AI investment isn’t returning value because most organisations run it through an operating model built for projects, which is the argument Matthew Hodgson makes in Evolve: The Operating Model AI Demands (Hodgson, 2026). This page summarises the book across its four parts, from the AI value gap in Chapter 1 to the four executive commitments in Chapter 15, for two readers: the executive deciding whether the book deserves a weekend, and the reader who has finished it and needs to brief a CEO, CFO or board in twenty minutes. Each section names the chapter that carries the full argument and evidence.
In Brief
- About 5% of organisations capture AI’s value at scale, and they have redesigned their processes around AI capability (Chapter 1).
- AI investment stalls for three structural reasons: Plan-Build-Run absorbs the gain, transformation programs change the vocabulary and leave governance where it was, and AI debt builds where no dashboard looks (Chapters 2, 3, 11 and 12).
- Four executive commitments remove those reasons: One point of accountability; Persistent teams, outcome funding; Consumption-speed finance; and From order-taker to order-maker. Each fails without the others (Chapter 15).
- The changes come in sequence: investment governance and performance measurement first, team design and decision rights second, AI augmentation at scale third. The first move is one product, funded for an outcome (Chapters 13 and 15).
Operating model design decides AI returns
Chapter 1
Four national audit offices across three continents have examined how public bodies use AI and reached the same finding: agencies are demonstrating their adoption of AI and are not demonstrating its returns (Hodgson, 2026, Chapter 1). When Australia’s Auditor-General examined AI at IP Australia, it found the governance largely effective, with a policy, completed risk assessments and defined accountabilities. The same audit found the benefits inconsistently defined and measured, and strategic oversight of what the AI produced incomplete (Australian National Audit Office [ANAO], 2026).
Boston Consulting Group’s survey of 1,250 senior executives across nine industries sorts companies into four maturity stages. Grouped here into three tiers, about 60% are laggards, about 35% are scaling and about 5% are what BCG calls future-built, and BCG reports that the value gap between them is widening (Boston Consulting Group [BCG], 2025).
| Tier 1 · Personal tool | Tier 2 · Team accelerator | Tier 3 · Enterprise capability | |
|---|---|---|---|
| Distribution | ~60% | ~35% | ~5% |
| AI role | Human-initiated, single-step interaction. Full human control at every step. | AI accelerates existing workflows. Underlying process design unchanged. Visible but limited. | Agents execute multi-step workflows autonomously within designed governance boundaries; processes redesigned around AI capability. |
| Governance model | None designed. Full human control at every step replaces it. | Controls added to the existing operating model: policy, access rules, and review gates. The governance is designed; the model it sits within is not. | Governance and operating model designed together. Decision boundaries revised each cycle based on what the cycle reveals. |
| Enterprise return | Individual efficiency gains (faster drafting, summarisation, research) with no enterprise-level capability advantage. | The ceiling equals the existing process’s ceiling. | Returns grow with each operating cycle. Governance capability grows. |
Source: Hodgson (2026), Table 1, adapted from BCG (2025).
Tier 3 organisations generate 1.7 times the revenue growth of laggards, 3.6 times the three-year total shareholder return and 2.7 times the return on invested capital, and 62% of their AI initiatives have been deployed against 12% at laggards (BCG, 2025). The difference is the rate at which pilots become operating capability, and that rate reflects how the organisation is structured.
An organisation running a technical operating model, with functional separation, project funding and milestone measurement, sits at Tier 1 by structural constraint. Designing AI governance inside that model lifts it to Tier 2 and no further, because learning stays with the person or team that acquired it and is lost when the project closes. Moving from Tier 1 to Tier 3 takes the same structural changes as moving from the technical operating model to the product operating model, applied to teams that now include AI agents.
Reason 1: Plan-Build-Run absorbs the gain
Chapters 2 and 3
Most large technology functions still run Plan-Build-Run, a design built in the 1980s and 1990s for periodic, predictable work. Plan turns demand into requirements, Build delivers them and Run sustains the result. Each handoff runs one way. What Build learns about the real problem is what Plan needed, and it arrives after the specification has been costed and signed. Run inherits the system without the context that shaped it, and without the mandate or budget to send a change back upstream.
In project-based organisations, work is actively progressed for less than 20% of its cycle time (Kersten, 2018). Evolve calls this overhead the Silo Tax: capacity absorbed by coordination at functional boundaries, learning lost at those boundaries, and management attention redirected from outcomes to utilisation. Activity reports cannot show it. An organisation paying a 20% Silo Tax and one paying 2% produce identical activity reports, and the difference appears only in what customers receive.
