The Problem

Scrum, Kanban, XP, Shape Up: all designed for teams where coding was the slowest thing in the room. AI broke that assumption in 2023. The bottleneck moved. The process did not.

The CNC Analogy

This shift is not new. Manufacturing went through it in the 1970s. Before CNC machines, a machinist’s value was in their hands: steady cuts, feel for the metal, years of muscle memory. CNC automated the cutting. It didn’t replace machinists. It changed what “machinist” meant.

The Axioms

These are the non-negotiable principles the methodology is built on. A team that accepts the axioms can adapt the practices. A team that rejects the axioms should use a different methodology.

The Economics

Datum front-loads effort into specification because defects caught upstream cost orders of magnitude less than defects caught downstream. This is not opinion. It is the most replicated finding in software engineering economics.

Boehm’s Defect Cost Curve

Barry Boehm’s research,^[1] validated by IBM Systems Sciences Institute^[2] and NASA JPL^[3] across decades, quantifies the cost escalation. A missing transaction wrapper costs nothing to add to a spec. It costs 15 minutes during coding, 2–4 hours during testing, and days-to-weeks plus thousands of dollars in production.

When Specification Was Skipped

These are not hypotheticals. These are the dollar costs of missing specs.

The AI Amplification Effect

Without specification discipline, AI makes the economics worse, not better.

AI-generated code that matches a spec is verifiable. The reviewer checks behavior against written criteria. AI-generated code without a spec requires full white-box review, which does not scale. The 98% increase in PRs flows into a review pipeline that is 91% slower, producing no delivery speed gain and introducing quality risk.

Capers Jones’s research across thousands of projects:^[4] poor requirements are the single largest source of project failure, responsible for approximately 40% of all defects. AI does not fix requirements. It amplifies them, for better or worse.

The Capacity Trap

A 2026 joint study by Accenture and the Wharton School^[5] analyzed task-level data across 18 industries covering more than 120 million workers. For a modeled $60 billion company, they estimated $6 billion in potential annual revenue growth from agentic AI at full maturity, alongside $1.7 billion in annual productivity gains.

The catch: roughly two-thirds of those productivity gains materialized as direct cost savings, but the remaining third appeared as cost avoidance: capacity freed for different, higher-value work. Without intentional redeployment, that freed capacity does not become growth. It evaporates.

When to Use It

Before investing in the mechanics, check the fit. Datum is designed for a specific context. If your team matches these four signals, the methodology will produce results within one cycle. If it doesn’t match, save yourself the friction.

Context Variants

The full model assumes a 7-person pod with weekly cadence. These three contexts require adaptation, but the core holds.

Genuine Poor Fit

Three conditions where Datum is the wrong choice. Not “hard to adopt” — genuinely wrong.

What Datum Is Not

For the skeptic in the room.

How It Works

Datum runs on a weekly cadence. The cadence is a rhythm, not a commitment. Stories carry over when they must. The carryover is explicit and communicated.

Delivery Styles

The previous chapter describes the weekly cadence — Monday spec, Tuesday build, Thursday harden, Friday retro. That is one delivery style. Datum’s core principles (spec before build, review against spec, standards as AI governance) are delivery-agnostic. They work in at least three distinct rhythms, each suited to different team shapes and business contexts.

The Team

Datum is designed for a 7-person pod. Five personas, each with a distinct function. The composition is deliberate: enough senior capacity to write specs and review, enough junior capacity to leverage AI for high-volume building, and a single point of architectural accountability.

RACI Matrix

Responsibility assignment across pod roles for every major activity.

Team Sizing

The 7-person pod is the full model. Not every team starts there. The axioms hold at any size. The practices compress.

Growing Engineers

Datum requires a structured growth path because the methodology's leverage depends on juniors who can eventually write specs, not just build from them. The risk without an explicit path: permanent junior executors who produce high-volume, low-accountability output indefinitely.

The four stages are performance-gated, not time-gated. Transitions happen when the engineer demonstrates the gate criteria, not when a calendar threshold passes.

Skills Are the Currency, Not Titles

A 2026 Accenture/Wharton study developed the WAsX (Wharton–Accenture Skills Index) to measure how skills translate into economic value in an AI-enabled economy. The finding: as AI automates routine cognitive work, the market increasingly rewards judgment, coordination, and domain-specific execution. Exactly what Datum’s growth path develops.

