Insights and perspectives
Industrialising AI in IT: The Three Pillars CIOs Can No Longer Afford to Ignore
Why AI-augmented Agile is a dead end — and how the DevOps–FinOps–Governance triptych gives CIOs back operational and financial control of AI.
Driven by now-mature tooling — Claude Code, GitHub Copilot, Cursor — and the emergence in 2026 of autonomous development agents such as Kiro (AWS), Copilot Coding Agent (GitHub) and Devin (Cognition), CIOs have formally committed to the industrialisation of AI in software delivery. The signal is unambiguous: according to JetBrains’ AI Pulse survey (January 2026, 10,000 developers worldwide), 90% of developers already use AI in their day-to-day work — and 22% rely on autonomous coding agents. The strategic question is therefore no longer “should we adopt AI?”, but “on what methodological foundation should it rest in order to generate quantifiable value?”.
The Illusion of AI-Augmented Agility
A particular school of thought is positioning itself as the new state of the art by promoting an Agile approach — assumed to be mature — augmented with artificial intelligence. The concept: plug AI agents into well-drilled Scrum teams, enrich ceremonies, and multiply backlog throughput. Our field engagements tell a different story. In the vast majority of organisations, Agile adoption has remained incomplete: teams do produce features, epics and user stories — yet when the sprint begins, the essentials are missing, with development driven by urgency rather than design. Layering AI on top of this heterodox Agile practice amounts to automating imperfection — or worse, amplifying it — thereby increasing operational risk. Scrum is structurally incompatible with agentic AI because it was designed to address constraints that AI naturally resolves:
| Scrum Constraint (2001) | What AI Changes in 2026 |
|---|---|
| Two-week sprints — designed to synchronise humans working 8 hours a day | Agents execute 24/7, without pause or milestones. McKinsey documents the emergence of a new model: a morning daily sprint for humans, overnight agent execution — a continuous loop that renders the two-week sprint structurally obsolete. |
| Teams of 7 ± 2 — to contain the cost of human coordination | When agents assume coordination responsibilities, the team size constraint disappears. BCG Platinion documents the Spotify case: 650 AI pull requests merged per month, with zero lines of code written manually since December 2025. |
| User stories — to break work into units comprehensible by humans | Agents do not need narrative accounts — they need precise, executable specifications. McKinsey and Thoughtworksconverge on spec-driven development as the new standard. |
| Velocity — the reference KPI for human productivity | Agents operate continuously, without sprints or breaks. If teams are not organised to absorb that pace, the organisation slows down precisely where it expected to accelerate, creating a bottleneck. |
DevOps: Native Compatibility with Agentic AI
Where Agile stumbles on a collective posture that is difficult to standardise, DevOps offers the exact opposite: a tooled, sequenced and measurable chain in which every step produces artefacts that a machine can act upon — code versioned in Git, structured tickets in Jira, automated tests, infrastructure described as code (Infrastructure as Code — IaC), operational metrics. This is precisely what makes AI operable: it requires standardised inputs in order to produce reliable outputs. Each pillar of the CALMS framework illustrates this:
| CALMS Pillar | What It Means for AI Agents |
|---|---|
| Culture of shared responsibility | This is the prerequisite for the orchestration framework — agent harnessing — that connects, supervises and controls AI agents to function without silos. |
| Automation (CI/CD, testing) | CI/CD pipelines are no longer simply code conveyors: they become the execution loops within which agents operate, iterate and deliver — potentially around the clock. |
| Lean (continuous flow, WIP reduction) | Where Scrum imposes end-of-sprint milestones, Lean DevOps aligns naturally with the continuous cadence of agents. |
| Measurement (DORA metrics, SLOs, traces) | Observability is what makes the harness effective. Without reliable data, the agent is inclined to fabricate rather than verify. Hallucination is not an AI bug — it is the consequence of a measurement failure. |
| Sharing (runbooks, IaC, documentation) | Shared context is the agent's "memory" via artefacts — DevOps runbooks become its operating instructions (sometimes formalised as AI skills). |
Three realities that make DevOps the native framework for agentic AI:
- A sequenced, tooled chain of roles.Business Analysts (BA), Software Engineers (SE) and Quality Assurance (QA) intervene at reproducible stages — design, development, testing, deployment, operations — each underpinned by a dedicated tool and a dedicated agent. Value does not reside in each role in isolation, but in their interconnection: the outputs of a BA agent directly feed the Software Engineer’s context, test results surface in real time to the QA/SRE, and overall velocity improves accordingly.
- A pioneering vendor market.Harness, Atlassian, GitLab and Digital.ai had integrated specialised agents at every stage of their CI/CD platforms by 2025. Where AI-augmented Agile remains a methodological promise, AI-augmented DevOps is already being adopted by technology leaders (Synapsys, 2025).
- A “FinOps-compatible” decomposition of work.An agent is not billed by the hour, but by tokens, model calls and inference cycles — a reality that remains massively overlooked: 96% of organisations report AI costs exceeding their projections FinOps Foundation, “State of FinOps 2026″). DevOps, by naturally sequencing work, is the privileged framework for carrying the economic modelling of agents.
