A white paper

The AI Value Illusion

Operationalizing AI beyond adoption

Adam Fisher · October 2026

Quick overview

The core argument · 4:53

Full paper

Every section, read aloud · 51:01

Figure 1. The inversion. Conceptual, not data. Every modern delivery framework was built to protect scarce building capacity. AI collapses the cost of first drafts; what is scarce now is judgment, verification, and context.

Work machines produce Human judgment and accountability Lost or unrealized value Captured value

Executive summary

Enterprises are not failing at AI adoption. They are failing at AI operationalization.

AI has made producing work cheap, but the operating model around that work (intake, ceremonies, hand-offs, governance, and measurement) was built for a world where producing work was expensive. So the capacity AI frees gets absorbed instead of redeployed. More licenses and more training will not fix this. What fixes it is a new operating model for the enterprise: a business address for every artifact, versioned context in place of meetings, a governed knowledge graph that people and agents can interrogate, test-first verification, and named humans who are accountable as adjudicators of machine output.

Figure 2. Adoption is up, value is flat. Executives track seats, weekly active users, and prompts. None of these shows whether capacity was captured. The gap comes from the operating model, not the technology. Operating profit is McKinsey’s EBIT measure: earnings before interest and taxes.

AI budgets are heading into the same scrutiny as every other line item. CFOs are being told to measure realized value, not deployment volume. Organizations that can’t show captured value at the next budget cycle won’t be arguing about adoption rates. They’ll be defending why the spend should continue.

The argument in three beats

The shift: expensive to build, cheap to think. AI inverts that.

Agile, Scrum, SAFe, stage gates, and change advisory boards are machinery for deciding what to build, because building the wrong thing was the most expensive mistake available. Generative AI collapses the cost of first drafts of code, tests, requirements, summaries, and reconciliations. What is scarce now is judgment (is this right?), verification (can we prove it?), and context (does whoever is doing the work, human or agent, know what the enterprise knows?).
The hidden failure: adoption is up, value is flat, because the time AI frees never comes back.

An engineer who finishes a two-week story in two days has no structural reason to say so. A product owner who drafts a PRD in an hour still attends the same four refinement meetings. In Gartner’s October 2025 release, 88% of HR leaders report no significant business value from AI tools. McKinsey finds 88% of organizations use AI, but only about 6% credit AI with 5% or more of their operating profit, a share that “has remained flat at about 6 percent.” Gartner’s May 2026 guidance to CFOs puts it plainly: “They must not mistake activity for impact.”
The vision: living versioned context, a business address for everything, humans as adjudicators.

Every PRD, decision, test, runbook, and policy attaches to a position in the business architecture. Context lives in versioned artifacts that people and agents can query, so a 30-person status meeting shrinks to a 10-minute decision forum. AI produces the work, and named humans adjudicate it: they judge it, sign it, and own it. Capacity freed by AI is measured, declared, and planned against instead of evaporating.

The address graph and the governed knowledge layer underneath this vision are set out in full in the companion paper, The Addressable Enterprise (June 2026).

Where the current model breaks

AI does not break the traditional model by failing. It breaks it by succeeding locally while everything around it stays the same. The breakage shows up in six places, each with a measurable mechanism and cost.

Capacity evaporationFreed time is never detected, declared, or redeployed, so it is absorbed by slack. Proxy work and workslopPeople relay AI output without review; the recipient pays to verify it. Lossy context transferContext lives in heads and meetings, decays with every hop, and is invisible to agents. Adoption theaterSeats and usage get reported because captured value was never instrumented. Ceremony built for expensive buildingTwo-hour debates about one-hour builds while the bottleneck moves to review and integration. Governance lagEvidence is assembled after the fact; agentic AI outruns both assumptions.

What replaces it

The replacement is a concrete system that can be built with tools that exist today. It has nine parts. Each layer references the one beneath it, and the top layer feeds measured capacity back into planning. Select any component to jump to its section.

Figure 3. The system on one page. Teal components produce; the amber component judges. The dashed boundary is governance that every agent call passes through. Almost no one has assembled all nine pieces yet.

What it is worth

$140M–$280M

A year in meeting-based context sync for a large enterprise running 200 projects (illustrative model, Part V). Even the most conservative savings case in the model is $28M a year, before counting faster decisions or fewer relay errors.

67,500 hours

Freed each year in a 500-engineer organization at a 30% time reduction on AI-amenable work, about 37 FTE. At a 0% capture rate, the enterprise pays for the licenses and banks none of it.

The value case rests first on meeting and context costs, which hold regardless of the individual-productivity debate. Individual speed-ups must be measured per work type, never assumed.

Where to start

Diagnose, 6–8 weeks
Baseline adoption against value, model meeting cost on real calendar data, score maturity, and pick one bounded pilot domain with a willing leader.
Pilot one domain, 12–16 weeks
Map the domain into the address graph, move its context into versioned artifacts, write executable specs, and replace status meetings with cited digests. Scale only if it clears every gate: at least 30% fewer recurring meeting person-hours, a capture rate of at least 40%, change-failure rate flat or better, and a risk partner who attests the evidence trail is at least as good as before.
Scale, months 5–24
Expand to three to five domains across two lines of business, then make capture rate a KPI the CFO reports.

Almost nobody has built this yet. Firms have deployed copilots, stood up AI councils, and written acceptable-use policies. Very few have rebuilt the plumbing that turns individual speed into enterprise throughput.

