Clare Kitching

AI risk isn't where you think it is.

It's not in the boardroom.
It's not in the policy document.

It's in the everyday decisions no one sees.

A prompt typed into a tool.
A file uploaded with best intentions but without thinking it through.
A shortcut taken to save time.
That's where things go wrong.

Most organisations focus on governance first.
And that makes sense.

But governance alone doesn't change behaviour.

What works is building guardrails into how work gets done.

Four layers matter:
1/ Governance
Who owns the risk and what is acceptable…

Growth, risk, efficiency. All hinge on data governance.
Here’s how to explain it to execs:

I worked with a company where the data team spent months trying to convince executives to assign data owners.
They failed. Why?
Because they talked about process, not outcomes.

Here’s what leaders need to hear instead.

Good data governance will give you:
✅ Better decisions: Reliable data means you can act with confidence.
✅ Risk control: Avoid fines, breaches, and reputational harm.
✅ Efficiency: Free your teams from fixing issues.
✅ Value creation: Accelerate ROI on your AI investments.

BCG says 70% of the value in an AI transformation comes from people moves, not the technology.

That should change where you spend your attention. Most leadership teams are deep in tool selection and model choice, but the return is elsewhere.

BCG lays out seven principles to consider:

1/ Agree on a few workflows, not "AI everywhere." Successful companies prioritise three or four use cases. Less successful ones spread across six or seven.

2/ Give managers a real say in redesign. Their roles change most, so let them shape how the work gets done. They run it better from day one.

We worry about AI taking our jobs.
Perhaps we need to worry about it taking our friends.

The latest research published in HBR shows therapy and companionship is once again the most common AI use case, and its share has more than doubled from 5% to 11% in a single year.
This is the 3rd year that AI in the Wild have looked at how humans are evolving alongside AI, with 12,000+ use cases analysed.

For years we've talked about AI as a productivity tool. It’s faster for writing, code, analysing and researching.

Agentic AI has moved from hype to hard questions.

As companies look to implement, they’re asking:

Where does it create real advantage?
How do we stop our company looking like every other company?
What happens to roles, workflows and accountability?
And what foundations do we need before we scale?

I pulled together 10 recent reports that helped me make sense of where agentic AI is heading next.
It starts with value and advantage, then moves into risk and leadership, then finishes with operating model and scale.

Most teams pour budget into dashboards and warehouses, then wonder why decisions still crawl.

It's a reminder that data doesn’t create action,
people do.

The highest leverage in any AI or data program isn’t collection or tooling.

It’s translation. The humans who turn rows into reasoning.

Here’s what I see strong leaders doing differently:
1️⃣ Funding insights and change, not just infrastructure

2️⃣ Asking “What will we change?” instead of “What does the dashboard say?”

3️⃣ Measuring the decisions made, the value implemented and not just the data gathered.

Before AI can reason with your data,
your business needs to agree on what that data means and
what you want to achieve.

The data isn't the hard part but understanding each other is.

I've been in data for over 15 years, and sometimes even I feel like I'm decoding a foreign language.
We've turned simple ideas into jargon that makes non-data people tune out.

Here's what these terms actually mean and why they matter for AI:

▶️ Ontology
A shared definition of your core business concepts and how they relate.
It gives AI clear concepts to reason about instead of guessing.

AI works best when it’s used like a knife, not a miracle.

AI is starting to look a lot like digital transformation all over again.
Big promises, huge investment, hundreds of pilots. Very little impact.

And I think the problem is surprisingly simple:
Most companies misunderstand what AI actually is.

They treat it like magic when it behaves more like a tool.

I see three mindsets:
🪄 The magic wand
“AI will solve it.”

This is the hope driven approach.
Buy the platform, hire consultants, announce the transformation.

Expect productivity to appear.

What catches most companies off guard with AI?
The speed with which small costs become massive ones.

There’s been a lot of talk about the surprisingly high cost of AI.
It seems many teams skipped the business case.

AI isn’t cheap, but the reality is nuanced.

Some AI tasks cost cents.
Some cost dollars.
Some run thousands of times a day.
Others only a few times a week.
Some save hours of work.
Some are not worth automating at all.

And the cost isn’t just driven by usage volume.

The most misleading AI metric might be adoption.

Lots of people using AI doesn’t automatically mean your business is improving.

“We rolled out AI to 5,000 employees.” Great. What changed?

This is the question many organisations still struggle to answer.

Too often, AI measurement starts and ends with adoption metrics:
Number of users, prompts per day, licences activated.