AI makes the Silo Tax worse. A developer who writes code three times faster still waits for the same handoff, the same approval chain and the same queue, so active work time shrinks while waiting time stays where it was. Mao and Mirhoseini (2026) measured the same overhead in AI multi-agent systems, where no organisational politics apply. In their centralised design, where a main agent assigned subtasks and rebroadcast findings, constraints were softened as they passed through the controller and agents spent their budgets rediscovering the same failures. With one of the models tested, coordinating through shared, verified state cut cost per task by roughly 50% and raised single-attempt accuracy by 9.3 percentage points over the strongest baseline; with a stronger model the gains were smaller.
Australia’s Robodebt scheme used automated decision-making to raise welfare debts people did not owe, and by late 2016 it was issuing more than 20,000 debt collection notices a week. It ran until 2019, affected 453,000 people and cost the Australian Government A$565 million net (Holmes, 2023). The Department of Social Services, which owned the policy, reported paying almost no attention to how the measure was implemented (Hannah & Botterill, 2025), and the Department of Human Services, which operated it, had no authority to question the policy. Complaints reached operations from the start and had no route back to anyone who could act. It took four years and a Royal Commission to stop a scheme running at human speed. The same design running at AI execution speed would produce harm faster than any human compliance layer could intercept.
Reason 2: Transformations leave governance unchanged
Chapter 12
Most transformations reach Plan-Build-Run’s management coordination layer and stop there. Investment governance, approval processes and executive accountability sit outside the program’s scope. Consensus forms around the most visible symbols of the change, such as its language, task boards and dashboards, while the investment logic beneath them stays intact. Bromley and Powell (2012) call this decoupling.
Three functions sit outside almost every transformation, and each can pull it back within twelve months. Procurement reasserts project logic with every fixed-scope Statement of Work. Annual capital and appropriation cycles fund a scope, a cost and a handover certificate. Executive committees keep their agendas, decision rights and reporting cadence, so their governance behaviour does not change whatever the operating model document says.
Every scaled framework, whether SAFe, LeSS, Nexus or Scrum@Scale, coordinates teams and leaves open who answers for outcomes. Four failure modes appear whichever framework is in use, and each has a signal an executive can test.
| Failure mode | What happens | The signal |
|---|---|---|
| Vocabulary without structure | Coordination improves inside an unchanged governance structure | The last investment decision above a team’s Sprint was made by a committee that predates the transformation |
| Appointment without accountability | A Product Manager without control of the Product Backlog, releases or budget | The Product Manager seeks upward approval for decisions they are accountable for making |
| Advisory capture | The advisory firm supplies capacity inside the executive’s existing mental model | The advice matches what the executive asked for rather than what the diagnosis shows |
| Portfolio reversion | The portfolio stays a project list with red-amber-green status | Investment still flows through annual allocation, reported by the people accountable for the status |
Three questions test whether a transformation has reached the governance layer. Has the funding model moved from project allocation to persistent product funding tied to outcomes? Has measurement moved from deliverables to value realised? Have decision rights moved from committees above the teams to the boundary where the consequences sit? Three “no” answers mean coordination changed and governance did not, and one or two “yes” answers are not enough, because the unchanged conditions will pull the changed ones back within a budget cycle.
Reason 3: AI debt goes undetected
Chapter 11
An AI team’s governance lives in its instruction set: skill files, prompts and quality specifications, all answerable to the Definition of Done. When the business moves and the instructions do not, the agents keep doing what they were told with full confidence. Evolve calls the result AI debt: the organisational liability that builds when governance over AI agents degrades undetected, measured as the gap between what the portfolio evidence shows and the conditions producing that evidence. It builds through four linked mechanisms.
| Mechanism | What it looks like |
|---|---|
| Ordering decay | The Product Backlog stays unchanged across several Sprints while new evidence arrives, so each item delivered is work that will be revisited or written off |
| Instruction drift | Governing instructions carry no revision history, and agents follow stale rules to produce compliant output with no customer value. A model version change, including one the provider makes, alters how every instruction is read while the text stays the same |
| Quality standard erosion | Each agent’s instructions drift away from the others and from the Definition of Done, and the irreversibility boundary that standard protects is the next thing to fail |
| Trust inertia | Scrutiny relaxes on a track record that no longer holds. Agents emit none of the ambient signals a human team would, so nothing flags the other three |
The first instinct is to buy an observability platform. It traces agent actions, logs tool calls and produces the audit trail a regulator expects, but it records only what an agent did. Whether the agent was working on the right problem, and whether its instructions still matched the business, sit outside the dashboard. Adding human review at every step fails for a different reason. Human inspection capacity is fixed while AI production capacity scales, so review at every step becomes the constraint on the whole system.
At terminal scale, AI debt becomes governance theatre: every artefact exists and none governs. Four questions at the quarterly portfolio review test for it. Do governing instructions carry a revision history, or have they gone unrevised across several Sprints? Is the Definition of Done still the standard every skill file and agent instruction answers to? Is the escalation path carrying operating model signals to the Product COO, the role accountable for how a value stream’s operating model runs, and are they acted on? Has a budget cycle eroded the team persistence that makes pattern detection possible? An adverse finding on any of them shows that the conditions for detecting failure have degraded.
