.png)
The AI-Native Operating Model (Artefact-Driven Governance)
by Didier Vila, PhD – Founder and MD of Alpha Matica and by Marta Smith,Poland MD, McGregor Boyall Associates Ltd
AI does not create value by itself; it amplifies the quality of the underlying operating model. In well-designed organisations it accelerates value flow. In bureaucratic ones it accelerates waste, technical debt, and downstream chaos. This is the central finding of Google Cloud’s 2026 DORA research into AI-assisted software development. [2]
The core problem is coordination. Traditional middle management carried many responsibilities — negotiation, resource allocation, people development — but a substantial share of it was spent moving information, aligning work, and resolving dependencies. When companies remove those layers without redesigning how work and truth move across the organisation, they risk what we call an instability tax: local bursts of AI-generated output that overwhelm review capacity and create new bottlenecks. We use this term as our own extension of DORA’s amplification finding, not as a separately measured result — it is the mechanism we believe explains why delayering-without-redesign fails, and it is a hypothesis this paper argues rather than one that has independently been tested against case data.
The alternative is an artefact-driven operating model — one in which verifiable digital artefacts (code, logs, data pipelines, automated evaluations, and machine-readable agreements) replace much of the human coordination bureaucracy. This builds directly on our earlier analysis of how generative AI breaks the traditional link between headcount growth and management-layer growth. [5]
The Verification Tax
Even capable organisations experience a temporary productivity dip when they scale agentic systems. DORA describes this as a J-curve, driven by learning-curve effects, downstream process adaptation, and — the largest factor — the verification tax: the cognitive load of reviewing, testing, and correcting large volumes of AI-generated work. [2]
The economics reinforce the point. McKinsey’s 2026 analysis of agentic AI costs finds that roughly 60% of agentic task expense already goes to checking and repairing model output rather than to inference itself [1] — the verification tax shows up on the P&L as much as on engineers’ calendars.
Without stronger automated guardrails and clearer ownership, increased output velocity is often offset by increased rework and review burden. The organisation generates more, but delivers less net value.
Escaping this pattern requires moving away from management by status reports and toward an operating model in which the quality and validity of work can be inspected directly through artefacts.
What artefact-driven actually means
The alternative is not “fewer meetings.” It is a designed set of objects that carry intent, interfaces and proof. In practice, five artefacts replace much of the coordination bureaucracy:
A mission card — the outcome, constraints and goals the team and its agents are allowed to optimise for.
An interface contract — the machine-readable agreement between teams: APIs, data contracts, service levels, and ownership of failure.
An evaluation suite — automated checks that define “good enough” before a human reviews.
An execution trail — pull requests, pipeline runs and agent logs that show what happened, not what was reported.
An exception ledger — the human decisions that cannot yet be encoded, so judgement stays visible instead of disappearing into chat.
If an artefact cannot be inspected, it is not part of the operating model. It is folklore.
Three Practical Shifts
To successfully transition to an artefact-driven operating model, organisations must move beyond simply adopting new tools and instead fundamentally reorganise how work is executed, coordinated, and validated. The three shifts below are our proposed design response to that diagnosis. They draw on established team-design research [3][4], but their combination — and their application to agentic AI specifically — is a framework we are putting forward for testing, not yet a result validated at scale. Organisations adopting it should track outcomes deliberately rather than assume the pattern transfers automatically.
Full-Cycle Teams
Instead of fragmented handoffs across specialised functions, small cross-functional teams (typically 3–7 people) own outcomes end-to-end. AI absorbs many narrow execution tasks, allowing the team to move from idea to delivery with fewer external dependencies. This reduces coordination overhead and gives agents a clearer mission.
Teams as Interfaces
When hierarchy is reduced, alignment cannot rely on meetings and escalation. Dependencies between teams should be made explicit through clear, preferably machine-readable interfaces and automated quality checks. This is consistent with the emphasis in Team Topologies on stable interfaces and reduced cognitive load as foundations for fast flow. [3][4]
Grounded Verification
In flat structures, leaders can no longer rely on polished summaries. Oversight must include direct sampling of real artefacts — pull requests, agent logs, pipeline definitions, evaluation results. This is closer to how a CFO samples raw ledgers than to traditional management reporting. It keeps the organisation honest.
This only works if the sampling itself is disciplined. Asking leaders to review artefacts exhaustively simply moves the verification tax up a level rather than removing it. The practical answer is risk-weighted sampling: automated checks and anomaly detection triage the bulk of output, and human review — leadership included — is reserved for what is flagged as high-risk, high-impact, or statistically out of pattern. Grounded verification is a targeting discipline, not a mandate to read everything, and it extends the grounded-leadership model we set out previously: algorithmic literacy paired with the judgement to know what still needs a human look. [5]
The Human Shift: From Information Router to Socio-Technical Coach
A significant share of traditional middle management time was spent gathering status, translating information, and routing decisions. AI systems now perform much of that work.
This does not mean coordination was middle management’s only value — judgement, mentorship, and conflict resolution were always part of the job. What changes is that the lowest-value share of the role, routing and translating information, is now largely automatable. That frees, and requires, leaders to spend more of their time on what was always the harder part.
The remaining leadership role is different. It is less about controlling information flow and more about shaping the socio-technical system: developing people who can work effectively with agents, maintaining psychological safety, resolving complex human and ethical issues, and ensuring the overall system remains coherent.
This is not the elimination of management. It is a change in what management is for.
What Changes in Practice
Dimension | Legacy Model | Artefact-Driven Model |
Coordination | Meetings, escalations, status reports | Explicit interfaces and automated checks |
Source of truth | Presentations and summaries | Code, logs, data, and evaluation artefacts |
Primary risk | Slow decisions and hidden problems | Unmanaged output volume and review overload |
Leadership focus | Headcount and activity tracking | System health, talent, and value flow |
These are directional shifts, not a single cutover. Most organisations will run a blend of both columns during a multi-year transition, and the failures we see most often happen when the legacy column’s coordination habits are removed before the artefact-driven column’s discipline is in place.
Conclusion
Structural simplification only creates advantages when it is paired with a new operating model. Removing layers while leaving coordination, verification, and ownership unchanged simply moves the bureaucracy around.
The organisations that will extract real value from agentic AI will treat the operating model itself as a design problem. They will reduce reliance on human information routing, make interfaces and quality visible through artefacts, and redefine leadership around the stewardship of socio-technical systems.
Architecture — not headcount or tool count — becomes the constraint that matters.
A note on evidence. The diagnosis in this paper — that AI amplifies existing organisational quality, and that a verification tax is the primary drag on ROI — rests on current, credible third-party research [1][2]. The prescription — full-cycle teams, machine-readable interfaces, and risk-weighted grounded verification — is our design response to that diagnosis. It is informed by established team-design research [3][4] and extends our own earlier work [5], but it has not yet been independently validated at scale for agentic AI specifically. We believe it is the right architecture to test, and we would welcome case data — ours or others’ — that confirms or corrects it.
References
1. Hämäläinen, L., Patel, M., Blumberg, S., Catlin, T., & Lala, W. (2026, July 13). Is that AI agent worth it? Agentic economics and the modern operating model. McKinsey Quarterly.
2. Google Cloud / DORA. (2026). The ROI of AI-assisted software development.
3. Skelton, M. (2026, July). The organizational foundation for AI ROI (Executive Report SG-808). Team Topologies.
4. Skelton, M., & Pais, M. (2019). Team Topologies: Organizing business and technology teams for fast flow. IT Revolution Press.
5. Vila, D., & Smith, M. (2026, August 20). The generative AI organisational reset. Alpha Matica and McGregor Boyall
Comments