AI made building cheap. So "that looks good too," "this looks good too" multiply endlessly. The scarce resource is no longer implementation — it's the judgment of what NOT to build, what to stop, what to decide, and what to commit to. Yet human cognition is wired to add and to overlook what could be removed. This is the real shape of a new organizational debt — read its lineage and current state, with citations.
Agrawal, Gans & Goldfarb's Prediction Machines (2018) read AI through economics, not technology. What AI cheapens is "prediction" (≈ building); its complement — judgment — becomes scarce and expensive. As the marginal cost of building approaches zero, an organization's edge moves to the layer that decides what to build, and what not to.
"Can't decide" is not weak will. It's structure. From individual cognition to group aggregation, four demonstrated mechanisms push an organization toward addition whenever it is left unmanaged.
The additive-bias experiments measured individual cognition. Extending it to organizations is the authors' own reasonable inference (red tape / bureaucracy), not direct organizational evidence. A 2024 replication shows dependence on task, culture and age, weakening "universal default" — the bias reproduces on average, but isn't "always." Kahn's original mechanism is also a market aggregation across many consumers, a different category from a single organization. Tying these studies together as "organizational decision debt" is an interpretation, not a claim the originals make themselves.
→ Read it as "left unmanaged, there is structural pressure toward adding," not "people always add."
When you think through decision-making in the AI era, a set of well-tested prior concepts is already there to lean on. Below is the lineage — from the cognitive bias to codified decision rights.
The core remedy for decision debt converges on Bezos's reversibility frame. And the decisive point: this frame isn't a 2018 relic — in 2026 a tier-1 consultancy re-cites it as a current recommendation.
Irreversible (Type 1 / one-way door): decide methodically, slowly, with consultation. Reversible (Type 2 / two-way door): high-judgment individuals or small groups should decide quickly — not by committee or consensus.
"As organizations get larger, there seems to be a tendency to use the heavy-weight Type 1 decision-making process on most decisions, including many Type 2 decisions. The end result of this is slowness, unthoughtful risk aversion, failure to experiment sufficiently, and consequently diminished invention." — Amazon 2015 Letter (verified against the primary PDF)
Additive bias works by never searching for the subtractive option. So the fix is design, not willpower: structurally force "stop / reduce" options onto the agenda, and put "added vs. retired" side by side on the roadmap.
People most overlook subtraction when not cued to consider it, when they have only one shot, and under higher cognitive load (Nature 2021). I.e., the busier the org, the more it adds.
Deloitte's 2026 edition (published 2026-03) cites Amazon's one-way / two-way door model as a current recommendation for matching decision speed to reversibility in the AI era, advising organizations to "classify choices and pre-assign owners, data, guardrails, and speed for each category." So this core frame hasn't aged out — it is being re-requested for the AI era.
Note: Deloitte uses "one-way/two-way door" / "Amazon's 2015 shareholder letter," not "Type 1/Type 2" or "Bezos." The Type 1/2 label belongs to the Bezos letter.
Decision debt also lives in structure. More middle layers breed the sense that "the decision is made elsewhere," eroding ownership; annual-locked budgets can't track changing conditions. Here are 2025-26 cases, with hedges attached.
The 15% is Amazon's self-report, with no external audit. The true cause is contested: analysts point to $2.1–3.6B in manager-pay savings, while Jassy frames it as "not really financially driven, not even really AI-driven" — about culture, agility, ownership. This note treats the stated rationale, not a verdict on the real cause.
→ Case evidence that fewer layers can speed decisions — not causal proof of effect.
56% / 30% is correlational, as McKinsey itself frames it. Reverse causation is plausible (higher performers hold more cash, so they can reallocate more). The data is 1990–2005 — over 20 years old. The adjacent "10% vs 6%" / "10.8% vs 2.5% CAGR" figures often quoted nearby could not be verified here and are not used (only the 56%/30%/15-year version is).
→ Nimble reallocation is "promising but not causally established." Use it to prompt a budget rethink, no more.
Most organizations that can't decide have ambiguous decision rights. The way out is to codify who decides, who inputs, and that once decided everyone commits. Three of the most canonical mechanisms.
What could be verified here is the stated purpose (reducing ambiguity), no further. The phrasing that Bain asserts a direct causal link from decision accountability to organizational performance could not be supported and was rejected. Stated adoption is widespread, but peer-reviewed or quasi-experimental evidence that it improves velocity or outcomes is thin.
→ "Codify the rights" is sensible, but the effect awaits proof. Don't equate adoption with improvement.
The prior concepts are all here. But the questions specific to the AI era still have thin-evidence areas. Here, honestly, is what the research found to be not yet certain.
Lines that land — for meetings, slides, or a single sentence.
Sources checked against primary documents. Correlation is flagged as correlation, and claims that couldn't be substantiated are not used as support. Numbers are kept identical across JA/EN.