Executive Summary
AI projects fail far more often on organisational alignment than on technology. Whether users adopt a system is the strongest predictor of whether it survives past the pilot stage. How clever the algorithm is matters less.
Core conclusions
- The pattern holds across cases. AI succeeds when subject-matter experts and frontline operators define the problem it should solve. It fails when technology teams design solutions apart from the people who’ll use them.
- Organisations that scale reverse the sequence. They start from a real operational problem rather than “which AI tool should we use,” and that is what sets them apart from those stuck in pilot programmes.
- Technology teams stop deciding what the organisation needs. Their job becomes to support domain experts. That means infrastructure people can reach, governance built in, one shared data foundation, and translation between business outcomes and technical capability.
The argument, in ten slides
Save it, share it, or send it to whoever owns your next AI rollout.
Abandonment of AI initiatives more than doubled in a year
Share of organisations that abandoned most of their AI initiatives in the year. Source: MIT NANDA 2025 study, as cited in the Evidence and Methodology box below.
For eighteen months, we watched enterprise AI projects multiply. Millions went into vendor products, consultancy work, proof-of-concepts, and pilot programmes. The promises were huge: automation at scale, predictive intelligence built into operations, an edge over competitors. Yet data from 2025 shows the sobering reality. Abandonment is running at the rate shown above, and a further 75 to 85% of projects never reach production. MIT research confirms the technology is not the cause. The failure is in execution, rooted in organisational misalignment: I go through the specific failure patterns behind that number elsewhere.
The pattern is now unmistakable. AI succeeds when it solves real operational problems for the people who must use it daily. AI fails when technology teams design solutions on their own, when business leaders and technology are disconnected, and when the organisation treats implementation as a technical exercise.
The lesson that separates successful organisations from those stuck in pilot programmes is plain. Put subject-matter experts and frontline operators in charge of choosing which problems AI should solve. Then set up technology teams to build the infrastructure that supports those choices.
The Misalignment Trap
Most organisations treat AI as a technology problem rather than an alignment problem. When deployment begins, three stakeholder groups (CEO, HR, and Finance) each hold a different, often unspoken idea of success. If they have not agreed these definitions in advance, every unexpected AI behaviour turns into a political fight. The project collapses even when the technology works.
This pattern repeats across the enterprise. Data scientists celebrate a model that reaches 94% accuracy in the demo environment. Finance cannot say whether the system produces measurable ROI, because no one set baseline metrics. Operations cannot explain why staff are sabotaging the system, because no one aligned incentives before deployment. 54% of executives cite cultural resistance as the main barrier to AI adoption. What they are describing is the absence of alignment, not the absence of capability.
What they are describing is the absence of alignment, not the absence of capability.
The deeper problem appears when technology teams lead discovery. The question “Which AI tool should we use?” comes before anyone understands the business problem. This technology-first approach is the single largest driver of AI project failures. It leads to misplaced investments, wasted resources, and solutions cut off from what operations need.
The case for user-centred design
When organisations reverse this sequence, starting with real operational problems and working back to technology, adoption changes completely. A financial services organisation building fraud detection AI set up a “user council” of fraud analysts from operations. The council met bi-weekly throughout development. It gave immediate feedback on interface design, alert structures, and how the tool fitted the workflow. Adoption reached 94% within three months. Earlier analytics tools typically struggled to reach 40%.
A manufacturing company rolling out predictive maintenance AI first deployed a dashboard full of technical metrics and probability scores. When adoption stalled, they redesigned the screen to show red/yellow/green status indicators. The deeper analytics sat behind them for maintenance teams to open when needed. Daily system usage rose 380%, and maintenance outcomes improved measurably.
These cases point to a critical finding. User adoption is the single most reliable predictor of AI project success. It matters more than how sophisticated the algorithm is, how good the data is, or how well the technical infrastructure performs. The constraint is not building good AI. The constraint is building organisational agreement on what “working” means and proving that the system consistently delivers that outcome.
The constraint is not building good AI. The constraint is building organisational agreement on what “working” means.
Domain Experts as Decision-Makers
Subject matter experts know the details, limits, and daily realities of their field in a way no one else can. A healthcare organisation building clinical decision support did not start with new algorithms. It started by shadowing physicians and finding their most pressing daily pain point: documentation was taking up 38% of clinical time. That finding moved the AI focus away from exotic diagnostic features and toward ordinary but valuable workflow automation. Physicians adopted it by choice because it dealt with the real constraint they faced.
Domain experts keep AI solutions accurate, relevant, and grounded in how things work in the field. They spot edge cases, unclear scenarios, and exceptions that raw data alone cannot capture. They understand safety regulations, resource limits, real-time processing needs, and business rules. In machine learning terms, they are your most reliable source of signal in the training data. They are also the most credible people to check model outputs.