The result: where AI-augmented Agile remains a slogan, AI-augmented DevOps is already an industrialisable reality. That said, this compatibility is not unconditional. AI-augmented DevOps requires concrete prerequisites: robust, battle-tested agents (eliminating hallucination risk) and teams trained in new AI configurations (models, context windows, etc.). Without these foundations, the CI/CD pipeline becomes an error amplifier, not a value lever (Eficode, “Transforming software development with AI and DevOps”).
FinOps: Making the ROI of AI Quantifiable
AI-augmented DevOps — by decomposing work into “AI-measurable” tasks and relying on a fully instrumented CI/CD chain — opens the door to FinOps modelling extended to AI: a decisive advantage for executive committees.
According to the “McKinsey Global Survey on the State of AI” (November 2025), 88% of organisations already use AI in at least one function — yet only 6% of them are able to attribute a measurable impact to their operational results.
Conversely, the FinOps Foundation’s “State of FinOps 2026” reveals that 98% of “AI-mature” organisations now actively manage their AI expenditure (compared with 63% in 2025 and just 31% in 2024).
In two years, AI cost management has evolved from a peripheral concern to an existential responsibility. Billing models have multiplied: platform subscriptions, consumption-based pricing (tokens yesterday, multi-model credits today), cloud costs per deployment, on-premises inference costs. This complexity makes AI spend governance not merely optional, but structurally necessary — failing which, AI OPEX rapidly spirals beyond any meaningful control.
A well-instrumented DevOps chain enables every task, every agent and every pipeline to be associated with a resource consumption figure and a unit cost:
« For this type of feature, our AI cost is €X, our infrastructure and services cost €Y, the ROI is Z%” — The conversation every executive committee should be able to have with its CIO in 2026.
This view — native to DevOps — enables decision-makers to visualise the ROI of their AI OPEX in concrete terms, to rationalise and optimise their AI usage, and to transform it into a controlled, measurable asset.
The CIO’s Operating Model: the Blind Spot of AI Adoption
Adopting DevOps and FinOps is necessary — but not sufficient. A CIO organisation can master its pipelines, manage its AI expenditure and still fail to industrialise AI for another reason: its operating model and associated governance are simply not fit for purpose. This is the blind spot in most of the AI roadmaps we encounter. Not all configurations are equal in the face of AI:
- Decentralised organisations by subsidiary or business unit — still prevalent in industrial groups — are structurally incompatible with an AI strategy. Licences are purchased in a fragmented manner, without pooling or economies of scale; AI governance becomes impossible to operate (traceability, EU AI Act compliance, security): AI degenerates into a costly patchwork of local initiatives.
- Shared Services (CSP) or Bimodal configurations (theorised by Gartner) deliver economies of scale but struggle to absorb the velocity of generative AI. The CSP produces static service catalogues, disconnected from real business needs. Gartner’s model, which separates run IT (stability) from innovation IT (velocity), traps AI in the latter and makes POC industrialisation impossible, for lack of a bridge between the two IT cultures.
- Hybrid configurations (centralised + decentralised) and Hub & Spoke models (centre of excellence + local teams) emerge as the only genuinely AI-compatible architectures: a central hub that concentrates AI components (data, foundation models, FinOps, governance, EU AI Act compliance) and autonomous product teams or spokes that consume these services and develop their own use cases. By mastering the AI environment at group level, it becomes possible to tailor tools and solutions to each business need — translating central strategy into local operational value.
The regulatory constraint reinforces this logic: the EU AI Act mandates group-wide traceability of AI systems, classification of use cases by risk level, and real-time monitoring of agentic processes — obligations that are impossible to fulfil without a central governance body. This is precisely the purpose of the “AI-Authority”, addressed in our insight “The Challenges of AI Adoption in the Enterprise.” This central body operates in consultation with all dimensions of the executive committee in order to define the scope of the possible and set the strategic guardrails for AI at group scale.
DevOps and FinOps are the two pillars of operational performance in the AI era — one industrialises delivery, the other masters its cost. But both frameworks only realise their full potential when underpinned by fit-for-purpose governance, including an AI-Authority capable of carrying this strategic ambition.
Three Questions to Address Before the End of 2026:
- Are our CI/CD chains ready to mobilise AI agents at every stage — or are we still deploying an augmented yet imperfect Agile practice?
- Are we capable, today, of producing a TCO per feature that includes AI costs?
- Is our IT governance — often still decentralised or fragmented by subsidiary — truly capable of driving an AI strategy at scale, or are we about to layer another incompatible technology on top?
If the answer to any one of these questions is “no”, the priority is not to accelerate adoption — but to build a foundation designed for AI. At Valthena, we are convinced that the right question is: “What DevOps–FinOps–Governance triptych do we put in place so that AI is operationally effective, economically controlled, and governable at scale?” Addressing these three dimensions is the purpose of our AI Transformation offering, which accompanies organisations from adoption through to full industrialisation.
“AI is not a product you install — it is a transformation you lead.” — Valthena, “The Challenges of AI Adoption in the Enterprise”, March 2026
Get in touch with our teams: contact@valthena.com