Part I

Anatomy of the traditional operating model

1.1The founding assumption: building is expensive

The Agile Manifesto (2001) reacted to waterfall projects whose big up-front specifications went stale before any code shipped. Its economic insight was that building is expensive and feedback is valuable, so teams should build in small increments and learn as they go. The Scrum Guide turned this into a fixed cadence of Sprint Planning, Daily Scrum, Sprint Review, Retrospective, and ongoing Backlog refinement. SAFe added Program Increment (PI) planning, typically a multi-day event every 8–12 weeks that aligns dependencies across dozens of teams.

AI breaks the assumption directly AI breaks it indirectly
Ceremony or artifact	Problem it solves	Hidden assumption	Under AI
Backlog refinement	Avoid building the wrong thing	Engineering hours are the scarcest input	Direct
Story points and estimation	Forecast an expensive resource	Effort per person is stable	Direct
Sprint planning	Commit scarce capacity	Capacity is fixed per sprint	Indirect
Daily stand-up	Synchronize context	Context lives in heads, not artifacts	Indirect
Sprint review	Validate expensive output	Output is scarce	Indirect
PI planning	Resolve cross-team dependencies	Dependencies can only be negotiated live	Indirect
Definition of Done	Quality gate on costly work	Verification follows building	Direct

1.2Intake, planning, building, verification

Demand arrives through annual or quarterly funding, executive asks, regulatory mandates, incident fixes, and Jira or ServiceNow intake forms. Each request is restated several times (business case → epic → feature → story), and every restatement passes through someone else’s understanding. Planning cascades from strategy (annual) to portfolio (quarterly) to PI (8–12 weeks) to sprint (2 weeks). The PMO overlays stage gates with templates and steering committees.

Capacity is set as headcount × historical velocity. That self-reported proxy is quietly invalidated by AI: a team that gets twice as fast but keeps its velocity flat looks exactly the same on every dashboard. Test-driven development is widely endorsed and rarely practiced, because writing tests first feels like double work when typing code is the bottleneck. As a result, verification trails building, and “done” means merged rather than proven.

1.3Coordination

Coordination is mostly synchronous:

Interruptions and ad hoc meetings. Microsoft’s 2025 Work Trend Index telemetry found that employees are interrupted every two minutes during core hours, about 275 times a day. It also found that 57% of meetings are ad hoc and that 48% of employees and 52% of leaders say work feels “chaotic and fragmented.”
Searching for information. Atlassian’s State of Teams 2025 (12,000 knowledge workers and 200 Fortune 1000 executives) found that leaders and teams spend about 25% of the workweek searching for information.
People as the retrieval system. In the same survey, 56% of workers said the only way to get what they need is to ask someone or schedule a meeting.
Executive visibility. Only 7% of executives feel confident they know how each team’s work supports the company’s biggest goals.

The theory behind this has been stable for fifty years. Brooks’s Law (1975): adding people to a late project makes it later, because communication overhead grows faster than capacity. Communication channels: n people have n(n−1)/2 channels. Conway’s Law (1968): systems mirror the communication structure of the organization that builds them, so hand-offs between silos become seams in the product. Information decay and tacit knowledge: every verbal relay loses fidelity, and knowledge held only in heads leaves when people do. Humans are a lossy medium.

Figure 4. Why a 30-person status meeting cannot work. Every line is a channel through which context can be lost. Six people have 15; thirty people have 435.

1.4Governance and measurement

In regulated financial services, governance comes in layers:

Change management. ITIL change management, usually run in ServiceNow, with a Change Advisory Board.
Model risk management. For fifteen years this was governed by SR 11-7 and OCC Bulletin 2011-12. On April 17, 2026, the Federal Reserve, OCC, and FDIC replaced them with principles-based guidance (Fed SR 26-2, OCC Bulletin 2026-13, FDIC FIL-15-2026). The new guidance states: “Generative AI and agentic AI models are novel and rapidly evolving. As such, they are not within the scope of this guidance.”
Third-party risk and internal audit/SOX. Both run on top of the layers above.

Governance today runs on documents and meetings, and evidence is assembled by hand after the fact. Measurement covers delivery (velocity, DORA metrics), portfolio (on time, on budget), and AI (seats, usage, self-reported hours saved). None of it measures captured capacity.

1.5Beyond the SDLC

The same anatomy repeats outside engineering: an operations analyst reconciling dealer-funding exceptions, a risk partner reviewing a new product’s controls, a servicing team updating call-center procedures after a policy change. Each follows the same pattern: intake by form or email, restatement by intermediaries, synchronous coordination, after-the-fact evidence, and effort-based measurement. Most elapsed time builds up in the hand-offs between product, risk, legal, operations, and technology. The work itself often takes hours, while the queues between functions take weeks.

Part II

Where it breaks: a failure-mode taxonomy

AI does not break the traditional model by failing. It breaks it by succeeding locally while everything around it stays the same.

Failure mode 1Capacity evaporation

Mechanism. AI compresses task time for the individual, but nothing in the operating model detects, declares, or redeploys the time it frees. Sprint commitments, velocity baselines, and staffing plans stay where they were. The freed time is absorbed by slack, gold-plating, extra meetings, or simply waiting to report “done.” This is Parkinson’s Law (“work expands so as to fill the time available for its completion”) operating at enterprise scale. Nobody announces they are faster, because announcing speed invites more work or a smaller team. Ethan Mollick calls employees who quietly use AI without telling their organizations “secret cyborgs.”

