The arrival of generative and agentic AI has created tensions for data leaders as organisations want employees to experiment, innovate, and identify new sources of value; but greater AI autonomy creates questions around accountability, cost, data quality, risk, and human oversight.
These challenges brought senior data and AI leaders together at an exclusive DataIQ Peer Exchange in Brussels to examine what responsible AI and data governance look like in practice. Across presentations and roundtable discussions, leaders examined how governance must move into the flow of work.
Make value the gateway to production
The ability to build an AI solution is not evidence that an organisation should deploy it. David Reed, Chief Knowledge Officer and Evangelist, DataIQ, challenged leaders to demand stronger evidence before moving proofs of concept into production. Successful experimentation needs to be followed by evidence of adoption, sustainable operating requirements, and measurable organisational value.
The roundtables reinforced David’s point. Leaders argued that organisations should assess business need, strategic alignment, technical feasibility, risk, and longer-term economics rather than allowing technologically impressive use cases to progress without sufficient scrutiny.
This reframes AI governance beyond controlling risk after a solution has been built. It starts with determining whether AI is the right solution to the business problem at all.
Governance must evolve as AI gains autonomy
Agentic AI changes the governance challenge because greater autonomy changes where control takes place.
Hesam Shokouhi-Jah, Associate Director, Data Governance at Coca-Cola Europacific Partners (CCEP), described the need for more systemic governance as humans step back from individual interactions with AI. Observability, intervention mechanisms, embedded guardrails, and the ability to investigate failures become increasingly important as systems gain greater autonomy.
Yet one principle remains fundamental: accountability stays with people.
FMO is confronting a related challenge as employees create AI agents using Microsoft Copilot tools. Felipe Cascaldi Garcia, Head of Data Management at FMO, shared an approach that combines an inventory of agents with automated risk assessment, monitoring, and escalation of more consequential cases for human review.
Organisations do not need the same amount of governance everywhere; they need proportionate governance, with friction and scrutiny calibrated to potential consequences.
Strong AI governance begins with visibility
An organisation cannot govern an AI estate it cannot see. FMO is using administrative information to identify citizen-developed agents, their owners, the tools used to create them, and usage. This visibility allows the organisation to remove unnecessary experimentation before applying deeper governance controls to the agents that remain.
The Peer Exchange roundtables raised the same issue at enterprise scale. Leaders discussed automatically registering agents when they are created, triggering reviews when they change, differentiating review requirements according to risk, and retiring inactive assets.
As AI becomes embedded within existing software, bought from suppliers, and built directly by employees, maintaining this visibility will become an essential governance capability.
Data accountability must become real
Poorly defined ownership and weak data quality become even more significant when AI can propagate their consequences. Ana Aguero Garcia, Data Governance Lead at Axpo, challenged organisations to distinguish between naming data owners and genuine accountability. Owners need the authority to influence processes, secure support, make decisions, accept risk, and escalate problems when necessary.
Axpo is making that accountability tangible through data quality. Rather than expecting data to be generically “good”, the organisation is working with owners to define asset-specific expectations, thresholds, remediation responsibilities, and escalation criteria connected to business impact. This shifts data leaders from proving that governance structures exist to demonstrating that governance produces outcomes.
Embed governance where decisions happen
Governance is moving closer to operational workflows. CCEP has extended established data and technology governance rather than creating an entirely separate AI framework, connecting AI governance with existing procurement, development, privacy, security, and project processes.
Axpo identified opportunities to embed appropriate governance checks into product development, change approval, operational performance, risk and controls, incident management, investment decisions, and AI development.
FMO demonstrates how technology can operationalise that ambition. Its approach uses Microsoft Teams, Copilot, Power Apps, Power Automate, SharePoint, and other automation to support assessment, registration, and monitoring, while retaining human oversight for higher-risk applications.
The ambition discussed in the roundtables was for governance to become sufficiently integrated into normal operations that it keeps the organisation moving rather than appearing as a separate layer of bureaucracy.
Human judgement becomes more important
Deciding where human authority must remain is pivotal for leaders. David warned that human-in-the-loop approaches become ineffective if people do not have the capacity or expertise to challenge AI-generated outputs. Organisations need explicit decision rights around consequential decisions and areas where human judgement remains essential.
However, automating routine early-career work could remove some of the experiences through which employees develop the expertise they later need to question automated outputs. Leaders need to consider which tasks AI can perform and how roles and development pathways should change alongside it.
Moving from governance activity to governance outcomes
Organisations need to judge governance differently. Counting policies, forums, owners, and controls demonstrates activity, but it does not necessarily demonstrate value. Discussions in Brussels focused on measures such as problems resolved, risk reduced, business continuity protected, remediation completed, investment returns supported, and projects delivered more effectively.
For senior data and AI leaders, that represents a significant evolution of the governance mandate. The goal is to create an operating discipline that allows organisations to make deliberate decisions, experiment safely, establish genuine accountability, and scale the applications that produce sustainable value.
Go deeper with DataIQ
The Peer Exchange discussions went further, covering practical approaches to AI inventories, automated risk assessment, decision rights, data-quality accountability, human oversight, token consumption, vendor dependency, skills, and the transition from proofs of concept into production.
DataIQ members can access the full Peer Exchange write-ups on the private Hub, including detailed insights and practical next steps shared by senior data and AI leaders.
Not yet a DataIQ member? Join data and AI leaders and access the full learnings and participate in future conversations.
