AI Leadership in Practice: Strategy, Risk, and Reward

Most organisations are not behind on AI. They just think they are.

That feeling, the creeping anxiety that Meta and Amazon are lapping you, that you're somehow failing by not deploying cutting-edge models at scale, is one of the most persistent and damaging myths in enterprise AI right now. In a recent session hosted by Data & AI Literacy Academy, Greg Freeman (CEO and Founder) sat down with Stephanie Gradwell and Jessica Bell, Directors at Pendle, an AI governance consultancy, to cut through it.

What followed was a grounded, practically useful conversation about what AI leadership actually looks like, and what gets in the way of it.

You don't need to be at the cutting edge, but you do need to understand your business

The latest research suggests that only around 4% of organisations have developed the kind of horizontal, business-model-reshaping AI capabilities that dominate the headlines. The instinct for many leaders is to treat that statistic as a gap to close. Jessica pushed back on that directly.

"99.99% of organisations don't need cutting edge technology," she said. "What I think causes decision paralysis is people thinking: if I can't be the best at this, I'll wait. But the value lies in a deep understanding of your business, which no cutting edge technology can actually solve for you."

Once you have that understanding, you can make educated decisions about which technologies actually matter for your context. Some of those will be  nascent while others will be standardised automations that have been available for a decade.

The point is: always start with the problem, not the technology.

Technology is the easy part

This is the claim that tends to raise eyebrows, especially with technical data leaders. But it's one Greg has arrived at through years of working inside large enterprise programmes, and it's increasingly backed up by research.

Mid-2000s studies on data projects found that 55% of failures came down to people problems. The early evidence from AI suggests that number is even higher, because in many cases, rolling out the technology itself is now genuinely quite straightforward. You can give everyone Copilot tomorrow. The harder question is whether they'll use it in a way that changes anything.

"The stuff that's really going to determine whether this is successful, and whether it's done in a risk-managed way, is the culture, the behaviours and the actions that individuals take ten thousand miles away that you can't do anything about," Greg noted.

Jess added a dimension to this that leaders often underestimate: AI causes an emotional reaction in people. It touches the world they've always known. If you want adoption, you need to deeply understand what people in your organisation are fearful of, what they value about their current roles, and make sure that as you introduce transformation, you're solving problems for the people who are ultimately going to be responsible for making it work.

The technology change is exponential. What this increasingly demands is a psychology of teams.

What strong AI leadership looks like

It's probably not what you expect. The strongest AI leaders aren't the most technically fluent people in the room. But they can set a clear operating model, communicate a coherent strategy, and crucially build teams and structures that can adapt as the landscape shifts under them.

Steph made a great point: the operating model you design today won't be right in 18 months. So the question isn't just what the model looks like now but how you design it so that any tool, any architecture, is interchangeable. Organisations that become significantly wedded to a specific platform they can't easily extract themselves from are storing up a real business risk.

And then there's the question of what AI is actually for.

Just embedding AI into existing workflows is not strong AI leadership. Tools like Claude or Copilot embedded into everyday tasks are useful, in the way Excel is useful, but they're a productivity layer, not a transformation. The organisations that get real value are the ones that step back from an existing process entirely and ask: if we were designing this from scratch with AI, what would it look like?

Here's one compelling example. One of Pendle's client used to take three weeks to produce a complex technical report for HMRC submission. By re-engineering that process end-to-end, and keeping a human in the loop, but using AI throughout, they got it down to an hour and a half. It had the same output and quality at a fraction of the time.

The commercial logic is obvious. They're charging the same amount for a piece of work that now takes a fraction of the resource, delivering immediate margin gains to their business.

The Governance question everyone avoids until it's too late

AI governance sounds like the part of the conversation where everyone dozes off. It shouldn't. Getting it right or wrong has consequences that can compound overtime.

Pendle's approach distils the major global frameworks into core principles, each of which needs to be understood in the context of your organisation rather than as a generic checklist.