Capture rate

Figure 5. Where the freed hours go. Illustrative engineering case from Part V: 500 engineers × 1,800 hours × 25% AI-amenable × 30% faster. The split among the four evaporation paths is illustrative.

Evidence.

Gartner (press release, October 28, 2025): “Eighty-eight percent of HR leaders report their organizations have not realized significant business value from AI tools.” Only 7% of organizations give employees guidelines on how to use time saved by AI (July 2025 survey of 114 HR leaders).
Atlassian State of Developer Experience 2025: 68% of developers save 10+ hours a week with AI, yet 50% lose 10+ hours a week to organizational inefficiencies such as finding information and context switching.
Gartner (September 2026; 1,303 respondents): productivity was a key target outcome for 75% of functional leaders and takes roughly 30% of functional AI spend. Yet only 22% of organizations have scaled AI across multiple business units, and roughly 11% do not know what their function spent on AI in 2025.

Cost. In a 500-engineer organization, about 67,500 freed hours a year are at stake. At a 0% capture rate, the enterprise pays for the licenses and banks none of those hours.

Failure mode 2The proxy problem and workslop

Mechanism. People relay AI output without discernment, acting as a proxy between a question and a model rather than as the professional accountable for the answer. A leader using Copilot over Jira, Confluence, and ServiceNow cannot just say “give me what my senior leader wants.” The model does not know the leader’s priorities or which ticket fields can be trusted. Output that looks finished moves downstream, and whoever receives it pays the cost of verifying it.

40%

of workers received workslop in the prior month

15.4%

of the content they receive is estimated to be workslop

1h 56m

to deal with each instance

$186

per employee per month; $9M+ a year at 10,000 people

42%

of recipients saw the sender as less trustworthy

Figure 6. Work that masquerades as work. Each square is one piece of received content; the hollow squares look finished but do not advance the task. BetterUp Labs and Stanford Social Media Lab, 1,150 U.S. desk workers, September 2025, in HBR.

Evidence. BetterUp Labs and Stanford Social Media Lab define workslop as “AI-generated work content that masquerades as good work, but lacks the substance to meaningfully advance a given task.” Microsoft Research and Carnegie Mellon (CHI 2025; 319 knowledge workers) found that “higher confidence in GenAI is associated with less critical thinking, while higher self-confidence is associated with more critical thinking.” The HBS/BCG “jagged frontier” experiment with 758 consultants shows why that matters.

Figure 7. The jagged frontier. HBS/BCG experiment, 758 consultants. This matches the automation-bias pattern long documented in aviation and clinical decision support.

Cost. Rework, eroded trust, and, in regulated contexts, errors that reach customers or regulators under a human signature that did no real review.

Failure mode 3Lossy context transfer

Mechanism. Enterprise context lives in heads and meetings: why a decision was made, what a capability does, which systems a change touches. It is passed along by voice and decays with every hop and every departure. AI makes this worse unless it is fixed, because an agent can only reason over context it can retrieve, and context locked in a meeting is invisible to it.

Figure 8. Humans are a lossy medium. Conceptual. Each restatement passes through someone else’s understanding; by the time intent reaches code, much of the original signal is noise. Context that never leaves a meeting is invisible to an agent.

Evidence.

Atlassian’s State of Teams 2025 (June 2025) found “Fortune 500 companies waste 2.4 billion hours each year just searching for information,” and 43% of workers say they could work faster if information were easier to find.
A McKinsey Global Institute analysis (2012) put searching for and gathering information at 1.8 hours a day.
Steven Rogelberg’s 2022 study for Otter.ai (632 employees across 20 industries) found organizations invest an estimated $80,000 per employee per year in meetings and “could save an estimated $25,000 or more per employee per year” by cutting unnecessary attendance.
Doodle’s 2019 estimate of $541 billion in global cost of pointless meetings comes from a scheduling-software vendor and is directional only.

Cost. In the illustrative model, roughly $140M a year for one recurring meeting series per project across 200 projects, and more once the full ceremony stack is counted (Part V).

Failure mode 4Adoption theater

Mechanism. Organizations report what is easy to count: licenses, active users, use cases launched. What matters (captured capacity, cycle time, P&L effect) needs instrumentation they have not built.

Figure 9. High performers redesign the work and define the human check. McKinsey State of AI 2025, 1,993 respondents. High performers are nearly 3× as likely to have redesigned workflows.

Evidence.

McKinsey State of AI 2025: 88% of organizations use AI in at least one function, but only 39% report any impact on operating profit (EBIT). About 6% are high performers (109 of 1,993 respondents). McKinsey’s August 2026 update says the share “has remained flat at about 6 percent,” while the share reporting any operating-profit impact fell to 37%.
MIT NANDA, The GenAI Divide (July 2025): 95% of organizations saw no measurable P&L return on $30–40B of GenAI spend. The report is preliminary, not peer-reviewed, and contested, so treat it as directional.
Gartner (May 2026; 204 finance leaders) urged CFOs to measure “realized value rather than deployment volume.” A related Gartner keynote reported that 62% of CFOs say fewer than a quarter of their AI initiatives deliver measurable benefits.
Gartner (September 2026) found that high performers who track ROI continuously and reallocate funding report positive returns on 81% of AI initiatives. Low performers do not know the return on 29% of theirs.

Cost. Capital goes to the wrong places, and executives lose credibility when budgets come up for review.

