The leadership accountability gap

Who owns AI readiness?

As AI features within the everyday workings of the enterprise, the question of who owns it is becoming harder-but more necessary-to answer.

There is a familiar question being asked in boardrooms as companies race to adopt AI: where is it?
That can be quickly followed by others. How many proofs of concept are under way? Why are competitors moving faster? When will the investment start producing a return? The pressure is understandable. Few leadership teams want to discover that a technology with potentially profound implications for productivity and competitiveness has passed them by. But fear of missing out can be a poor basis for capital allocation or business strategy.

For technology leaders, this is creating tension. They are being asked to move faster while simultaneously building and owning the governance, data foundations, and operating models required to deploy AI responsibly. That tension was explored during a recent executive discussion convened as part of the Leading Through Uncertainty series, in partnership with Hitachi Vantara.

From AI experiments to AI workers

As AI moves deeper into organisations, accountability spreads with it. Nowhere is that more apparent than with AI agents.

The first generation of enterprise generative AI largely assisted people. The next iteration from frontier AI organisations, agentic AI, performs tasks on our behalf. Agents access systems, interrogate information, and execute workflows with varying degrees of autonomy that can be set by the business.

This has created a highly unusual management problem: an AI agent is software, but it can increasingly behave like a member of the workforce. Leaders are split on whether these realities are mutually exclusive, we heard.

Organisations already know how to manage access permissions for software and how to assess the performance of employees; agents blur those two disciplines. They require security controls governing what they can access, but they also need to be monitored for whether they are performing the job expected of them.

What’s more, their performance can drift. Outputs can deteriorate, and unlike conventional software its behaviour cannot “always be captured in a fixed functional specification”.

1. This raises questions that stretch well beyond the technology department:
   

2. Who monitors an agent's performance and why?

3. What happens when it consistently underperforms?

4. Are you happy with the measurement for success or performance?

5. Who is accountable for a decision the agent makes?

6. And, most interestingly, how does an organisation prevent the slow attrition of human expertise?
   

The temptation is to regard these as entirely new problems. Some executives take the opposite view. Businesses have long introduced technologies that automate work and alter jobs. Managers remain responsible for the outcomes of their functions regardless of whether the work is performed by a person, conventional software, or an AI system. 

The FOMO problem

These qualifying questions are compounded by the speed at which companies believe they must act.

Leaders complain that activity itself can become a proxy for progress. Elsewhere, a large number of pilots looks reassuring on a board paper even if relatively few eventually create value.

Some organisations are beginning to push against this, we heard. One technology leader described moving away from an approach that encouraged numerous proofs of concept across the business. Instead, the organisation returned to something less fashionable: identifying its most important customer and operational problems first, then determining whether AI was the appropriate technology to solve them.

It is another example of basic management discipline, but one easily lost amid enthusiasm for a new technology or change programme. The inexhaustive list for a successful programme requires leadership, funding, usable data, security, process redesign, and a means of measuring whether anything improved. Global organisations also have to contend with different regulatory, cultural, and operational conditions.

The relevant metric, perhaps, is not how much AI an organisation is deploying. It is whether customers are more satisfied, employees are more productive, costs have fallen, or revenues have increased. These take time-and patience.

It also requires leadership teams to become more comfortable with failure. A blunt consensus from the roundtable: AI projects will fail. Some will work technically but prove to have little business value; others may produce value but at an uneconomic cost; leaders need to distinguish between the two and, crucially, understand what happens next.

A credible AI strategy therefore needs a Plan B alongside its ambitions for Plan A. Organisations should know whether they can return to a previous process, (re)introduce human intervention, or abandon a use case without creating unacceptable operational consequences.

When governance becomes the bottleneck

The instinctive response to uncertainty is often greater central control. With AI, that can create its own problems.

One large organisation represented in the discussion had established central governance and a centre of excellence to oversee AI across multiple businesses. The intention was sensible: create consistency and reduce risk. In practice, a small central team was expected to control activity across a sprawling organisation containing large businesses with very different requirements. Decision-making slowed and experimentation became harder.

The organisation is now moving towards a more federated model.

Common guardrails remain: legal, privacy, security, and data requirements are not optional. But business units have greater freedom to pursue lower-risk use cases within agreed boundaries. More consequential initiatives-particularly those crossing multiple systems or business units-now receive greater central oversight. This may point towards a broader evolution in AI governance for enterprises as many come to the realisation that they have no single level of AI readiness.

One function may have high-quality data, experienced teams, and clearly understood use cases-another may still be struggling with basic data ownership. Imposing the same governance process on both can either expose the organisation to unnecessary risk or slow the more mature business to the pace of the least prepared. The challenge? Governance needs to be proportional without making it optional.

Educating the C-suite

Much discussion about AI skills rightly focuses on employees but it’s important to consider leadership capability, particularly as AI ownership remains a nascent topic itself.

Do senior executives need to become machine-learning engineers? Perhaps not. Do they need enough understanding of AI to ask sensible questions about where it can create value, what risks it introduces, and what investment it requires? Absolutely.

But that can be difficult in organisations where senior leaders have deep expertise in policy, operations, or commercial management but limited technology backgrounds.

Technology professionals across upper management are increasingly finding themselves educating upwards, sometimes through one-to-one sessions built around the executive's own priorities. Rather than explaining large language models in the abstract, the conversation starts with the problems a particular leader is responsible for solving.

Reverse mentoring may have a role, too. Younger employees who use AI tools routinely can help senior colleagues understand what they can and cannot do, lowering the barrier to asking informed questions. There is a leadership dynamic here that organisations should not underestimate: seniority can make admitting ignorance harder, just as the consequences of that ignorance become greater.

In other words, an AI-literate C-suite or board is one capable of distinguishing a meaningful business opportunity from a fashionable experiment through collaboration and discourse with the rest of the business.

The economic dimension of accountability

There is one more important dimension within enterprise AI conversations: token-omics.

The apparently simple act of asking an AI system a question has a cost. As organisations move from occasional employee use towards agents executing repeated workflows at scale, those economics become more significant.

1. Leaders therefore need greater observability over what their AI systems are actually consuming. Questions raised included: 

2. How many tokens does a workflow use?

3. What infrastructure is required to support it?

4. Where does the data sit?

5. Who owns it?

6. Does sovereignty affect where workloads can run?

7. Most importantly, does the value generated by the use case exceed its cost?
   
   

These questions naturally turn AI economics into an accountability issue as the technology introduces consumption patterns that can vary according to the model, workload, and frequency of use. An agent that appears inexpensive during a pilot may look rather different when operating thousands or millions of times across an enterprise with significantly poorer data infrastructure.

This is another reason why ownership cannot reside exclusively with technology. IT may understand infrastructure costs, but the business must understand the value of the outcome those costs are buying.

Closing thoughts: nobody owns AI alone

So who owns AI readiness? The emerging answer is that nobody does at least not alone.

Technology has responsibility for the enabling environment and appropriate guardrails. Data leaders must establish ownership, quality, and trustworthy foundations. Security, legal, and privacy functions define important boundaries. Business leaders remain accountable for the outcomes AI is being asked to deliver, whether they understand those decisions or not.

Businesses seem to be choosing between distributed accountability or ambiguous accountability. The former gives different leaders explicit responsibilities within a common operating model; the latter allows everyone to assume somebody else is in charge. AI will eventually force organisations to confront that difference whilst increasing the number of functions that coordinate together.

AI is now an operational capability which has relegated speed a red herring metric for the executive measuring AI success. Far more important is where to impose control and, following accepted leadership practice, where to say no.

In partnership with