Decision Architecture

Most architecture conversations stall on the wrong artefact. The team debates whether to draw the diagram in C4 or boxes-and-lines, then files it, and within a quarter nobody can say why the boundary sits where it does. As we argue in Architecture Is Not About Technology — It Is About Decision-Making, the diagram is residue; the architecture is the set of decisions and the trade-offs accepted to reach them. So the practical question is not what should we draw but how do we make and record a decision well. Decision Architecture answers that. It is three moves — classify, decide, capture — built from three external models that fit together more neatly than their authors planned.

The model: classify, decide, capture

Read the framework as three stages, each governing one question.

Classify by reversibility (Bezos). Before deciding, ask which kind of door this is. A two-way door is reversible and cheap to unwind; a one-way door is consequential and irreversible. The classification sets the budget of deliberation you are allowed to spend.
Decide via the advice process (Harmel-Law). A named individual owns the decision and seeks advice — not agreement, not sign-off — from everyone meaningfully affected and everyone with relevant expertise. Advisors inform; they do not veto.
Capture as an ADR (Nygard). Record the single decision as a short document — Title, Status, Context, Decision, Consequences — held in source control beside the code it governs. The output is rationale, not a picture.
Supersede, never edit. An accepted decision is immutable; when reality moves, write a new ADR that links back and marks the old one superseded. The chain becomes a navigable history of why the system is as it is.

The sequence matters because each stage calibrates the next. Reversibility sets how much advice is proportionate; the advice process produces the context and consequences; the ADR makes both legible to whoever inherits the system. Skip the first and you over-engineer trivial choices. Skip the third and you re-litigate settled ones.

Stage one — classify by reversibility

In his 2015 shareholder letter, Jeff Bezos drew the line that should govern architectural tempo. Type 1 decisions are one-way doors: consequential, near-irreversible, to be made slowly and methodically. Type 2 decisions are two-way doors: changeable, to be made quickly by high-judgement individuals or small groups. His warning is the one most enterprises ignore — as organisations grow, they apply the heavy Type 1 process to nearly everything, including the abundant Type 2 decisions, and so manufacture slowness and risk-aversion. Bezos's companion rule is to act on about 70% of the information you wish you had; waiting for 90% is too slow. Reversibility is precisely what licenses that bias to action, because a suboptimal two-way-door choice can be reopened and walked back. Classify first, then you know whether you are buying a quick reversible bet or committing to a boundary you will live with for years.

Stage two — decide via the advice process

Most organisations still route significant decisions through a review board, which couples speed to consensus and removes accountability into a committee. Andrew Harmel-Law's architecture advice process, set out in Facilitating Software Architecture (O'Reilly, 2024), breaks that coupling. Anyone in the system may take and own an architectural decision, on one condition: while shaping the options they must seek advice from two groups — everyone meaningfully affected, and everyone with relevant expertise. The advisors give advice, not agreement and not a veto. The decider stays named and accountable. Described as decentralised and fast, the process meets modern decision criteria by involving the right number of people, optimising who is entitled to initiate a decision, prioritising trust over control, and minimising the need to broadcast decisions after the fact — all of which support team autonomy and fast flow. This is the operating mechanism behind Architecture Governance Without Bureaucracy: rigour without a bottleneck.

A review board asks permission to be slow. The advice process asks expertise to be heard — then holds one named person accountable for the call. The first protects the committee; the second protects the decision.

Stage three — capture as an ADR

Michael Nygard coined the Architecture Decision Record in a November 2011 essay: a short document capturing a single decision as Title, Status, Context, Decision and Consequences, deliberately lightweight enough to suit agile work. Fowler and Nygard frame it as a document that captures and explains one decision and its significant ramifications — and insist that once accepted, an ADR is never reopened or edited, only superseded by a new ADR that links back, creating a historical log. The value, they are explicit, is the recorded rationale and trade-offs, not the diagrams. The discipline has hardened into a standard. Thoughtworks places Lightweight ADRs in the Adopt ring and recommends storing them in source control so the record stays in sync with the code. MADR 4.0.0 (September 2024) is what the Nygard template grew into after a decade of operational experience, adding structured option and consideration sections. Academic treatment (CEUR-WS) defines each architectural decision as a concrete, architecturally significant issue with several solution options and rationale arguing whether and how the desired quality attributes are achieved. And the Azure Well-Architected Framework formally adopted ADRs as recommended practice in November 2024 — enterprise governance catching up to a practitioner habit. Cover the standard separately in Architecture Decision Records: The Missing Link Between Architecture and Delivery; here it is simply stage three.

