Vinod Bidarkoppa: AI in a Broken Process Is Just Faster Inefficiency

AI adoption is no longer the scarce asset. Enterprise scale is. Closing Day One of Auckland’s CIO Innovation Summit, former Walmart International chief technology officer Vinod Bidarkoppa argued that the distance between the two has much less to do with model capability than the data, processes, platforms, economics and operating choices sitting underneath it.

TL;DR

• Widespread AI use is not the same as transformation. Bidarkoppa’s test is whether it has been built into an end-to-end process that produces a measurable enterprise outcome.
• Fix the data foundation first: named owners, active stewards, clear lineage, consistent definitions and reusable data products.
• Build common platform components once and deploy them many times, then add only the local regulation, currency, branding or business variation that is genuinely necessary.
• Plan for agent consumption costs before production. Per-seat software budgeting becomes unreliable when agents call models, tools and other agents as usage grows.
• Separate assist, augment and autonomise use cases. The largest gains sit in redesigning the process, not simply automating the existing task.
• The modern CIO has to hold speed and governance, innovation and resilience, and AI capability and human judgement at the same time.

Adoption is not transformation

Bidarkoppa arrived with a scale of experience few CIOs ever encounter. His career has crossed airline reservations, international retail, healthcare and Walmart International. Walmart most recently reported US$713 billion in FY2026 revenue, approximately 2.1 million associates and operations across 19 countries. His keynote drew on that scale, but its central warning applied just as readily to a New Zealand organisation: using AI somewhere is not evidence that the enterprise has changed.

McKinsey’s 2025 State of AI survey, based on 1,993 respondents, found that 88% reported regular AI use in at least one business function. Yet only about one-third said their organisations had begun scaling AI programmes across the enterprise. Bidarkoppa also pointed to a widely debated MIT NANDA report, which was reported as finding that 95% of generative AI pilots in its dataset produced no measurable profit-and-loss impact. He explicitly called the research controversial. The useful point was not to turn that figure into a universal law, but to confront the gap between activity and outcomes.

His diagnosis had two parts. Organisations are trying to build AI on fragmented, duplicated and inconsistently governed data. They are also laying it over processes designed years or decades before autonomous systems were possible. Faster execution does not make a poor process valuable.

"When you try to put AI in a broken process, all you get is faster inefficiency."

That distinction is especially important when pilots are easy to launch. A demonstration can satisfy a management team, produce a board slide and still disappear because nobody designed it for production, ownership, integration or scale. Bidarkoppa called these “zombie pilots”: initiatives that remain visible enough to consume attention but never become part of how the business works.

Start with data products, not model selection

The models were not where Bidarkoppa started his architecture. He began with the data layer, and with the organisational behaviour that determines whether that layer remains usable.

Large enterprises commonly have a central repository that appears coherent from a distance. Business teams then take snapshots for local analysis, build their own definitions and create parallel pipelines. Over time, the organisation accumulates multiple versions of the same customer, item, inventory or financial truth. Each new AI use case then needs fresh integration work, more compute, more storage and another argument about which number is correct.

"The prerequisite for great AI is really data."

His prescription was concrete: give data a named business owner, appoint stewards who maintain quality, preserve lineage so people can see where information came from, and standardise the semantic definition as well as the physical implementation. Package high-value domains as reusable data products so the organisation consumes the same governed customer, inventory or item foundation rather than rebuilding it for every business unit.

That is not a data-office clean-up exercise sitting beside the AI strategy. It is the AI strategy’s foundation. If ownership, quality and access are ambiguous for people, agents will inherit the ambiguity at machine speed.

Build the chassis once

Bidarkoppa used the car industry’s common chassis as his platform analogy. Manufacturers do not design every vehicle from the ground up. They reuse the expensive foundation, then change the suspension, transmission, body and local features needed for a particular market or segment.

He described applying the same logic to Walmart’s international e-commerce estate. Catalogue, search, cart, checkout and payments are common platform components. Regulation, compliance, currency and local branding create necessary variation, but they do not justify rebuilding the entire commerce stack country by country. The same principle applies inside one country when an enterprise has many divisions or product lines.

The operating-model implication is significant. Platform teams own durable capabilities, while product teams assemble those capabilities around customer or business outcomes. McKinsey’s Global Tech Agenda 2026 similarly found that top-performing organisations are adopting product and platform models more deeply, reducing hand-offs and connecting technology delivery more directly with enterprise strategy.

For CIOs, the question is not simply whether the organisation has an AI platform. It is whether teams can reuse identity, access, data, orchestration, observability and governance services without creating a fresh technical estate every time a promising use case appears.