Failure mode 5Ceremony built for expensive building

Mechanism. When a draft implementation takes an hour, a two-hour refinement debate about whether to build it is upside down. Building two variants and testing them is often cheaper. Yet the ceremony stack is unchanged, while the real bottleneck has moved downstream to review, testing, and integration.

Figure 10. Keyboard speed is real but uneven. Change in time to complete a task with AI; left is faster. Hollow markers are perceptions or estimates their authors call unreliable. The spread is why time reduction must be measured per work type, never assumed.

Evidence. DORA’s 2025 State of AI-assisted Software Development (nearly 5,000 respondents; correlational) found 90% of respondents use AI. AI adoption now correlates positively with throughput, reversing the 2024 finding, but it still correlates with higher delivery instability. About 30% report little or no trust in AI-generated code. DORA calls AI an amplifier of whatever system it lands in.

Individual-productivity studies conflict. GitHub’s controlled experiment (Peng et al., 2023) found developers completed a scoped task about 55% faster. METR’s randomized trial (16 experienced open-source developers, 246 tasks, early 2025) found AI made tasks take 19% longer (CI +2% to +39%), even though developers believed they were 20% faster. METR’s February 2026 follow-up (57 developers, 800+ tasks) produced point estimates of −18% time for returning developers and −4% for new recruits; METR calls these results unreliable because of selection effects, but says it is “likely” developers are faster now.

Taken together: keyboard speed is real but uneven, and it does not become system throughput unless verification and integration are redesigned. Product organizations, the “thinking” side of the inversion, have barely changed their cadence.

Cost. Engineering speed piles up as unreviewed merge requests and unstable releases, while the ceremony tax stays the same.

Failure mode 6Governance lag

Mechanism. Governance assumes a human wrote the artifact and that evidence can be gathered after the fact. Agentic AI, which takes actions and calls tools, outruns both assumptions. With the rules unsettled, firms either freeze (no value) or proceed informally (unmanaged risk).

Figure 11. Governance has a hole exactly where agents live. Since April 2026, U.S. model-risk guidance excludes generative and agentic AI, while supervisors are already asking about shutdown, audit trails, and provenance.

Evidence.

U.S. bank regulation. SR 26-2 and OCC 2026-13 exclude generative and agentic AI from scope and promise a request for information “in the near future.” Govern365’s full-text Federal Register review found no such RFI in the 155 days to September 19, 2026. Reuters reported on October 13, 2026 that supervisors are informally asking banks how they govern AI agents: shutdown, audit trails, and model provenance.
EU AI Act. The Digital Omnibus (Regulation (EU) 2026/1744, in force July 27, 2026) moved Annex III high-risk obligations, which cover creditworthiness assessment, from August 2, 2026 to December 2, 2027. General-purpose AI obligations have applied since August 2, 2025.
NIST. The AI RMF 1.0 (January 2023) organizes risk work into Govern, Map, Measure, and Manage. Its Generative AI Profile (NIST AI 600-1, July 2024) lists 12 GenAI risks, including confabulation, human-AI configuration, and value chain and component integration.

Cost. Either value is blocked or risk goes unpriced. Supervisors are already asking for audit trails that meeting-based processes cannot produce.

Part III

The new system

Nine components, each closing one or more failure modes.

Figure 12. Every failure mode has more than one fix. Hover over or focus any item to trace its connections.

3.1The business address graph

Every artifact (PRD, ADR, test suite, runbook, control, ServiceNow change, Jira epic, policy, dataset) carries an address. Each layer references the layer below it, which gives traceability both upstream (from a function to the strategic objective it serves) and downstream (from an objective to every artifact, test, and control that implements it). The companion paper The Addressable Enterprise gives the full reference design for this spine, including ownership and freshness at every address, inheritance by position, and the brownfield build path.

Figure 13. A business address. The graph is directed, not a tree: shared capabilities such as Identity and Notifications have more than one parent. Every artifact hangs off a node, so tracing upstream and downstream is a query, not a meeting.

Design rules.

It is a directed graph, not a tree. Shared capabilities such as identity, payments, and notifications have many parents.
Anchor on capabilities. They are the most stable layer, consistent with BIZBOK practice and TOGAF capability-based planning. Value streams are ordered paths through the graph.
Align team ownership to the graph. Following Team Topologies, stream-aligned teams own domains and platform teams own shared capabilities, which turns Conway’s Law into an asset.
Treat maintenance as governance. Every node has an owner, a review cadence, and a last-attested date. Stale nodes are flagged.

Closes: lossy context transfer, adoption theater, governance lag.

3.2Versioned context artifacts in GitLab

Artifact	Purpose	Format
Constitution	Non-negotiable domain principles (security, accessibility, data)	Markdown
PRD or spec	What and why, acceptance criteria	Markdown with YAML front matter (address, owner, status)
ADR	Decision, options, rationale	Numbered Markdown, immutable once accepted
Executable spec	Defines done	Playwright, Cypress, or Maestro
Runbook	Operational procedure	Markdown linked to a ServiceNow CI
Decision log	Forum outcomes, signed by the adjudicator	Markdown
Status digest	Weekly state from Jira, ServiceNow, and GitLab	Machine-generated, never hand-written

This is docs-as-code applied to the whole operating model. It follows the pattern of GitHub’s open-source Spec Kit (constitution → specify → plan → tasks → implement, with a converge step that checks code against the artifacts) and AWS Kiro’s spec-driven IDE. It also mirrors written cultures such as GitLab’s handbook-first practice and Amazon’s narrative memos. The principle is that the spec, not the chat history, is the source of truth. GitLab merge requests provide review, approval, diffs, and immutable history, which is exactly the audit trail regulators ask about. Confluence remains the reading surface.

