Key Takeaways:
- Most organizations are still between piloting and selective production on AI — and the blocker is almost never the model. It’s the data underneath it.
- Agentic AI changes the stakes: when AI acts instead of recommends, bad data no longer produces a bad report — it produces a compounding chain of bad decisions at machine speed.
- The organizations moving fastest from pilot to production involve legal, compliance, and HR at design time, and name one person accountable for the full agent lifecycle.
I recently joined twenty of the UK’s senior data, analytics, and AI leaders gathered at the Institute of Directors in London for a conversation I’ve been wanting to have for a while. No sales presentations or vendor roadmaps, just a peer-led roundtable about the thing that’s keeping everyone up at night: you’ve proven your AI pilot works. Now what?
I’ve run enough of these sessions to know that the first thirty minutes are always polite. People give the safe answers. They talk about pilots and promising results and governance frameworks “in progress.” Then something shifts, someone tells the truth, and the room gets interesting. This one shifted faster than most.
Here’s what I took away.
The honest answer to “Are you in production?” was illuminating
I asked the room a simple question to open: “How many of you have an AI use case today that you would genuinely describe as production – not a pilot, not a proof of concept, but something the business depends on?”
The pause before the hands went up told the whole story.
Most organizations in that room were somewhere between piloting and selective production. A few were scaling. Almost nobody claimed to be truly business-critical. And when I pushed on what was standing between where they are and where they need to be, the answer was almost never the model. It was the data underneath it.
The math changes when AI stops recommending and starts acting
I’ve been using a GPS analogy for years. For as long as I can remember, I’ve told data leaders: I’d rather have no GPS than one with bad information – at least then I know I’m lost. Bad data in a GPS only gets me to the wrong place faster.
But that analogy always assumed something important: you’re still the one driving. You can question the instruction. You can pull over. You can recalculate.
Agentic AI changed the analogy. Now the car turns the wheel itself. When it takes the wrong exit, you’re already on the motorway … and the car is still accelerating.For years, AI recommended. Now it acts. Agents query your systems, write to your records, and trigger
workflows – in chains, at machine speed, without a human reviewing each step. Bad data plus a recommendation is a bad report. Bad data plus an autonomous agent is a compounding chain of bad decisions at scale.
Every person in that room felt the weight of that. Several of them were already living it.
The semantic layer gap is bigger than most people will admit
This was the conversation I didn’t want to end. Someone in the room described a moment that I’ve heard variations of more times than I can count: their agent was returning results that looked technically correct but were contextually wrong.
It turned out their organization had multiple definitions of “active customer” embedded across different systems. Humans knew this and navigated around it. The agent didn’t. It picked one definition and ran, and the output made no sense to half the business.
The fix isn’t more data. It isn’t a better model. It’s a semantic layer: governed definitions, business context, and AI-readable metadata that lets an agent act on the right version of the truth, not the most accessible version of it.
The candid admission from several leaders in the room was that they have glossaries and wikis and data dictionaries. What they don’t have is something an agent can actually use. That gap is widening fast, and it’s one of the most underinvested areas in enterprise AI readiness today. Even the most data-mature organisations in that room hadn’t built a semantic layer that could properly scale for agentic AI.
Legal, compliance, ethics, and HR belong at the table – earlier than you think
This topic generated real energy. One of the most common failure patterns when moving an AI pilot to production isn’t technical. It’s that the pilot was designed, built, and validated without the people who will ultimately determine whether it can actually go live.
I’ve seen this at major enterprises. You run a successful proof of concept, the data team is confident, the model performs well. And then legal looks at the training data, compliance asks about the decision trail, and someone in HR raises questions about how the model affects people in the process. Suddenly a six-week timeline becomes six months.
The organizations that move fastest from pilot to production are the ones that involve legal, compliance, ethics, and HR at the beginning – not as a final gate, but as active participants in the design. The conversation in the room confirmed this isn’t theoretical. It’s the difference between a pilot that scales and one that stalls.
Not every problem needs an AI agent, and the room knew it
Some of the most useful governance conversations I have aren’t about how to govern AI, but whether to use AI at all for a particular problem.
One of the sharpest observations from the afternoon in London: a lot of organizations are deploying AI on problems that are fundamentally deterministic. Problems that could be solved reliably, cheaply, and auditably with rules-based logic, traditional analytics, or simple automation. The agent adds cost, complexity, explainability challenges, and governance overhead, and it doesn’t add value.
Strong governance frameworks don’t just evaluate whether an AI is safe to deploy. They evaluate whether AI is the right tool for the job. The room had clearly started grappling with this. There’s real risk in doing AI for AI’s sake: chasing the technology rather than the outcome.
Nobody owns it end-to-end. Almost nobody.
I asked who actually owns AI governance once it goes to production: CDO, CTO, Legal, Risk, or nobody? The answer from most of the room was some version of: contested, which effectively means nobody. Engineering is deploying. Legal is worried but reactive. The CDO wants to own the data layer but doesn’t have authority over deployment decisions.
The organisations making the most progress had done one specific thing: they had named someone accountable for the full lifecycle of an agent, from the data it touches, to the decisions it makes, to what happens when something goes wrong. Not accountable for a piece of it. All of it. That clarity alone was making a measurable difference.
What happens next
I left the Institute of Directors more convinced than ever that the UK data leadership community is ready for this conversation – and hungry for more of them. The quality of thinking in that room, the willingness to admit the gaps, the genuine peer exchange – this was exactly what I hoped it would be.
We’ll be running more of these sessions across the UK and Europe. If you’re a senior data, analytics, or AI leader navigating this transition – from promising pilot to trusted, business-critical AI – I’d like you in the room next time.
The problems are real. The solutions aren’t as far away as they feel. And the conversation is better when you’re in it.
Interested in joining a future session? Get in touch.
