The mistake: assuming data that’s clean enough for a human dashboard is clean enough for an autonomous agent. It isn’t, she said, and closing that gap is now the central challenge for businesses trying to scale AI beyond pilot projects.
Ahmad’s argument builds on a thesis she laid out earlier this year in a Fortune commentary piece, where she described enterprise software as undergoing what she called the biggest collision in the history of software: a shift from deterministic systems that always produce the same output, to generative AI that reasons probabilistically and can produce different answers from the same inputs. The throughline connecting that piece to her remarks in London is that this unpredictability isn’t a flaw to be engineered away, but property that businesses need to learn to manage.
In practice, Ahmad said, that means accepting that AI agents are no longer confined to advisory roles. Agents are increasingly deployed autonomously, booking orders in ERP systems, publishing marketing campaigns and sending emails without a human approving each step. But autonomy isn’t binary, she argued: it depends on the risk attached to a given decision.
An agent might be trusted to spend $5,000 on SEO optimisation without oversight but flagged for human review if it proposes spending a million. “There’s an element of risk to every decision,” she said. In some domains, such as a doctor reviewing an agent’s analysis of medical imaging, she said a human stays in the loop on every single decision.
Rather than relying solely on people to perform that oversight, Ahmad pointed to a growing pattern of “guardian” or “verifier” agents, AI systems built specifically to police the decisions of other AI systems.
She cited Deutsche Telekom, which has deployed swarms of agents to analyse network data and suggest configuration changes, with a separate guardian agent applying business logic to approve or block those changes before they go live. She gave a similar example from a financial services customer running distributed trading agents, where a verifier agent can kill a trade at any point if market conditions shift. Guardian agents, she said, are becoming part of how agent swarms are designed from the outset, rather than a bolt-on safeguard.
Much of Ahmad’s talk focused on why enterprise data readiness is harder than it looks. She described what she calls the “dashboard fallacy”: the assumption that data curated well enough to inform a human decision-maker is automatically good enough to train or run an agent. A dashboard works because the human reading it brings context the data itself doesn’t contain, for instance, finance and marketing teams at the same company often define a basic term like “active user” differently.
Agents don’t have access to that unwritten context unless it’s explicitly captured, which is why Google Cloud has been pushing customers to evolve their technical metadata catalogues into what it calls a knowledge catalogue, one that captures business rules and definitions rather than just data lineage.
That context problem compounds at scale, Ahmad warned. Hand-coding business context into a single agent is manageable; doing it across thousands of agents becomes a mess, since the context ends up scattered across every individual agent rather than managed in one place.
On where companies are getting AI deployment right, Ahmad pointed to businesses using agents themselves to do the underlying data engineering: building data pipelines and governance structures with AI rather than manually, an approach she called “intent-driven engineering.”
On where it goes wrong, she flagged dark data: the 80 to 90% of enterprise data that’s unstructured and was never built for analysis, let alone for feeding autonomous agents with appropriate access controls.
RELATED STORIES
New Google Cloud and Deloitte AI Studio to fuel the agentic enterprise
Google Cloud declares the pilot era over as agentic AI takes centre stage at NEXT 2026