Closes: lossy context transfer, governance lag, the proxy problem.

3.3The governed knowledge layer

The knowledge layer is a proposed governed knowledge graph that would index every addressed artifact and the relationships among capabilities, features, services, APIs, datasets, controls, tests, incidents, and owners. It is designed around an addressable map: every node in the business architecture can be resolved, and every artifact hangs off a node. It is a design, not a running system; the requirements below describe what it must do. The Addressable Enterprise describes this layer in depth: ingestion on merge, authority tiers, and the personal, institutional, and product knowledge tiers.

Graph plus retrieval (GraphRAG). Vector search finds relevant passages, and graph traversal finds relevant relationships, such as what depends on this capability or which controls test this function.
Provenance on every node. Source repository, commit, author, attestation date, and classification.
Entitlement-aware. An agent querying on a user’s behalf sees only what that user is allowed to see.
Freshness signals. Stale content is flagged, never silently served.

Figure 14. Graph plus retrieval, with governance built in. The knowledge layer is a design. The figure shows the behavior it must have, not a running system.

Closes: lossy context transfer, the proxy problem.

3.4The interrogation layer: Copilot, agents, MCP

Copilot over Jira, Confluence, and ServiceNow for the broad workforce.
Coding agents (for example, Claude Code) working in GitLab with the constitution and specs in context.
Model Context Protocol (MCP) servers that expose the knowledge layer, GitLab, Jira, and ServiceNow as governed tools, so context is pulled in rather than pasted.

The interrogation contract: any answer that informs a decision must cite addressed artifacts, and “the AI said so” is not provenance.

A leader asking “what should my senior leader know about Payments this week?” gets a digest in which every claim links to a Jira item, ADR, or incident, and the leader then adjudicates it.

Agent quality depends less on prompts than on what is in the context window. Constitutions, addressed specs, and the neighbors retrieved from the graph are the basic building blocks of context engineering.

Closes: lossy context transfer, the proxy problem.

3.5Agent IAM and governance

Agents are a new class of identity and need the same lifecycle controls as any other privileged principal.

Figure 15. A governed agent call. These controls answer the questions supervisors are reportedly asking (shutdown, audit trails, provenance), map to the NIST AI RMF Govern and Manage functions, and fill the gap left by SR 26-2’s carve-out.

Control	Pattern
Agent identity	Registered principal with owner, purpose, and risk tier
Delegated authority	Acts on behalf of a named human with scoped, time-bound tokens
Tool allow-listing	Governed MCP catalog; namespaces separate experimental from production
Least privilege	Read by default; write scopes require approval by tier
Audit trail	Every tool call logged with agent ID, on-behalf-of user, inputs, outputs, and addresses touched
Kill switch	Per-agent and per-tool disable, regularly tested
Model provenance	Registry of approved models and versions per tier

Closes: governance lag.

3.6Model routing

Atlas, a local-first AI platform and model-routing architecture the author has prototyped, contributes three design ideas:

Route by task class, sensitivity, and cost. Classification and extraction go to small or local models, complex reasoning goes to frontier models, and restricted data stays inside its boundary.
Evaluate continuously. An evaluation harness runs representative tasks per route, so model swaps are based on evidence.
Design memory deliberately. Keep session context, project memory (versioned artifacts), and enterprise memory (the knowledge layer) separate. Chat history is never the system of record.

Figure 16. Route by task, sensitivity, and cost. Atlas is a prototype; it was proposed and has not been adopted. On AWS and Bedrock, a scale-to-zero principle keeps experimentation cheap and lets cost be attributed to each address.

Closes: adoption theater, governance lag.

3.7Test-first verification

The definition of done becomes a passing executable specification, written before implementation: Playwright or Cypress for web, Maestro for mobile, and contract tests and data assertions for services and pipelines.

TDD rarely survived contact with practice because writing tests first doubled the expensive typing. AI makes drafting tests cheap, so the human’s job becomes adjudicating whether the tests capture the intent of the PRD. Agents then code until the tests pass.

Figure 17. Verification moves to the front. Amber steps are human gates; teal steps are machine work. “Done” means the executable spec passes, not that the code merged.

This moves verification to the front and is the strongest lever against DORA’s instability finding. The same pattern applies to operations: reconciliation rules become assertions, and policy changes become test cases against procedures.

Closes: ceremony built for expensive building, the proxy problem.

3.8The adjudication role model

AI produces; humans adjudicate.

Adjudication means judging, correcting, and owning output. The adjudicator’s name is on it, and errors are theirs.

Role	Produces, AI-assisted	Adjudicates, human-owned
Engineer	Code, tests, refactors	Spec fidelity, design, security, the merge
Product owner	PRD drafts, acceptance criteria	Intent, priority, trade-offs
Operations analyst	Reconciliations, exception summaries	Exceptions, root cause, customer impact
Risk and compliance	Control mappings, evidence packs	Control sufficiency, risk acceptance
Senior leader	Portfolio digests	Decisions, allocation, escalation

Norms.

No relay without review. Forwarding output means you have adjudicated it.
Understand, don’t memorize. Engineers need not recall syntax, but they must be able to explain what the code does and defend it without the AI present.
Tiered adjudication. Higher-risk addresses need deeper review or a second adjudicator.
Recorded adjudication. Approvals live in merge requests and decision logs.

