Delivered on time, then quietly reversed
Picture a familiar story. An AI roll-out goes live on time, within budget, and with executive applause. Six months later, customer complaints spike, manual workarounds creep back in, and the team that built the model has long since moved on. What fails is not the algorithm. It is the assumption that intelligence could be delivered, signed off, and left to run.
That moment is becoming familiar. Across sectors, AI initiatives that look successful at launch are unravelling in production. Not because the technology is immature, but because the way organisations execute strategy has not caught up.
AI breaks the project model
For decades, strategy execution has been organised around projects. Clear scope. Fixed timelines. Defined end points. Deliver, hand over, move on.
AI breaks every one of those assumptions.
An AI system does not stabilise at launch. It drifts. Data changes. User behaviour evolves. Models require retraining. Governance requirements tighten. The environment that shaped the original design is gone within months. Treating that reality as a post-project “operations issue” is one of the biggest causes of value erosion.
The evidence points the same way. MIT's NANDA initiative reported in 2025 that 95% of the organisations it studied had seen no measurable profit-and-loss impact from their generative AI pilots, with integration and operating model, rather than model accuracy, as the main obstacles. Regulation assumes the same lifecycle view: Article 72 of the EU AI Act requires post-market monitoring of high-risk systems for as long as they are in use.
The language of projects obscures this failure. When success is defined as delivery, degradation after go-live is invisible to governance. The project is closed. The metrics look green. Meanwhile, performance slips quietly until the system is either rebuilt or abandoned.
Leading organisations have reached a blunt conclusion: AI cannot be executed as a finite initiative. It must be operated as a living system.
Systems need different funding, governance and ownership
This realisation forces a structural rethink of how strategy is designed and governed.
In a project model, risk is assumed to decrease as delivery progresses. In AI systems, risk accumulates over time. Bias emerges. Feedback loops amplify errors. Regulatory expectations evolve. The moment of highest exposure is often long after launch, precisely when project oversight has ended.
By contrast, organisations that treat AI as a system design problem align execution with reality. They fund continuously rather than episodically. They assign long-lived owners rather than temporary sponsors. They govern through ongoing assurance rather than stage gates.
This shift is already visible. Product-centric operating models, long common in digital platforms, are becoming the default for AI. Instead of asking “what projects are we running?”, leadership asks “what systems are we responsible for, and how are they evolving?”
The implications are profound:
- Funding moves from one-off capital allocation to persistent investment. AI budgets do not end at go-live; they blend run and change into a single value stream.
- Governance shifts from milestone approval to continuous monitoring. Oversight becomes embedded in operations, not exercised periodically by committees.
- Ownership becomes explicit and durable. A named leader remains accountable for outcomes years after deployment, not just for delivery.
This is why 2026 matters. Boards are discovering that AI strategies framed as project portfolios look busy but compound nothing. Meanwhile, competitors operating a smaller number of well-owned systems are pulling ahead, not through superior models, but through superior operating discipline.
From a project portfolio to a system portfolio
The shift is easier to make than it sounds, because most of the information already exists in the project portfolio. Five steps convert it.
- Group projects into systems. List every AI project, live or planned, and group them by the operational system they change, such as claims triage, demand forecasting or customer service. Several projects usually turn out to be versions of one system.
- Name an owner for each system. One leader, accountable for its outcomes for as long as it runs, not a sponsor who moves on at go-live.
- Fund the system, not the release. Give each system a standing budget that covers running costs, monitoring, retraining and improvement, reviewed quarterly against its outcomes.
- Replace stage gates with reviews. Instead of approval at milestones, review each system's performance, risk and cost on a fixed cycle, with authority to invest more, change course or retire it.
- Set retirement criteria. Decide in advance what would end a system, so stopping is a planned decision rather than a quiet abandonment.
The controls each system needs through its life are set out in Your AI agents need lifecycle management, and the exit criteria that move a pilot into the portfolio in From Pilots to Permanence.
From delivery to stewardship
For leaders, this shift is uncomfortable because it challenges familiar sources of certainty.
Projects offer closure. They end. Systems do not. They expose leadership to ongoing accountability, to metrics that never settle, to performance that can improve or deteriorate at any time.
For teams, the change is equally stark. In a project world, success is delivery. In a system world, success is stewardship. The work does not peak at launch; it begins there. Engineers, data scientists, risk teams, and product leaders must collaborate continuously, not sequentially.
And for customers, the difference is tangible. You experience the gap immediately when an AI system has no owner left to learn from its mistakes. Errors repeat. Trust erodes. Confidence fades. Systems without stewardship feel indifferent; systems with it feel responsive.
The human challenge, then, is not learning to build AI. It is learning to live with it, to accept that strategy execution is continuous rather than episodic.
Advantage is maintained, not delivered
The dividing line is becoming clear. Some organisations will still be asking which AI projects they should launch next. Others will be focused on which systems they are prepared to own for the next decade.
The difference is not tooling or talent. It is mindset.
Projects assume stability. AI guarantees change. Strategy now belongs to those willing to design for evolution rather than completion.
In the AI era, competitive advantage is not delivered. It is maintained.
