade into the background, while high-stakes uses demand a more capable mechanism or a heavier Trust layer, which is part of why they stay named.

How It Relates to Marketing

Marketing stacks accumulate over-provisioned mechanisms quietly, because the impressive technique is easy to justify and hard to question once it’s in place. The discipline gives a marketing leader a standing rule to run against every AI line item: is this the lightest mechanism that clears the bar, or is it a heavier one bought because the vocabulary sounded advanced?

Applied across a portfolio, it tends to find two kinds of waste. First, closed, well-structured tasks — routing, scoring, templated sends — running on generative or agentic mechanisms that a rule or predictive model would handle deterministically and far more cheaply. Second, the opposite and rarer error: genuinely open-ended tasks under-built on a mechanism too light to clear the bar, where the visible cost is low but the output quietly misses.

The Accuracy Bar

The work is all in that last phrase — the task’s accuracy bar. Every task has a bar, a level of accuracy below which the output isn’t good enough to ship. The bar is not a property of the technology. It’s set by what being wrong actually costs, which is the cost-of-wrong line drawn down the middle of a funnel.

A subject-line variant sits low. The wrong one trims the open rate and nobody files anything. A refund decision sits high. A wrong one is a chargeback and a board question. Same kind of task, different bar, because the cost of error lies elsewhere. So the discipline says: find the leftmost mechanism that clears that bar, and stop there. Move right only when the input is too unstructured or too novel for the lighter mechanism to clear the bar without dropping under it.

Two things make this harder than “always go cheap.” Moving right is sometimes exactly correct — a two-part, unscripted customer message can defeat a rule tree and a predictive classifier while a grounded generative model handles it cleanly. And under-building an open-ended job misses the bar exactly the way over-building a closed one bleeds money. The rule is to fit the mechanism to the task, in both directions.

How to Apply the Discipline

1. Set the bar from cost-of-wrong. Decide what being wrong on this task actually costs, which sets how accurate the output has to be.
2. Start at the left. Ask whether a rule clears the bar. Then predictive. Then generative. Then agentic.
3. Stop at the first mechanism that clears it. Don’t keep moving right for capability you won’t use.
4. Justify every step rightward. Each step must be earned by the task’s openness, and each step is paid for twice — compute and verification.
5. Re-check when the task changes. A bar can rise (the task starts touching a decision) or an input can get more open, either of which can move the right answer.

Comparison to Similar Principles

Principle	Domain	Relationship
Lightest-mechanism discipline	AI implementation	Match the mechanism to the task’s accuracy bar
Occam’s razor	Reasoning	Shared spirit: don’t add capability without need
KISS / minimum viable solution	Engineering	Same instinct, applied to mechanism choice
Total Cost of Ownership	Procurement	Quantifies the cost side the discipline is guarding

The discipline differs from a generic “keep it simple” rule by naming the exact stopping condition: the accuracy bar set by cost-of-wrong. It isn’t “always go light.” It’s “go as light as the bar allows, and no lighter.”

Best Practices

• Anchor the bar to consequences, not to the tool’s ceiling. The question is what accuracy the task demands, not what the technology can reach.
• Watch both failure directions. Over-building wastes money; under-building misses the bar. Neither is safe.
• Make “why not lighter?” a standing question. Require every generative or agentic choice to defend itself against the mechanism to its left.
• Price the reach honestly. When you move right, count both bills — compute and verification — not just the license.

Future Trends

• Cheaper generative mechanisms raise the temptation. As prices fall, defaulting to heavy mechanisms gets easier to rationalize, which makes the discipline more necessary.
• Verification cost becomes the real constraint. Even as compute cheapens, the verification bill for right-side mechanisms stays stubborn, and increasingly that’s what caps how far right a team should go.

FAQs

1. What’s the rule, exactly? Use the lightest mechanism that clears the task’s accuracy bar.

2. What is the accuracy bar? The level of accuracy below which the output isn’t good enough to ship. It’s set by what being wrong costs, not by the technology.

3. Does it mean always use the cheapest option? No. It means use the lightest mechanism that clears the bar. For genuinely open-ended tasks, that can be a heavier mechanism — under-building is a failure too.

4. What does moving right cost? Twice over: more compute, in the Forces tier, and more verification, in the Trust layer.

5. How does it connect to the book’s thesis? It mirrors uneven disappearance. Light mechanisms suit the low-stakes uses that fade; high-stakes uses need heavier mechanisms or more trust, which keeps them named.

Related Terms

1. The Mechanism Axis
2. The Forklift and the Coffee Cup
3. Cost-of-Wrong
4. The Pillar × Mechanism Grid
5. The AI Capability Framework (Master Stack)
6. The Three Forces
7. Total Cost of Ownership (TCO)
8. LLM Tokens

Sources

• Kihlström, Greg. Stop Saying “AI”. https://amzn.to/4wilWcA
• Greg Kihlström — official site. https://www.gregkihlstrom.com