This model targets the Microsoft/CMU finding directly: it rewards confidence in one’s own judgment rather than confidence in the tool.

Closes: the proxy problem, governance lag.

3.9Capacity accounting

Re-baseline. Adjust team capacity each planning interval from measured cycle-time change per work type, not from self-report.
Declare. Teams publish a capacity ledger showing hours freed and where they went (new scope, tech debt, risk reduction, learning, or released budget).
Attribute. Every work item has an address, so captured capacity rolls up by capability, product line, and LOB.
Reward declaring. Teams that surface capacity get first claim on redeploying it. Headcount cuts must not be the default, or people will stop declaring.
Tie to P&L. Link redeployed hours to earlier revenue, avoided incidents, displaced vendor spend, or released budget.

Figure 18. The capacity ledger loop. Freed hours are measured, declared, attributed by address, and redeployed every planning interval.

The instrumentation combines Flow Framework metrics, DORA, and SPACE or DX Core 4, with capture rate as the headline KPI.

Closes: capacity evaporation, adoption theater, ceremony built for expensive building.

3.10Where the author’s work stands

Part of this system is in use, part is designed or prototyped, and one piece has not been started.

Work	Status	Component it informs
Lightspeed: in-house AI initiative; hundreds of engineers on Claude Code and custom tooling	In use	Interrogation layer (coding agents), test-first verification, adjudication norms
10-session weekly Copilot workflow program for delivery teams	Running	Interrogation layer for non-engineers; adjudication habits; workflow redesign
The knowledge layer and the addressable map	Design Capability map under way; graph layer not built	Business address graph; governed graph layer
Atlas: local-first platform, model routing, evaluation, memory design	Prototype Proposed, not adopted	Model routing, evaluation harness, memory architecture
Enterprise MCP governance and agent IAM	Contributor Contributed; did not own	Agent IAM and governance; governed MCP catalog
Capacity accounting	Not started Proposed in this paper	Capacity ledger and re-baselining

Lightspeed shows engineering adoption at scale, and the Copilot program is extending it to the rest of the workforce. The remaining components are designs: the knowledge layer would supply addressed context, agent governance would make it safe, and Atlas-style routing would make it economical. Capacity accounting is the missing piece that would make it pay. None of the arguments in this paper depend on the designed pieces already working.

Part IV

The system in practice

4.1Ceremonies: before and after

Figure 19. The status meeting is context transfer. Base case from Part V: 30 people, three one-hour meetings a week. The number of decision owners shown on the right is illustrative.

Ceremony	Today	Tomorrow
Daily stand-up	15 minutes of verbal status	Generated digest read async; 10-minute huddle for blockers only
Backlog refinement	1–2 hours a week debating stories	AI-drafted spec and executable acceptance tests reviewed in the MR; 30 minutes for contested items
Estimation	Planning poker	Measured cycle-time distributions per address for AI-amenable work
Sprint planning	Commit from velocity	Commit from the re-baselined capacity ledger
Sprint review	Live demo	Passing executable specs and a recorded walkthrough; live time for direction
Retrospective	Unchanged	Unchanged, plus review of adjudication misses
PI planning	Two days, many teams	Dependencies generated from the address graph; live time for real trade-offs
Project status (25–30 people, 3× a week)	Verbal round-robin	Weekly cited digest; one 30-minute decision forum with decision owners only
CAB	Weekly review of all changes	Low-risk changes auto-approved on evidence; CAB reviews high-risk only

What stays: decisions under real disagreement, relationship-building, onboarding, incident response, and creative exploration. The target is the context-transfer portion of meetings, not human contact.

4.2A day in the life

Before

2.5 hours of stand-up, refinement, and cross-team sync. A story whose intent has passed through four people. An afternoon spent in chat threads finding out which service owns a field. The story is finished Tuesday, but the board says “in progress” until Friday.

After

A five-minute digest. The addressed spec opens in GitLab, and MCP loads the constitution and ADRs into the agent’s context. The engineer adjudicates the AI-drafted Playwright tests against the acceptance criteria, the agent codes to green, and the engineer reviews the diff for design and security, then merges. Asking the knowledge layer “what else consumes validateDebitDate?” replaces the chat-thread hunt. Freed hours are logged in the capacity ledger and pull the next addressed item.

4.3Today and tomorrow, layer by layer

Layer	Today	Tomorrow
Intake	Forms and emails restated by intermediaries	Request attached to an address; AI drafts the spec from history; owner adjudicates
Planning	Velocity and story points	Re-baselined capacity ledger; cycle-time distributions
Building	Human-typed code; ad hoc AI	Agents implement adjudicated specs under the constitution
Verification	Tests after code; separate QA	Executable specs first; CI green means done
Hand-offs	Meetings, emails, re-explaining	Addressed artifacts; subscriptions by address
Measurement	Velocity, seats, self-reported hours	Capture rate, flow, DORA, P&L by address
Governance	Meeting approvals; after-the-fact evidence	Evidence by construction: MR approvals, adjudicator records, agent audit logs

Part V

Economics

All figures in this section are illustrative. Sourced benchmarks are labeled as such. Both models below are live: change the inputs to run them on your own numbers.

