AI Isn't Failing. Our Operating Models Are.

By Ari Magalhaes · 2 June 2026

Why boards need a different conversation about AI investment, and what to ask instead.

When personal computers entered organisations in the 1980s, economists tracked an unsettling phenomenon. Billions were being spent on computing infrastructure. Productivity statistics barely moved. The Nobel laureate [Robert Solow](https://en.wikipedia.org/wiki/Robert_Solow) captured the paradox in 1987: "You can see the computer age everywhere but in the productivity statistics."

The resolution came a decade later, once companies stopped automating existing workflows and started redesigning them. Once data became a shared enterprise resource rather than departmental property. Once leadership accountability structures caught up with what the technology actually required.

We are in the same paradox now, with AI. The Solow problem has not gone away, it's back with a vengeance.

The returns question is here

Everywhere I look, boards and executives are asking the same thing: we have poured money into *[insert new technology here]*, so where are the returns?

As an economist and transformation specialist, I am experiencing déjà vu. This is exactly the pattern I have seen for years in large-scale change programs. The problem is rarely the technology. It is leadership, incentives and the way we design work.

Three independent research reports, from BCG, McKinsey and MIT, have converged on a figure that should reframe every board conversation about AI:

> Fewer than 1 in 10 companies are capturing meaningful AI value at scale.

[McKinsey](https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai) finds that while most large organisations are investing, only a minority can tie those efforts to measurable earnings gains, and even then the impact tends to be incremental. [Gartner's October 2025](https://www.gartner.com/en/newsroom/press-releases) survey lands a sharper number: 88% of HR leaders say their organisations have not realised significant business value from AI tools, largely because they have not rethought how work itself should change.

This is not a model-quality problem. It is a governance and operating model problem.

The translation gap, and the messy middle

Jeremy Korst, who co-authors the [Wharton/GBK Enterprise AI Adoption Study](https://ai.wharton.upenn.edu/) with researchers from the Wharton School, frames the diagnostic precisely: "The gap isn't primarily a technology problem. It's a translation problem. Investment is flowing faster than the organisational infrastructure required to absorb it."

Their research makes the internal fracture visible. 74% of business leaders report perceived positive returns from early AI deployments. 45% of senior executives say those returns are significant. Among the middle managers actually running these programs, only 27% agree. The researchers call this [the "messy middle"](https://ai.wharton.upenn.edu/).

How I would say it: the strategy looks great in the board pack, the technology demos well, but the real work gets stuck with managers who are already overloaded and rarely have the bandwidth, incentives or support to redesign how their teams actually operate. I see this in every non-AI transformation I have worked on, capital is deployed faster than the organisation can absorb it productively. The management information reaching the board does not reflect what is actually happening on the ground. AI is the latest, more expensive, version of that story.

AI is exposing every seam in the C-suite

IBM's research provides the structural explanation. Beneath near-universal enthusiasm for AI at the top of any large organisation, five executive agendas are running in parallel, and rarely reconciled into a single design for value creation.

The CEO wants growth and reinvention. The CFO wants measurable ROI and capital discipline. The CHRO wants workforce readiness and cultural stability. The CIO and CTO want architectural integrity and security. The COO wants operational continuity. These tensions exist in every organisation. What AI does is expose them at scale and price them.

> "It's exposing every pre-existing seam in the C-suite. If the CFO and the CHRO weren't aligned before AI, they are definitely not aligned through the AI transformation.", Manish Goyal, IBM

This is exactly what I see in turnaround work. When cross-functional challenges are treated as everyone's problem, they end up becoming no one's job. Material AI investment without a coherent design that reconciles these five agendas, with explicit trade-offs, named accountabilities and board-level oversight, is not a management challenge, but rather a governance gap.

Tools without workflow redesign: random acts of proof

Most organisations are layering tools onto existing processes and calling it transformation. Copilots get deployed, but productivity barely moves. Why is that? IBM's Caroline Roche puts it plainly:

> "You cannot just give your employee an AI and expect to get huge productivity, you must fundamentally redesign the workflow."