The study also found a persistent signaling gap: workers overwhelmingly signal broad, generalist traits, while employers pay for specialized, execution-oriented capabilities. In Datum terms: calling yourself a “senior engineer” signals nothing. Demonstrating that your specs produce low-rework implementations and your reviews catch drift. That is what the market pays for.

This is why the growth path is performance-gated, not time-gated. Gate criteria like “senior revisions are cosmetic, not structural” (Co-Specifiers) and “low rework rate on own specs” (Specifiers) measure skills that carry economic value, not years spent.

The Spec Revision Is the Coaching Session

Accenture/Wharton studied the organizations they call “Talent Reinventors.” Leaders were 1.3× more likely to delegate and coach, even when it slowed execution. These organizations grew revenue 1.8 percentage points faster, strengthened culture (7× more likely), and increased adaptability (4× more likely).

In Datum, this coaching is the spec revision loop.

Talent Reinventors data: Accenture & Wharton School, The Age of Co-Intelligence, March 2026.

The Artifacts

Documentation is infrastructure. Not a nice-to-have, not a chore for after the sprint. Infrastructure, the same way a CI pipeline is infrastructure. Every AI session loads these documents. Every review verifies against them. Every retro improves them. When the documentation is wrong, every AI session produces wrong output at scale. When it is precise, every session inherits the team’s accumulated decisions.

The methodology produces seven named artifacts, each with a clear owner and update cadence. Each artifact is a control surface: one person’s decisions become another’s constraints. Click any artifact to see its structure, an example, and the failure mode it prevents.

Where Artifacts Live

Artifacts are version-controlled, not scattered across wikis and Slack threads. They live in git, get reviewed like code, and have blame history. A concrete convention:

Metrics

Datum produces specific, measurable signals. Not vanity metrics — diagnostic indicators that reveal whether the methodology is working or decaying. Track these five. Ignore everything else.

What Good Looks Like

Expected metric ranges across the first 4-week adoption cycle.

Metric Traps

Metrics that look healthy in isolation can mask dysfunction when read together.

Agent Governance

Every AI coding session is a new employee who has never seen your codebase, has no memory of yesterday’s decisions, and will confidently generate plausible-looking code that violates your architecture. The standards document is the onboarding packet. Agent governance is the HR policy.

Adapted from Fuller, “Create an Onboarding Plan for AI Agents,” Harvard Business Review, March 2026.

The Fresh-Context Problem

What Goes Wrong Without It

The Volume–Risk Spectrum

Not all agent work carries the same risk. A 2026 Accenture/Wharton study argues that governance intensity should scale with where tasks fall on a volume–risk spectrum: high-volume, high-risk domains need more rigorous controls than low-volume, low-risk ones. Uniform governance wastes architect attention on config changes while under-governing payment integrations.

Datum applies this principle at story level. The spec session is where risk is assessed, and the governance response is calibrated accordingly.

Risk classification happens in the spec session, not after the build. The Lead Architect or spec author tags each story as low, medium, or high risk based on three questions: How many systems does this touch? What happens if it’s wrong? Can it be rolled back? The answers determine which governance track the story follows.

The Accountability Asymmetry

“Intelligence may be scalable, but accountability is not.” That line from the 2026 Accenture/Wharton study of AI agents across 18 industries is the core of why governance is structural.

The study found that 50%+ of U.S. working hours are subject to reshaping by AI agents. By 2027, half of business decisions will be augmented or automated, with 15% fully autonomous by 2028. Customer operations are trending toward 80% autonomous resolution by 2029. Agents are spreading “rapidly across the enterprise value chain, often ahead of formal strategy and governance.”

The report’s sharpest line: “Capability parity does not imply responsibility parity.” AI generates PhD-level reasoning but carries no moral weight, no institutional accountability, no long-term obligation. Those stay human. When intelligence distributes across humans and AI, responsibility does not.

The study’s case modeling revealed a pattern that Datum teams will recognize: the function with the largest revenue opportunity (Sales) was also the area with the highest volume of risk-sensitive decisions. Value and risk scale together. Governance must be designed before agents touch high-value systems. You do not get to learn from failure at this scale.

