Governance is your next competitive advantage: What we learned at AI Uncorked

A recap of our virtual fireside chat, "AI Uncorked: Guardrails, Compliance & IP Protection for the Enterprise," held July 23, 2026.

When we polled the audience at the start of AI Uncorked, the results told a familiar story: 32% of organizations are just getting started with AI, another 32% are piloting in select teams, and only 21% say AI is established in how they work. Most enterprises, in other words, are somewhere in the messy middle — past curiosity, short of confidence. And the thing standing between the two, the panel argued, isn't the technology. It's governance.

That was the thread running through the entire conversation. Hosted by John Lees-Miller, co-founder of Overleaf, the fireside chat brought together Dr. Mark Bloomfield, Fellow at Cambridge Judge Business School, who spends most of his week advising boards on AI transformation, and Maximilian Freier ★, economist and Secretary of the Monetary Policy Committee at the European Central Bank, where he also coordinates the ECB's own AI rollout.

Here's what stood out from an hour of genuinely unscripted discussion.

Two camps, one honest answer: Nobody knows where AI is taking us

Max opened with the economist's view: The profession has split into an enthusiast camp, which sees AI as a steam-engine-scale transformation that will carry humanity into a new age, and a realist camp that compares it to the internet, the PC, or electricity — significant, but familiar in shape.

His own position? Both are right, depending on two variables: Whether technological capabilities keep compounding, and how fast diffusion actually happens. In the near term, he expects productivity effects rather than upheaval. But as adoption deepens and capabilities grow, "I do see scope for really transformational effects to our societies," which is exactly why policymakers and researchers need to be investigating the implications now.

Mark framed the uncertainty differently, around two axes: An organization's ambition — whether it uses AI merely to become more efficient or to genuinely rewire what it does — and trust, which he argued won't simply arrive but must be built through understanding how these tools work and where their limits lie. Plot those two axes and you get four possible futures. The organizations that combine high ambition with deliberately built, high-trust infrastructure are the ones likely to succeed.

Everyone says they've adopted AI. Only 7% actually have.

"Almost all firms say they've adopted AI. But that probably means they've bought a couple of licenses." – Maximilian Freier

The panel found the opening poll refreshingly honest. Max noted that in surveys of Euro-area firms, nearly everyone claims to have "adopted AI", but dig deeper and only around 7% report significant adoption.¹ Buying a few licenses isn't transformation, which is why so many CFOs are underwhelmed by the returns: "Without deep adoption, we don't see productivity gains."

John added a view from the technology side of the business, where the transformation has been faster and deeper than anywhere else: "The way that we build software today really looks nothing like it did 18 months ago." A useful reminder that adoption curves vary widely not just between organizations, but within them.

Mark shared a candid admission from a large California client: "I genuinely thought once we rolled out Copilot, we would all be transformed." He called this technological solutionism — the seductive belief that you can plug in the tech and the organization transforms in parallel.

In his experience, organizations move from curiosity to confidence, i.e., the point where they give themselves permission to redesign workflows or stop doing things entirely. And that transition is where the real friction lives: AI policy, governance, data quality, and internal dissent.

Rethinking the business case: Be problem-first, not AI-first

When John asked whether ROI is even the right frame for AI investment, Mark identified a structural lag: Give someone a tool that makes them 20% more productive, and nothing guarantees the organization coordinates that freed capacity into net gains. His bigger warning was about narrative traps and chasing "AI-first" or "agentic-first" labels instead of being problem-first, starting from clearly defined organizational problems, and letting AI be part of the solution.

Max agreed the value has two dimensions — expanding the analytical toolkit and classic efficiency — and highlighted use cases where returns can be quick to arrive. His flagship example: The ECB's Corporate Telephone Survey tool, which automates transcription, summarization, and sentiment extraction for a quarterly survey of firms across Europe. Within three quarters, he could demonstrate the FTE savings precisely, achieving break-even in months, not years.

Mark matched it with a case from a New York investment bank suffering 33% attrition. Exit interviews showed people were leaving for lack of coaching and development, while analysis revealed 92% of the HR team's time was consumed by transactional questions. A modest AI agent built on the employee handbooks freed that capacity, and within 18 months attrition fell to 4%.

The takeaway for your own business case: Start from the problem, not the technology. You'll know what to measure, and you'll know when it's working.