5.1Meeting and context-cost model

C_meeting = P × F × D × W × R C_portfolio = N_projects × C_meeting × M_stack Savings = C_portfolio × T × E

P is attendees, F is occurrences per week, D is hours, W is weeks per year, and R is fully loaded hourly cost. M_stack accounts for the rest of the ceremony stack. T is the share of meeting time spent on status and context transfer, and E is the share of that time eliminated by versioned context and digests.

Base case (author’s directional estimate): 30 × 3 × 1.0 × 52 × $150 = $702,000 per project per year. Across 200 projects that is $140M with one series per project, $211M at a stack multiplier of 1.5, and $281M at 2.0. The author’s directional $200–300M figure therefore implies a ceremony-stack multiplier of about 1.5–2.0. Validate it with calendar telemetry.

$702,000

Cost per project per year

$140.4M

Portfolio cost per year

$28.1M

Addressable savings per year (portfolio × T × E)

Figure 20. Meeting-cost calculator. Starts at the base case with one meeting series per project and the most conservative savings cell (T = 40%, E = 50%).

Scenario	P	F	D	W	R	Annual cost per project
Low	20	2	0.75	48	$110	$158,400
Base	30	3	1.0	52	$150	$702,000
High	35	3	1.0	50	$185	$971,250

Addressable savings on the $140M base
	T = 40%	T = 55%	T = 70%
E = 50%	$28M	$39M	$49M
E = 70%	$39M	$54M	$69M

Even the most conservative cell is material, and none of them counts faster decisions, fewer relay errors, or quicker onboarding. Sourced benchmarks for triangulation: Atlassian (about 25% of the week spent searching), Rogelberg/Otter.ai (632 employees; “$25,000 or more per employee per year” in potential savings from cutting unnecessary meetings), and Microsoft (57% of meetings ad hoc).

Worksheet. To apply the model, sample 10 recurring meetings, record P, F, D, W, and R for each, classify each minute as status and context or as decision and discussion to estimate T, and multiply through.

5.2Capacity-capture model

H_freed = N × h × a × s H_captured = H_freed × c Value = H_captured × R_value

N is people, h is productive hours per year, a is the share of work that is AI-amenable, s is the measured time reduction on that work, and c is the capture rate. Illustrative engineering case: 500 × 1,800 × 0.25 × 0.30 = 67,500 hours (about 37 FTE).

67,500

Hours freed per year (37 FTE)

33,750

Hours captured

$5.1M

Value at the hourly cost floor

Figure 21. Capacity-capture calculator. Run it at s = 10%, 30%, and 50%, then plan on the measured value.

Capture rate	Hours captured	At $150 an hour (cost floor)
0% (status quo)	0	$0
25%	16,875	$2.5M
50%	33,750	$5.1M
75%	50,625	$7.6M

The evidence on s ranges from METR’s 2025 slowdown to GitHub’s 55% speed-up, which is why s must be measured per work type and never assumed. For non-engineering Copilot users, a is typically higher and s is less well measured, which is an argument for instrumenting early.

5.3The investment side

The main costs are:

address-graph build and ongoing attestation (budget roughly 2–5% of a domain architect’s time per domain),
the knowledge layer platform (graph, indexing, entitlements, scale-to-zero inference),
the MCP catalog and agent IAM,
test-first enablement and CI capacity,
change management,
licenses and routed inference.

The case is positive when meeting savings, captured capacity value, and risk avoided together exceed the investment. In most large enterprises, meeting savings alone fund the platform.

Part VI

Implementation roadmap

Each phase de-risks the next, so the enterprise never makes a large bet before it has seen a small number. Every phase has an entry condition and an exit gate.

Figure 22. Twenty-four months, four gates. Diamonds mark the exit gates. The pilot does not scale unless it clears every gate.

Phase 0: Diagnostic Entry: executive sponsor; aggregate calendar telemetry; Jira, ServiceNow, and GitLab data; AI usage data.

Adoption-vs-value baseline
Meeting-cost model on real data
Maturity assessment (Appendix B)
Failure-mode heat map by LOB
Pilot domain selection: bounded, measurable, willing leader, moderate regulatory exposure

Exit: quantified baseline; signed pilot charter with target capture rate and meeting reduction; risk partner engaged.

Phase 1: Pilot domain Entry: charter; GitLab structure; knowledge layer access; MCP tools approved in the pilot namespace.

Map the domain into the address graph and assign owners
Migrate PRDs, ADRs, and runbooks into versioned artifacts
Write a domain constitution
Build executable specs for the top journeys
Replace status meetings with digests and decision forums
Record an adjudicator on every merge and decision
Launch the capacity ledger and re-baseline

Exit, all required: ≥30% fewer recurring meeting person-hours; measured cycle-time change per work type; capture rate ≥40%; change-failure rate flat or better; risk partner attests the evidence trail is equal or better.

Phase 2: Scale

Expand to 3–5 domains across at least two LOBs, including one operations domain
Productize the graph tooling, digests, and MCP catalog
Train adjudicators at scale through the workflow program
Integrate capacity accounting with quarterly planning and finance

Exit: LOB-level capacity reporting; evidence-based CAB auto-approval for low-risk changes; ≥90% of graph nodes attested within the cycle.