The Accenture/Wharton report proposes “humans in the lead, not in the loop.” Datum operationalizes the same principle: the Lead Architect does not review every line — but they set the standards document that governs every line. The spec author does not write the code — but they define the contracts the code must satisfy. Human authority is exercised through artifacts, not through direct supervision of every keystroke.

Anything else breaks. When intelligence scales and accountability does not, governance must be structural: documents, checklists, and review gates that run regardless of volume. Any model that depends on a human personally reviewing every output will collapse at the speed AI enables.

Source: Accenture & Wharton School, The Age of Co-Intelligence, March 2026. Fortune coverage.

When It Breaks

Every methodology has characteristic failure modes. These are the five ways Datum collapses when teams adopt the rituals without the substance. Each violates a specific axiom. Trace the tag to find where it breaks.

A Week in Datum

The best way to understand Datum is to watch one story move through it. This is PAY-41: Add Apple Pay to Checkout. Five days, from brief to production.

Getting Started

The smallest thing a team can do to experience Datum: run one spec-first cycle on one story. Write a brief (problem, acceptance criteria, constraints). Write a lightweight tech spec (contracts touched, boundary schemas). Build from the spec with AI. Review against the spec. One story, start to finish. If the team cannot see the value after one story — if the spec did not prevent at least one mistake the AI would have otherwise made — the model needs adjustment before scaling.

What to Expect Emotionally

Do Not

Migration

Getting Started covers your first cycle. This chapter covers what happens after — when you need to transition a team, convince a manager, or migrate from an existing methodology.

Every team arrives from somewhere. The three most common starting points each have different friction profiles. What maps cleanly, what changes, and where the resistance concentrates.

Common Objections

Glossary

Accountability Asymmetry

Intelligence scales; accountability does not. AI can generate output at infinite volume, but humans still own every outcome. Governance must be structural (documents, checklists, gates) because direct supervision cannot keep pace. See Ch. 13.

2.5x Multiplier

AI code estimate × 2.5 = total effort including hardening, testing, and verification. The 60% that isn’t code generation takes the same time regardless of how the code was produced.

Builder

Stage 1 of Datum’s growth path. Builds from specs with AI, self-verifies, flags gaps instead of guessing. See Ch. 10.

Brief

Business intent document written by EM/PO. Contains: problem statement, scope, constraints, testable acceptance criteria. The upstream input to the spec session.

Capacity Trap

AI acceleration frees hours, but freed hours do not become growth automatically. Without deliberate redeployment into specification, verification, and governance, the capacity evaporates. See Ch. 3.

Carryover

Stories not completed in the current week, explicitly moved to next. Distinguished from hidden scope creep by being visible at Monday alignment.

Datum

The fixed reference point. In CNC machining, every measurement is relative to the datum. In this methodology, the spec is the datum — the reference for building and reviewing.

Defensive Checklist

8-category quality scorecard (0–3 per category, 24 max). Categories: input validation, error handling, idempotency, timeouts, logging, state management, security, testing.

Co-Specifier

Stage 2 of Datum’s growth path. Drafts specs under senior review. The gap between draft and revision shrinks over time. Gate: senior revisions are cosmetic, not structural. See Ch. 10.

Delivery Style

The cadence pattern: Sprint (weekly batch), Kanban (pick-and-go), or Continuous (per-story flow). The quality model is invariant across styles.

EM/PO

Engineering Manager / Product Owner. Writes briefs with business intent, scope, constraints, and acceptance criteria. In small pods, may be part-time or combined with other responsibilities.

Fresh-Context Problem

Every AI session starts from zero. Without a standards document loaded as context, each session invents its own patterns. The standards document solves this.

Governor

Stage 4 of Datum’s growth path. Multi-lens review, standards contributions, mentors builders. The shift from individual contributor to force multiplier. See Ch. 10.

Hardening

Thursday quality verification. Code is checked against the defensive checklist, acceptance criteria are verified, integration testing runs. Required before Friday demo and release.

Lead Architect

The single point of architectural accountability in a pod. Owns the standards document, leads spec sessions for high-risk stories, reviews complex work. Two functions: governance (setting standards) and coaching (growing juniors via the spec revision loop). See Ch. 7.

Multi-lens Review

Review approach checking multiple dimensions: correctness, security, performance, maintainability, ops readiness. Not just “does it work.”

Pod

A self-contained team of 2–7 people running Datum. Scales by replication (more pods), not growth (bigger pods).