This means being able to explain, in plain language, what an AI system does, what data it uses, what risks it carries, and who owns it, from board level to the person clicking accept. If someone can't summarise this clearly, that's a signal the understanding isn't there.

• ‍Fairness means recognising that large language models are trained on historical data that carries historical biases. This is a technical reality a lot of businesses don't want to engage with but it's crucial work. If you're using AI in recruitment, performance management, or any decision that affects people, you need bias testing built into the process. And you need to understand that an AI system doesn't just inherit your bias, it perpetuates it at scale, indefinitely.‍
• Accountability is the one that trips organisations up most often. "Everyone is accountable for AI" is, in practice, a way of ensuring nobody is. Accountability needs to be structural and named, specific to each stage of the process. As Greg put it: whose name goes on the lawsuit? That level of seriousness is exactly what should accompany these decisions.‍
• Safety and Privacy are less abstract than they sound. Most organisations have employees using AI tools in ways they haven't fully considered. The obvious risk (like don't put sensitive data into a public chatbot) is generally understood (albeit bares repeating on an ongoing basis!). The less obvious risk is how the AI tools your organisation uses are architecturally connected to your other systems. There are documented cases of AI chatbots being prompt-injected to expose data from connected systems. Understanding your Data Processing Agreements with third-party AI providers is not optional, it's essential due diligence.‍
• Human Agency is straightforward in principle: AI doesn't replace human judgement. The human accountable for an output can override it. That principle needs to be designed into processes, not assumed.‍
• Sustainability matters because token use and data centre energy consumption are a significant cost, both environmentally and financially. How you build and communicate about your solutions matters, not just whether it works.

These principles exist on a lifecycle. Governance isn't a document you write and file. You have to continuously revisit it at every stage, from conception, development, deployment, to monitoring, and retirement.

A four-question framework for any AI use case

The problem with governance in practice is that it can feel abstract until you need it. Pendle's shared their practical four-part framework that operationalises these principles for any individual AI use case.

1. What problem does this actually solve?

Too many AI projects begin with pressure from above ("we need to be doing AI") or anxiety about competitors, rather than a clearly articulated problem that everyone across the organisation agrees is worth solving.

2. What are the dependencies?

This determines feasibility. If unlocking the value of a use case requires capabilities your organisation doesn't currently have, like data quality, skills, an AI policy or a defined culture, you either have a genuine blocker or a clear roadmap for what needs to be addressed first. The importance of the problem should drive how urgently you address the dependencies.

3. How do you monitor it end to end?

Once deployed, how do you know it's doing what it was supposed to do? How do you assess whether it's staying true to the business case? This is where a lot of organisations struggle.

4. Who is accountable, and at what stage?

Not one person for everything. You need a selection of specific named individuals, accountable for specific stages, with the tools and authority to actually maintain that accountability and change things if needed.

Pendle has built this framework into a publicly available prioritisation tool at pndl.ai, with common use cases pre-loaded across industries. It's a good starting point for thinking through where AI genuinely creates value in your context, and for having the governance conversation before you need to have it under pressure.

On AI Literacy: It's not about the buttons

There's a version of AI literacy that's only focuses on "button training:. Here's how to use Copilot, here's how to write a prompt, here's where the settings are, and so on.

That's now become table stakes and isn't where the real work is.

The technology moves too fast to train people on specific tools and expect that investment to last. What you can train people on, and what actually determines whether your AI transformation sticks, is the mindset and behaviours that make someone willing to pick up a new tool, adopt a new process, and think differently about how they work.

And critically: data literacy is and will always remain the foundation. People cannot effectively use AI and cannot evaluate its outputs, understand its limitations, or identify when something looks wrong, if they don't understand why data quality and data governance matter. These are not separate capability streams but belong in the same conversation.

‍

Data & AI Literacy Academy partners with enterprise organisations to build the capability, culture, and governance that makes AI transformation work. To find out more, visit dl-academy.com.

Pendle's AI governance prioritisation framework is publicly available at pndl.ai.