An AI pilot can work for a small team while leaving the wider organisation almost unchanged. A new AI News interview with NTT DATA executives puts that gap at the centre of the enterprise adoption debate: integration, process design and accountability matter alongside the model itself.
The interview accompanies the company’s ServiceNow Insights and Vision Report 2026/27. Its publicly available findings raise a useful question about evidence. Being able to demonstrate a tool is different from being able to demonstrate a repeatable business result, and neither should be assumed from enthusiasm alone.
Fragmented Deployment Is Not a Failure Rate
NTT DATA’s public report summary says 59% of respondents describe AI as confined to small pilots or individual functions rather than unified deployment. It also says 47% describe demonstrated AI returns as anecdotal, while 52% identify organisational change and adoption management as the most important partner capabilities.
Those figures describe different questions. They should not be added together or converted into a claim that most AI systems do not work. A functional deployment may remain isolated; an anecdotal benefit may be real but insufficiently measured. Neither category directly supplies a controlled estimate of financial return.
The open landing page presents selected results rather than the complete report, whose download requires registration. The full survey instrument and sampling details were not reviewed here. These are therefore vendor-published findings, not a representative estimate independently established for every enterprise. That limitation should travel with the numbers whenever they are used to support a broader argument about adoption.
The Executives’ Argument Starts With the Workflow
In the AI News interview, NTT DATA’s Jesus Ordonez Garcia argues that scaling means connecting a successful experiment to everyday processes, existing systems, security and governance. Jeroen van Hamersveld emphasises employee preparation and the transition from local experimentation to organisational value. Both are describing their company’s approach, not reporting independently audited customer outcomes.
Their argument also distinguishes redesign from acceleration. Automating an inefficient process can reproduce its weaknesses more quickly. Ordonez Garcia consequently places process redesign before the expansion of autonomous decision-making, while van Hamersveld stresses coordination between agents rather than a collection of disconnected tools.
That framing makes the demonstration itself a starting point. A practical assessment would need to ask what changed in the whole workflow: whether staff still have to repair outputs, whether exceptions move somewhere else and whether the benefit survives ordinary operating conditions. These are editorial questions suggested by the interview, not additional findings from NTT DATA’s survey or claims about an unnamed customer.
Older Research Helps Separate Causes From Headlines
A RAND study published in August 2024 provides relevant, explicitly older context. Based on interviews with 65 experienced practitioners, it identifies recurring problems including poorly specified business needs, inadequate data, technology-first projects, weak deployment infrastructure and tasks beyond a system’s capabilities. It is qualitative research into perceived causes, not a new census of AI project outcomes.
RAND’s scope also matters: projects that simply used pretrained language models were excluded. Its widely repeated background reference to estimates exceeding an 80% failure rate should not be treated as the measured result of those interviews, or transferred without qualification to today’s agent deployments.
The NIST AI Risk Management Framework core, originating in 2023 and currently under revision, offers a separate reference point. It connects governance, context, measurement and management throughout a system’s life. Its guidance includes documented responsibilities and assessment in conditions resembling deployment. It is voluntary guidance, not proof of commercial success or a certification of NTT DATA’s approach.
Read together, the sources support a narrower conclusion than an industry-wide failure headline. The new interview and selected survey findings identify integration and measurement as live concerns. Older research and guidance help explain how to investigate them. The decisive evidence would be a clearly defined process improvement, measured with its costs and limitations, rather than the number of pilots an organisation can announce.