Shadow AI

AI agents deployed without formal governance. 75% of knowledge workers use AI tools, often unsanctioned. In Datum terms: builders generating code without a standards document loaded. See Ch. 3.

Prompt Library

Shared, curated collection of effective AI agent prompts. Captures what works for the pod. Reviewed and updated at Friday retro.

Re-spec Session

30-minute session triggered by the spec-break rule. Patches the spec before continuing the build from a broken reference.

Review Findings Log

Running record of all review findings, spec gaps, and spec-break triggers. Source data for Friday retro. Turns the methodology into a learning system.

Spec Gap

An ambiguity, missing detail, or unspecified behavior in a technical specification discovered during implementation. Three or more spec gaps in a single story trigger the spec-break rule.

Spec Session

Collaborative design meeting where a brief becomes a technical specification with contracts, state machines, boundary schemas, and acceptance criteria. Held Mondays (sprint) or on-demand (kanban/continuous). See Ch. 5.

Spec-break Rule

If a builder discovers 3+ spec gaps in a single story, building stops. A re-spec session is held. The cost of re-speccing is always lower than building from a broken spec.

Specifier

Stage 3 of Datum’s growth path. Writes specs independently, leads spec sessions, reviews juniors’ code. Gate: low rework rate on own specs. See Ch. 10.

Standards Document

Single source of architectural truth loaded into every AI session. The primary control mechanism for AI output quality.

Spec Revision Loop

Datum’s coaching mechanism. A Co-Specifier drafts a spec, the Lead Architect revises it. The delta between draft and final spec is the teaching artifact. See Ch. 10.

Technical Spec

Contract-level specification: API contracts, state machines, boundary schemas, test expectations. The artifact the AI builds against and the reviewer verifies against.

Volume-Risk Spectrum

Governance intensity scales with risk. Low-risk stories (config changes) need minimal oversight. High-risk stories (payments, auth) need full tech spec, ADR, multi-lens review, and spec-break enforcement. See Ch. 13.

WIP Limit

Maximum concurrent PRs per pod. Typically 2 for a 7-person pod, 1 for a 2-person pod. When hit: review before pulling new work.

References

• Crowley, J., Close, K., Munie, K. and Karaca-Griffin, S. The Age of Co-Intelligence: How Humans, AI Agents and Robots Are Redefining Value. Accenture Global Products Practice & Wharton AI and Analytics Initiative, March 2026.
• Boehm, B. Software Engineering Economics. Prentice Hall, 1981. See also Boehm, B. and Basili, V. “Software Defect Reduction Top 10 List.” IEEE Computer 34(1), January 2001.
• Jones, C. Applied Software Measurement. McGraw-Hill, 3rd edition, 2008. Approximately 40% of all defects traced to requirements errors across 12,000+ projects.
• NASA JPL. Software defect cost studies, 1990s–2000s. 70–85% of rework costs traced to requirements errors. Validated Boehm’s defect cost escalation curve.
• IBM Systems Sciences Institute. Relative cost of fixing defects by phase. Cited in Boehm (1981) and widely replicated.
• Faros AI. Developer telemetry study: 10,000+ developers, 1,255 teams. 2025. +98% PRs merged, +21% tasks completed in high-AI-adoption teams.
• Stern, L. (Agoda). Developer productivity data: +91% increase in PR review time under AI-assisted workflows. Summarized in Stiller, E. InfoQ, March 2026.
• Stiller, E. “AI Coding Assistants Haven’t Sped up Delivery Because Coding Was Never the Bottleneck.” InfoQ, March 2026.
• Griffin, L. and Carroll, R. “Spec-Driven Development.” InfoQ, 2025.
• Fuller, J. “Create an Onboarding Plan for AI Agents.” Harvard Business Review, March 2026.
• Brooks, F. The Mythical Man-Month. 1975 (Anniversary Edition 1995). Chapter 4: Conceptual Integrity.
• Brooks, F. “No Silver Bullet: Essence and Accident in Software Engineering.” 1986.
• Singer, R. Shape Up: Stop Running in Circles and Ship Work that Matters. Basecamp, 2019.
• Anderson, D. Kanban: Successful Evolutionary Change for Your Technology Business. Blue Hole Press, 2010.
• Cagan, M. Inspired: How to Create Tech Products Customers Love. Wiley, 2018.