Phase 3: Two-year horizon

The address graph covers all LOBs, and portfolio planning runs on it
Status meetings are the exception; capture rate is a KPI the CFO reports
Agents are governed under IAM with audit evidence ready for supervisors
Ready for EU AI Act Annex III (December 2, 2027) and any U.S. interagency AI guidance

Organizational change

Norms: “No relay without review.” “If it isn’t addressed, it doesn’t exist.” “Decisions are written.”
Accountability: named adjudicators per address and tier. Honest misses are reviewed without blame, and rubber-stamping has consequences.
Re-baselining: every interval. Freed capacity is redeployed within the domain first, and releasing budget is an explicit portfolio decision.
Incentives: reward declared capacity and captured value, not utilization.
Roles: product owners become intent adjudicators, QA becomes verification architecture, and the PMO becomes a value-flow office that runs the capacity ledger.

Part VII

Risks, objections, and honest responses

“Some meetings matter.”

True. Decisions under disagreement, trust, dissent, and creative work need people in a room. The model targets only context transfer, which is explicitly parameterized as T. Retrospectives, decision forums, and relationship time stay.

“The graph will rot.”

It will if it is treated as a project rather than as governance. Anchor on capabilities, require an address when an artifact is created (enforced by templates and CI checks), assign owners with attestation dates, flag stale nodes automatically, and fund ongoing curation. A partially accurate graph with freshness signals beats tribal knowledge with none.

“AI will make errors.”

It will, confidently, outside its frontier. The answer is structural: executable specs catch functional errors, provenance exposes unsupported claims, tiered adjudication puts deeper review where the risk is, and stability metrics are a pilot exit gate.

“Deskilling.”

The risk is real, and the Microsoft/CMU findings are a warning. Adjudicators must be able to explain output without the tool. Rotate people through deep-review assignments, treat test design and architecture as core craft, and develop juniors through mentored adjudication rather than by removing their work.

“This is a headcount play.”

If freed capacity defaults to layoffs, people stop declaring it and evaporation returns. The model separates measurement from allocation: capacity is declared and first redeployed to backlog, debt, and risk, and any budget release is a transparent portfolio choice. Engage labor representatives early where they exist.

“Regulators won’t accept AI-produced work.”

Regulators require accountable humans, sound controls, and evidence. They do not require humans to type. Versioned artifacts with named adjudicators and agent audit logs are stronger evidence than meeting minutes. Because SR 26-2 excludes generative and agentic AI, firms need a broader governance framework for those systems, and this system is that framework, aligned to the NIST AI RMF.

“We already have Copilot.”

Copilot is one surface of the interrogation layer. Without addresses, versioned context, adjudication, and capacity accounting, its output stays local and evaporates.

“The productivity evidence is mixed.”

It is, and this paper claims no universal speed-up. The value case rests first on meeting and context costs, which hold up regardless of the individual-productivity debate, and it requires s to be measured.

Caveats

Status of the author’s work. The knowledge layer and Atlas are a design and a prototype, not production systems, and capacity accounting has not been started. Section 3.10 gives the status of each piece.
Illustrative numbers. All Part V models are illustrative. The $200–300M meeting figure depends on a ceremony-stack multiplier that should be validated against calendar telemetry.
Source quality. Doodle, Atlassian, BetterUp, and Microsoft have commercial interests in their findings. MIT NANDA’s 95% figure is preliminary and contested, DORA’s findings are correlational, and METR describes its 2026 follow-up as unreliable.
Regulatory flux. The U.S. interagency RFI on generative and agentic AI had not been published as of September 2026. EU AI Act dates reflect law-firm reporting on Regulation (EU) 2026/1744 and should be confirmed on EUR-Lex.
Anonymization. The author’s employer is not named. No internal figures, SOW terms, or unpublished data are included.

Appendices

Appendix A: Glossary

Adjudication: Judging, correcting, approving, and owning AI-produced output; the adjudicator is accountable for errors. Distinct from producing the work.
Address: A node in the business-architecture graph (Company → LOB → Product Line → Domain → Capability → Feature → Function) to which artifacts attach.
Capacity evaporation: AI-freed time absorbed by slack or low-value work because it is never declared or re-planned.
Capture rate: The share of measured AI-freed hours that is declared and redeployed or released as budget.
Constitution: A versioned document of non-negotiable domain principles that every spec and agent references.
Context engineering: Designing what goes into an AI system’s context window to improve output quality.
Executable specification: Acceptance tests written before implementation that define “done.”
Knowledge layer: The proposed governed graph layer that would index addressed artifacts and their relationships, with provenance and entitlements.
MCP (Model Context Protocol): An open protocol for exposing tools and data to AI agents.
Model routing: Assigning tasks to models by task class, data sensitivity, and cost.
Workslop: AI-generated content that looks like good work but lacks the substance to advance the task (BetterUp/Stanford).

Appendix B: Maturity rubric

Shading deepens with maturity. Level 4 is the full operating model in place.

Dimension	1: Experimenting	2: Adopting	3: Operationalizing	4: Operationalized
Value measurement	None	Seats, usage, self-report	Cycle time per work type in pilots	Capture rate reported to the CFO by LOB
Context	Heads and meetings	Wikis	Versioned artifacts in pilots	Enterprise address graph; meetings for decisions only
Verification	Tests after code	Some automation	Executable specs first in pilots	Test-first default, including operations
Accountability	Informal	AI-use policy	Named adjudicators in pilots	Tiered adjudication enterprise-wide
Agent governance	Shadow AI	Approved tools list	MCP catalog, agent IAM, audit logs	Full lifecycle; tested kill switch; supervisory-ready evidence
Model economics	Single model	Multiple vendors	Routing by task and sensitivity	Continuous evaluation; cost per address
Planning	Velocity	Velocity plus anecdote	Re-baselined in pilots	Capacity ledger drives allocation