When organisations place domain experts directly in AI development, innovation speeds up. Some organisations use decentralised governance models, where domain experts make the core decisions and technology teams support the infrastructure. They report ROI multipliers of 1.7 to 6.6 times those of technology-led models. Allegis Group restructured around a centralised Data Science Hub. Its hub-and-spoke model made business-unit experts the decision-makers, supported by data scientists and engineers. This let the company scale quickly across the organisation without copying specialist technical teams into every function.
Technology Teams as Enablers
This change asks technology teams to see their role in a very different way. They become infrastructure providers who help experts make the decisions.
Successful technology teams in this model do four specific things:
-
They establish accessible infrastructure. Platform engineering approaches treat developers and domain experts as customers. Containerisation, or data-pipeline orchestration, platforms hide complexity behind self-service interfaces. Intelligent internal developer platforms (IDPs) build in AI-assisted guidance, policy enforcement, and automated provisioning. They remove the barriers between an expert’s insight and putting it into operation.
-
They embed governance into systems. In mature organisations, governance is no longer bolted on after deployment. It is built into products through policy enforcement, model versioning, explainability by default, and audit-ready logging. Technology teams design these limits together with compliance and risk teams, so domain experts work inside guardrails.
-
They maintain unified data foundations. The most frequent cause of AI project failure in Asia-Pacific is data scattered across on-premises infrastructure, multiple cloud platforms, and edge networks. Technology teams that bridge these silos let domain experts build AI solutions without first solving the underlying data engineering problems. Clean, connected, accessible data infrastructure is non-negotiable. Everything else rests on it.
-
They translate between business outcome and technical capability. This role needs technical leaders with three skills. They need AI fluency to know what is technically feasible, strategic thinking to link capabilities to business goals, and cross-functional coordination to keep business, operations, and technology aligned. These people are the bridge between defining the problem and delivering the technology. They keep solutions focused on measurable outcomes.
Operational Structure for Scale
Organisations that have scaled AI successfully usually adopt one of two governance patterns:
The decentralised model gives business units ownership of their domains, each with its own AI capability. A central platform team provides data infrastructure, governance guardrails, and shared ML operations (MLOps) tooling. This approach brings the most domain experts into investment decisions and points innovation at high-impact operational problems. Scale grows because more teams can innovate in parallel, each guided by the limits of its own domain.
The hub-and-spoke model (also described as a Competency Centre of Excellence) keeps a central data science and engineering team that works in partnership with domain experts across business units. These partnerships run as ongoing collaboration rather than one-off project advice. Quarterly evolution workshops let domain experts share what they have learned, agree together what to improve next, and own where the capability goes. This approach keeps specialist talent together where it is most scarce, whilst handing decisions to the people who understand the operational impact.
Both models need explicit governance. People often see governance as a brake on innovation. In reality, organisations with mature governance frameworks are speeding up AI adoption, because governance gives them the confidence to scale. Executive ownership of AI governance roles reflects this shift. These executives make it possible to scale decisions across the enterprise with confidence, and they do not slow those decisions down.
The maturation cycle is putting subject matter experts back in charge
The AI hype cycle is fading, and a maturation cycle is taking its place. Organisations are learning that AI succeeds when subject matter experts drive strategy, define problems, check solutions, and own outcomes. Technology teams play an essential supporting role. They build the infrastructure, maintain governance guardrails, bridge data silos, and make fast experiments possible, without asking domain experts to become infrastructure specialists.
Evidence & Methodology
Only one number in this piece has a public source behind it. The rest, the user-council case, the dashboard redesign, the ROI multipliers, are illustrative examples I have seen the pattern of repeatedly, without a named company or public citation I can point you to. Here is which is which.
| Claim | Source | Grade |
|---|---|---|
| 42% of organisations abandoned most AI initiatives in a year, up from 17% in 2024 | MIT NANDA’s 2025 study | Measured |
| A fraud-detection user council lifted adoption to 94% in three months, versus 40% historically | Illustrative case, company not named, no public citation | Unverified |
| A dashboard redesign lifted daily maintenance-system usage by 380% | Illustrative case, no public citation | Unverified |
| Decentralised governance models report ROI multipliers of 1.7 to 6.6 times technology-led models | No public source I can point you to | Unverified |
Sources
- MIT NANDA. (2025). The GenAI divide: State of AI in business 2025.
Free tool
Human-AI Interaction & Decision Quality
Diagnose where the handoff between AI speed and human judgement is breaking down in your own teams.