Agent economics will punish vague ownership

Traditional software budgeting gave CIOs a relatively stable unit: a licence per user. Agentic systems disturb that predictability. Enterprises will consume agents embedded in software-as-a-service products, build their own agents, and allow one agent to call another. The cost grows with work performed, tools invoked, tokens consumed and chains of activity that may not be visible in the original user request.

Bidarkoppa warned about vendors offering agent features at no cost for the first year without making the following year’s economics equally clear. His message was not that organisations should avoid them. It was that architecture and commercial discipline need to arrive before widespread use creates an unplanned cost base.

"The only thing that's free in this world is parents' love. There's nothing free."

The practical response is an ownership model that can answer four questions before production: what business outcome the agent owns, which systems and other agents it may call, what a completed unit of work costs, and who can stop or redesign it when consumption or risk moves outside the expected boundary.

Assist, augment or autonomise

Bidarkoppa separated agent use cases into three levels. Assist is the quickest entry point, including service and call-centre work where an agent helps retrieve, summarise or prepare information. Augment keeps a person explicitly in the loop, such as a merchant using an agent to assemble context before a supplier negotiation.

The third level, which he called “autonomise” before joking that he might have invented the word, is where an organisation redesigns the process, roles and technology together. It is also the hardest. Automating one task may save minutes; reconstructing an end-to-end reconciliation, fulfilment or service process can change the unit economics of the function. It can also change jobs, decision rights and accountability, which is why the people question cannot be left until deployment.

This is the boundary between efficiency and transformation. A CIO can add AI to the work that exists, or work with the business to decide what the process should become when software can reason, act and escalate inside defined limits. The latter carries more value and more organisational consequence.

Governance is operating infrastructure

Bidarkoppa’s reference architecture placed an agent layer above enterprise platforms and below the people who use them. That middle layer has to handle agent registration, semantic routing, role-based access and the controls governing how agents find and call one another. Beneath it sit software platforms, custom applications, model choices, cloud infrastructure and operational data.

In that design, governance is not a committee positioned outside delivery. It is part of the runtime environment. Permissions, traceability, intervention points and decision ownership need to travel with the agent as it moves across data and systems.

"Governance should not be a checklist. It's the infrastructure. It's how you actually operate."

The exact balance depends on the industry. A recommendation error in retail may create inconvenience or lost revenue. In healthcare, Bidarkoppa noted, the same design choices can affect patient safety, clinical accuracy and regulatory compliance. Moving quickly does not remove those obligations. It changes how continuously the organisation must enforce them.

The CIO’s role now contains a series of “ands”

Bidarkoppa described the modern CIO role as a set of tensions that cannot be delegated away: technology and strategy, global and local, AI and people, speed and governance, preservation and innovation.

The preservation side remains unforgiving. A CIO cannot credibly lead a board conversation about AI acceleration while core operations are unstable. Customers do not accept outages because the technology team is working on an intelligent future. Resilience, security and service quality remain the permission to innovate.

The people side is equally structural. Bidarkoppa’s talent equation combined AI fluency, deep domain expertise and what he called the DNA of change. A technically capable workforce without operational knowledge will automate the wrong things. Domain experts without AI fluency will struggle to see what can change. Neither group succeeds if the organisation treats change management as communication added near launch.

That is where the session description’s idea of phronesis, or practical wisdom, earned its place. The CIO’s differentiator is not access to a model that competitors can also buy. It is judgement: knowing which process deserves to be rebuilt, where autonomy is safe, which platform component should be shared, when a pilot should stop, and how much change the organisation can absorb without losing trust.

Make the complexity disappear

Bidarkoppa closed by returning to the customer. An airline passenger does not care how crew, aircraft, scheduling and airport systems were assembled into a flight. A shopper does not need to understand the forecasting, distribution centres, transport and store operations that put an item on a shelf. They care that the experience works.

"The best technology in my mind is the technology that makes it invisible."

His challenge to CIOs was therefore practical. Find the zombie pilots. Identify where AI has been placed over an unreformed process. Test whether the data is genuinely ready. Decide which components form the common chassis. Make governance understandable from the executive team to the people doing the work. Put formal change capability beside the technology team. Then demonstrate one core process that has been simplified enough to produce a step change, not just a faster task.

AI may be the current wave, but the leadership promise is older: protect trust, quality and convenience while changing how the organisation delivers them. The enterprises that pull ahead will not be those that make the technology most visible. They will be the ones that make the complexity disappear while the outcomes improve.