How to use it, and how to measure it

Apply the framework to any decision whose cost of change is high — the test, after Booch, for whether something is architectural at all. Run it in order: classify the door, gather advice proportionate to the door, write the ADR before the choice hardens rather than after. Writing before forces the options, trade-offs and affected parties into the open, which is the point; an ADR written afterwards to satisfy process is dead paperwork.

Measure the practice, not the volume. The honest metric is decision latency — the gap between information being available and action being taken — which decomposes into signal-to-release time, release-to-execution time and end-to-learning time, exposing whether your bottleneck sits in data, governance or execution. Each added approval point lengthens the timeline, so protective governance can quietly become the very bottleneck it was meant to prevent. The cost of getting this wrong is rarely a single headline number; as CIO has observed of delayed AI-architecture calls, it shows up as repeated unresolved reviews, rising effort spent explaining the system rather than improving it, and slowed adoption because teams will not build on a system they do not trust. A healthy practice shows falling decision latency on two-way doors, stable rationale on one-way doors, and a clean superseding chain rather than silent edits.

A decision matrix

Illustrative — a composite, not a specific client: how the three moves play out across common decisions. The point is proportion, not the exact rows.

Library choice inside one service — two-way door · the decider’s judgement · no ADR. Do not convene a board.
Feature-flag rollout — two-way door · advice from the immediate team · no ADR. Decide fast.
New service boundary — one-way door · advice from affected teams and the relevant expert · ADR required.
Data-residency commitment — one-way door · executive and legal advice · ADR required, reviewed before the door closes.
Letting an agent merge a class of change without human review — a one-way door at the policy level · broad advice · ADR plus a provenance trail.

Who decides, who advises, who is informed

Every significant decision has exactly one named decider, a set of advisors (everyone affected, plus the relevant expertise), and a wider set merely kept informed. Confusing the three is the common failure: a decision with many advisors and no named owner collapses into a committee; one with an owner but no advisors hardens into a silo. The advice process works precisely because it keeps these roles distinct.

Decisions in the agentic era

When an agent participates the shape inverts: the agent proposes (advice), a human decides, and the decision must leave a trail. Classification still governs — an agent may own two-way doors under policy, but one-way doors keep a named human at the gate. This is where Decision Architecture meets Provenance Engineering: Reconstructing Who Decided What When Humans, Agents and Models All Contributed: the ADR records why a decision was made; provenance records who made it and on what evidence.

A maturity model: from boards to managed latency

Level 1 — Boards by default. Everything routes through a review forum, regardless of reversibility.
Level 2 — Records exist. ADRs are written, but after the fact, as paperwork.
Level 3 — Classified. Decisions are sorted by reversibility; two-way doors move fast.
Level 4 — Advice, not approval. Named deciders own the call and seek advice; the standing board is gone.
Level 5 — Latency-managed. Decision latency is measured, superseding chains are clean, and rationale is navigable.

Diagnose your decision practice

Do reversible decisions move at a visibly different speed from irreversible ones?
For your last significant decision, who was the named owner — or was it a committee?
Are decisions recorded as rationale before they harden, or as paperwork afterwards?
When a decision changes, do you supersede it with a linked record, or silently edit?
Do you know your median decision latency?

From insight to action

Decision Architecture is the discipline beneath much of our architecture and governance work, and it long predates AI — decision-making never goes out of date, which is part of why we trust it. We bring the classify-decide-capture practice and the experience of running it without a board; the agentic extensions are convictions we are working out in the open, not a proven method. Begin by classifying one live decision honestly — that single move usually reveals how much of your process is proportionate and how much is ceremony.

That is the whole framework: classify by reversibility, decide via advice, capture as ADRs that supersede rather than mutate. It optimises for decision quality, not diagram coverage. To see how this discipline scales without reintroducing the committee it replaces, read Architecture Governance Without Bureaucracy.