Those metrics matter to start, but you need to keep an eye on bigger outcomes.
AI isn’t just a productivity tool.
It can change customer experience, revenue generation, operational speed and employee experience.

Two ways to get AI wrong:
chase every shiny object, or play it so safe you sleepwalk into irrelevance.

Some organisations pile into agents and GenAI and hope it changes everything.

Others stay in their comfort zone.
Rules, reporting and incremental improvements.

Both create hidden risk.

If you only chase big bets, results get fragile fast.
If you only back safe bets, you slowly fall behind.

The AI strategies that hold up over time look more like portfolios.

You have dependable layers:
▶️ Clear rules that protect compliance.

Most companies still think AI is a technology problem.
It’s not.

PwC’s 2026 predictions remind us once again:
Only ~20% of value with AI comes from the tech.
The other 80% comes from how work is redesigned.

Yet most teams do the opposite.
I’ve seen teams with incredible tools still stuck with no measurable impact.
And others with basic tools driving real gains.

The difference?
They changed how the work gets done.

Buying feels like progress.
Pilots feel like progress.
But months later… nothing actually changes.

Redesigning work? That’s slower, messier and more uncomfortable.

AI isn’t a single technology you adopt.
It’s a set of building blocks you choose from.

AI feels sudden.
But what looks like an overnight breakthrough is really decades of layers quietly stacking up.

Thanks to my friend Luís Rodrigues for the image inspiration, it captures that story better than most explanations I’ve seen.

Here’s how we got here…

⏭️ At the bottom: Rules and logic...sometimes rebranded as Classical AI
Rigid rules. No learning.
Think eligibility checks in insurance or compliance rules in banking.
Useful, but fragile.

AI risk isn’t where you think it is.

It’s not in the boardroom.
It’s not in the policy document.

It’s in the everyday decisions no one sees.

A prompt typed into a tool.
A file uploaded with best intentions but without thinking it through.
A shortcut taken to save time.
That’s where things go wrong.

Most organisations focus on governance first.
And that makes sense.

But governance alone doesn’t change behaviour.

What works is building guardrails into how work gets done.

Four layers matter:
1/ Governance
Who owns the risk and what is acceptable…

Data isn't the hard part.
Understanding each other is.

Ontology. Lineage. Semantic layers. Vector databases.

I've been in data for over 15 years, and sometimes even I feel like I'm decoding a foreign language.
We've turned simple ideas into jargon that makes non-data people tune out.

Here's what these terms actually mean and why they matter for AI:

▶️ Ontology
A shared definition of your core business concepts and how they relate.
It gives AI clear concepts to reason about instead of guessing.

▶️ Entity
A real world thing like a customer, product or event.

The most dangerous phase of an AI program is right after the win.

The demo worked.
Stakeholders are excited.
Everyone wants to move on to the next shiny thing.

That’s when problems start if the long term hasn’t been factored in.

AI solutions succeed when they're safe & reliable.
That means:
→ Clear decisions on what it should and should not do
→ Humans who stay accountable
→ Time (& budget) set aside for fixes, tuning, and judgement calls
→ Acceptance that edge cases happen

AI requires an ongoing operating model, not a one off investment.

Everyone wants AI magic.
Few want to invest in the plumbing.

We live in a moment where the promise of AI
is louder than the reality of data.

I see it every week with teams.

The excitement is real. The budgets are flowing. The pilots look impressive.
But when you lift the lid a different story shows up.

Data that lives in twelve places.
Metrics that mean one thing in finance and another in operations.
Ownership that sounds like “we think it sits with them”.
Governance that feels optional.

And still we keep sprinting toward AI, hoping it will smooth over the cracks.
It never does.

Everyone talks about AI models.
Very few talk about AI systems.

When you look under the hood of most AI today, you rarely find just a large language model.

You find layers.
Context. Memory. Retrieval. Tools. Autonomy.

This diagram shows the progression.

▶️ Large Language Models (LLMs) are the foundation.
They are massive neural networks trained to understand and generate human language.
They take an input and produce a response based on patterns learned during training.
Powerful, but limited to what they already know.

The fastest way to understand AI governance?
Stop thinking about data. Start thinking about decisions.

There's a lot of noise around governance in the age of AI.
And leaders cut through some of that noise by understanding the split between inputs and outcomes.

Data governance protects the inputs.
It’s about accuracy, privacy, lineage, access and compliance.
It answers: Can we trust the data going into our systems?

AI governance protects the outcomes.
It deals with fairness, explainability, drift and human oversight.
It answers: Can we trust the decisions made on outputs of AI systems?