Change management with no point B

Why does AI transformation feel harder than previous technology rollouts? Mark's answer: Traditional change management moves an organization from A to B, but with AI "there probably is no point B" because the capability shifts too fast. He described redesigning an entire Cambridge executive program mid-week because of what shipped between the Tuesday and Thursday sessions. The goal isn't reaching a fixed end state, but building the muscle of continuous adaptation.

"Traditional change management goes from an A to a B. With AI, there probably is no point B — the capability is shifting too fast." – Mark Bloomfield

Max's favorite change-management story made the human side vivid: A colleague of 15 years' standing who insisted no machine could summarize a text faster than him and who, months after being pushed through a resisted rollout, spontaneously talked a different skeptical team through their doubts and highlighted his positive experience.

The lesson: Take people along and don't dismiss shallow adoption. Sometimes fixing the spelling mistakes is the gateway that builds trust for deeper change.

Both panelists warned against the "wait and see" posture. The capability gap a competitor builds while you wait, Mark noted, may become impossible to close. And the risks of mishandling the people side are real: He described one organization where employees, incentivized to document their skills so agents could use them, deliberately fed false information into the system as a form of protest.

Governance as an enabler — And shadow AI as the price of getting it wrong

"Governance could be our next source of competitive advantage." – Mark Bloomfield

The heart of the discussion challenged the reflex that governance equals "no." Mark is currently working with three clients whose explicit position is exactly that: Governance as their next source of competitive advantage. He distinguishes between responsible AI, i.e., what an organization believes — fair, accessible, inclusive AI, and governance, i.e., the operation of those beliefs: policies, actions, accountability.

Done well, governance answers the questions that actually unlock adoption: Where can people experiment, what needs risk review, and what's the recourse when the answer is no. Get it wrong in either direction and you pay. Default to "governance equals no", which is common in data-rich, regulated firms, and adoption stalls while shadow AI flourishes and talent leaves. Lift all the barriers, and with today's agentic capabilities the reputational risk is immense.

Shadow AI came up repeatedly, and the examples were sobering. Mark described a trading floor where someone built a personal agent that works from photos of the trading screen — mixing business and personal data in an entirely unmanaged tool. Max's observation was blunter still: talk to staff at any institution that doesn't offer the latest AI tools, and "every single person I talk to uses shadow AI of some sort." If sending a document to a private account saves someone two hours and gets them home to their children earlier, they will do it. As Mark put it: water always finds a way.

"Every single person I talk to uses shadow AI of some sort" – Maximilian Freier

If your teams aren't using the tools you gave them, it's worth asking what they're using instead.

The second audience poll underlined the point. Asked what would most help them move forward with AI adoption, 64% chose better tooling with built-in compliance controls — far ahead of peer examples (43%), executive sponsorship (36%), policy frameworks (21%), or external guidance (21%). People don't want permission slips; they want tools they're allowed to use.

What you can do on Monday morning

The session closed at the individual level, with each panelist offering three takeaways.

From Max:

1. Push on transparency and explainability. Opening the black box is the research frontier, and the answer to policymakers who distrust ML-based analysis.
2. Use AI openly and without fear, but stay accountable: He'd have no problem with a paper drafted by AI, as long as the researcher fully understands, owns, and can defend every result.
3. Protect your skills. It's dangerously easy to let the tool's capabilities quietly replace your own.

From Mark:

1. Always ask how the sausage is made: Don't be blinded by impressive demos; stay curious about what's underneath.
2. Treat AI as a contact sport: You only learn what it can and can't do by building with it.
3. Guard against what he calls AI obesity: Don't gorge indiscriminately on the AI buffet. He takes a weekly 45-minute "AI amnesty" walk to ask whether he's using AI in ways that reinforce the person he wants to be. Knowing when not to use AI, he argued, may be the most important judgment call of all.

Watch the full conversation

If you missed the live session or want to revisit it, the full recording is available here: Watch the recording →

Governance and IP protection weren't just what we talked about at AI Uncorked, but they're at the heart of how Overleaf is built. If you'd like to explore how Overleaf supports secure, governed collaboration for your research and technical teams, reach out to us at enterprise@overleaf.com.

AI Uncorked is brought to you by Overleaf, the collaborative technical writing platform trusted by enterprise teams in financial services, engineering, and R&D.

Disclaimer

★ The views expressed by Maximilian Freier should not be reported as representing the views of the European Central Bank (ECB). The views expressed are those of the speaker and do not necessarily reflect those of the ECB or the Eurosystem.

References

[1] https://www.ecb.europa.eu/press/key/date/2026/html/ecb.sp260323_1~1e06784a89.en.html