A workflow, in the technical sense, is an end-to-end flow of work that crosses departmental boundaries, source-to-pay, record-to-report, clinical assessment, client onboarding. A process lives within a single function. Most AI deployments have targeted processes: a writing assistant for marketing, a forecasting tool for finance, a CV screening tool for HR. These are productivity improvements at the margin. They are not the redesign of the cross-functional flows where AI can change the underlying economics of an operation.

IBM's own internal program, Client Zero, illustrates what redesign would actually require. They inventoried 400 workflows, selected those with the greatest operational use, stripped out accumulated work that no longer served a purpose, redesigned each end-to-end, and only then applied AI. The sequence is the lesson: workflow redesign before AI deployment, not after.

Goyal names the failure mode at the other end of the spectrum: "pilot proliferation" and "random acts of proof", lots of short-term wins, no enterprise-level impact. His remedy, which I share from both a consulting and board perspective, is P&L ownership. Each AI initiative needs a named business owner with a number they will be measured against. Without that, it remains an experiment at the edges of the organisation.

Data: the hidden constraint boards are underpricing

There is a further layer that boards rarely examine closely enough. AI requires more than clean reporting dashboards, it requires integrated, real-time information organised around what engineers increasingly call an ontology: a structured representation of how the enterprise actually fits together.

Roche offers a concrete banking example. The same customer might hold a childhood savings account under one name, a credit card opened at eighteen under a slightly different name, and a wealth management relationship opened decades later under a married name. Risk models and fraud detection are only effective if those three records are recognised as the same person, contextualised across the full relationship. Without that reconciliation, the AI is operating on fundamental ambiguity.

Her observation applies as directly to aged care and disability services, where a person's funding, clinical and personal history might sit across three different systems, as it does to banking. Roche's conclusion:

> "The organisations that view data as an IT problem are really the losers in the AI race. The organisations that view data as a critical business asset are the ones that will see the most value."

From a governance lens, this means boards need to ask whether data quality is being managed as a strategic performance metric rather than a technical one.

Five questions boards should now be asking

IBM's Enterprise 2030 research shows that speed has entered the top three C-suite concerns for the first time. Goyal's observation cuts to the core: "Everybody has the same exact technology. There is no proprietary technology that I have that you don't have access to. Ability to harness this at speed is what is going to differentiate."

Speed, in this context, does not mean more pilots. It means the organisational capacity to align the C-suite around a coherent value thesis, prioritise the right workflows, build data infrastructure that compounds over time, and genuinely support the middle of the organisation, not mandate it.

I am increasingly less interested in how many AI tools an organisation has deployed and far more interested in how it is governing workflows, data and incentives. Five questions I now consider essential in any board discussion about AI:

1. Who has explicit P&L accountability for AI outcomes? A business-side owner with a defined target. If the answer is a committee or a collective mandate, accountability has dissolved. 2. Which end-to-end workflows have been mapped, costed and prioritised? Not departmental processes, the cross-functional flows where the economics of the business actually live. 3. Is data governed as a board-level strategic asset? With visible quality metrics and executive ownership of the definitions, not as an IT cost line. 4. How are we managing the "messy middle"? Are managers being given bandwidth, skills and incentives to change how work gets done, or are they being handed another mandate on top of an already full workload? 5. What is the board actually seeing? Are AI programs reported as discrete spend items, or as a value story linked to specific revenue, productivity and risk outcomes? The reporting framework determines what the board can govern.

The amplifier

AI is not a magic layer we can apply on top of existing complexity. It is an amplifier. It amplifies broken incentives, fragmented leadership and poorly designed workflows just as readily as it amplifies good strategy and disciplined execution.

The Solow paradox resolved when organisations redesigned the work itself. The AI paradox will resolve the same way, for the organisations that treat it as a governance challenge, not a technology one.

AI isn't failing. Our operating models are. That is exactly where boards and executives need to focus their attention.

> Ready to pressure-test how your board is governing AI investment? [Request a confidential briefing](/#contact) or take our [AI Readiness Diagnostic](/tools/ai-readiness) to see where the gaps are.