Four commitments remove the three reasons
Chapter 15
Where AI returns keep growing, four design decisions sit above any technology choice. Each is a commitment the executive makes before teams can act.
One point of accountability. A single person holds ongoing authority over what a product team works on and what it releases. When three managers each hold partial authority over what an AI team releases, contested outputs generate meetings rather than decisions, and machine-speed execution queues behind a human-speed approval nobody owns. The Product Owner holds the release decision for a team and the Product Manager across a value stream.
Persistent teams, outcome funding. Teams stay with the product and are funded against an outcome. A team that has run forty Sprints has calibrated its governing instructions against forty cycles of inspection and adaptation. Disband it and the documents survive while the calibrated judgement disperses, and the replacement team starts again.
Consumption-speed finance. Spend is reported against evidence of value, and spending authority sits where AI consumption happens: with the team inside the Sprint, with the Product Manager inside the product’s investment envelope, and with the portfolio for reallocation. A single poorly scoped agent can consume a month’s budget in a day (Hodgson, 2026, Chapter 15). Suppliers are contracted for team capacity, outcomes or retained access to capability rather than fixed scope.
From order-taker to order-maker. The technology function diagnoses the problem behind each request instead of implementing it as specified. At one internal cloud platform, a consuming team requested a specific database configuration, and the platform team’s operational data showed that setup for that database type failed 40% of the time because of ambiguous networking configuration. The team automated provisioning for every consuming team, at a lower cost than the support tickets the old model had been absorbing.
Each commitment fails without the others.
| What is in place | The failure that follows |
|---|---|
| Accountability without persistence | A named Product Manager with no team once Build closes, and no accumulated governance knowledge to carry forward |
| Persistence without consumption-speed finance | The Sprint Review shows the product should pivot, and the pivot waits for the next budget cycle, often a year later, by which time the evidence has changed |
| Financial delegation without accountability | Spend escalates with nobody connecting it to the outcome the product was funded to produce |
| Any three without the order-maker shift | An efficient, persistent, well-governed team solving the wrong problem |
Executives govern AI by designing boundaries
Chapters 8, 9 and 10
The irreversibility boundary is the point after which a wrong decision cannot be corrected in the product’s next iteration. Decisions below it belong to the team, and decisions above it to someone with the authority and context to bear the consequences. For an AI team, the boundary sits at the release decision. The executive’s accountability moves from approving decisions to designing the boundary, and an executive whose teams escalate everything is seeing a signal about how precisely the boundary was drawn.
At the portfolio, four sentences written before money moves turn the Evidence-Based Management (EBM) measures (Scrum.org, 2024) into a control that operates before funds are committed (Hodgson, 2026, Chapter 10). The worked example in Chapter 10 is an onboarding product.
| Sentence | Question it answers | Worked example |
|---|---|---|
| Investment hypothesis | Why invest? | We are investing because 22% of customers abandon onboarding before completion, driven primarily by an 11-day approval time |
| Evidence | What is measurable within the first review cycle? | Within two Sprints, median approval time for the early adopter cohort falls below 8 days and abandonment in that cohort below 18% |
| Threshold | What justifies continued funding? | Funding continues while both measures trend towards the Product Goal targets of 4 days and 12% |
| Decision rule | What happens if the evidence does not arrive? | If neither measure moves within two review cycles, the Product Manager reports which of four explanations applies and the portfolio decides the posture |
Four postures replace the approve-or-cancel choice. The Product Manager decides two of them: Persevere, when the evidence shows progress and the goal still serves the strategy, and Pivot, when progress has stalled but the goal is sound. The portfolio decides the other two: Pause, timeboxed, only when lagging indicators are confirmed as the cause, and Pull, when the outcome has been achieved and the product has reached maturity, when the goal or the measures were wrong with no viable pivot, or when the strategy has moved.
Automation removes judgement's practice ground
Chapter 14
Routine work does a second, invisible job: it builds the judgement the organisation will need in five years. Competence develops through legitimate peripheral participation in real problems alongside experienced practitioners (Lave & Wenger, 1991), which needs genuine accountability, a problem at the edge of current capability, and experienced judgement on hand at the moment of uncertainty. Automating the work removes all three at once, and the AI now being deployed targets exactly that practice ground: compliance assessments, standard correspondence, document classification and contract review.
Evolve adds four fields to every investment case that automates work currently done by practitioners on a developmental pathway:
- Classification. Which developmental pathway does the automated work provide, and for which capability?
- Replacement mechanism. What deliberate structure recreates those conditions, such as rotations with real decisions, shadowing with prediction, or diagnosis of real failures?
- Accountability. Whose name is on the pipeline outcome at the five-year mark?
- Sequencing. Is the replacement operating before the incidental pathway is removed, with a date and a consequence if it is not?
Investment governance must change first
Chapters 13 and 15
The structural changes come in a fixed order, because people act on what the funding model pays for and measures.
| Order | Structural change | Why this order is fixed |
|---|---|---|
| 1 | Investment governance and performance measurement | Determines the accountability structure; every downstream behaviour follows from what the investment model makes rational |
| 2 | Team design and decision rights | Only durable once the funding model makes the new accountability structure rational; a product team funded through stage gates reverts to project behaviour |
| 3 | AI augmentation at scale | Productive only inside the product operating model’s accountability structures; without them, AI augmentation returns task-level efficiency at best |
Source: Hodgson (2026), Table 4.
An insurer in Zen Ex Machina’s (ZXM) practice changed accountability first, moving product management from eliciting requirements to answering for customer outcomes. Eight cross-functional teams then replaced the functional workstreams, and coordination was redesigned last. Over twelve months, features delivered per planning interval rose by 200%, unit cost per feature fell by 67%, cycle time from development to test fell from 122 days to 14, and average overcommitment fell from more than 40% to near zero (Hodgson, 2026, Chapter 13). The people and the technology stayed the same.
On 30 March 2020 the Prime Minister announced JobKeeper, a wage subsidy that would disburse A$89 billion. The Australian Taxation Office was five years into a product operating model, and six value streams pivoted to deliver it without standing up a single new governance structure (ANAO, 2022). In a federal agency’s cloud platform transition, decision latency fell from several months to one day and traceability from strategic objectives to team work rose from 10% to 92% within nine months (Hodgson, 2026, Chapter 13). When the coaching support was withdrawn, the organisation reverted to central decisions, project funding and milestone reporting within six months. Nothing was checking whether the conditions that made it work were still in place.
The first move is one product. Fund it against an outcome rather than a set of deliverables, name the person accountable for that outcome and give them the release decision, then attend the Sprint Review four weeks later and judge the working output against what the money was committed to. Read the next AI business case with two questions: who is accountable for the outcome it funds, and when will you see working output assessed against that outcome? If it cannot answer both, send it back with the one product, the one accountable person and the one Sprint Review that would let it answer.
Companion pages apply the argument
- AI governance assessment: the Appendix A diagnostic, scored against the four commitments.
- When a product changes lifecycle stage: funding by lifecycle stage.
- Moving from stage-gate approval to continuous portfolio investment.
- EBM portfolio dashboard: what to measure and how to read it.
- Governing cloud and AI spend at team level: consumption-speed finance in practice.
- Contracting for team capacity: the procurement boundary.
- Capability pipeline risk assessment: the four fields for automation cases.
- When technology decisions belong at board level.
References
- Australian National Audit Office. (2022). Administration of the JobKeeper scheme (Auditor-General Report No. 22 of 2021–22). Australian Government. https://www.anao.gov.au/work/performance-audit/administration-the-jobkeeper-scheme
- Australian National Audit Office. (2026). Artificial intelligence use in IP Australia. Australian Government. https://www.anao.gov.au/work/performance-audit/artificial-intelligence-use-ip-australia
- Boston Consulting Group. (2025). The widening AI value gap: Build for the future 2025. https://www.bcg.com/publications/2025/are-you-generating-value-from-ai-the-widening-gap
- Bromley, P., & Powell, W. W. (2012). From smoke and mirrors to walking the talk: Decoupling in the contemporary world. Academy of Management Annals, 6(1), 483–530. https://doi.org/10.1080/19416520.2012.684462
- Hannah, A., & Botterill, L. C. (2025). Ignoring harm, saving face: Non-knowledge, senior public servants and the Robodebt scheme. Australian Journal of Political Science, 60(2), 149–162. https://doi.org/10.1080/10361146.2025.2486108
- Hodgson, M. (2026). Evolve: The operating model AI demands. Zen Ex Machina. https://zenexmachina.com/evolve/
- Holmes, C. (2023). Report of the Royal Commission into the Robodebt Scheme. Commonwealth of Australia. https://robodebt.royalcommission.gov.au/publications/report
- Kersten, M. (2018). Project to product: How to survive and thrive in the age of digital disruption with the Flow Framework. IT Revolution Press.
- Lave, J., & Wenger, E. (1991). Situated learning: Legitimate peripheral participation. Cambridge University Press.
- Mao, Y., & Mirhoseini, A. (2026). Decentralized multi-agent systems with shared context (arXiv:2606.10662). arXiv. https://doi.org/10.48550/arXiv.2606.10662
- Scrum.org. (2024). The evidence-based management guide: Improving value delivery under conditions of uncertainty. https://www.scrum.org/resources/evidence-based-management-guide
