Most enterprises entered 2026 believing they had an AI strategy. David Garrett, a data and AI advisor with decades of experience across financial services, energy, and commodities, thinks the more honest question is whether they have a data strategy to back it up.
“AI ambition to data reality,” he says. “A lot of boards are very ambitious. But boardrooms by nature are also risk-averse. They need to tie that ambition to reality.”
Garrett spoke to Cititec as part of our Voices of Energy series. During our discussion, he revisited arguments he made in our 2025 Insights discussion: that companies needed to master their data, trace its lineage, and build the right teams.
His verdict: The core problem remains largely the same.
The data problem that hasn’t moved
A year ago, Garrett argued that unmanaged, poorly understood data was the primary constraint on meaningful AI adoption. In 2026, he sees little evidence that most large organisations have closed that gap.
“There’s still unmanaged data, poorly understood data, weak data foundations which are driving AI risk right now,” he says. “Full lineage is still beyond many companies. When did the data enter the organisation? What changed with it? Who touched it? That’s a big ask for most companies.”
What has shifted, he suggests, is awareness at board level — though he is careful about how much weight to place on that. “The boardroom is waking up and becoming more educated, but not enough. In large organisations, boards often see AI through layers of interpretation. That may come from internal teams with their own incentives, or from external consultancies that are not fully familiar with the business reality. Either way, there is a risk that the board hears a polished AI story rather than the operational truth.”
The structural difficulty is one Garrett traces back to a fundamental ownership problem. When IT holds data but doesn’t understand its business value, organisations end up sitting on what he describes as a gold mine they cannot access. The value gap is compounded when the originating team moves on, taking their legacy knowledge with them. “They should know exactly what data they have. They should know the business value of the data. They should know the regulatory consequences of not managing that data correctly.”
He draws a pointed distinction between large enterprises and smaller, entrepreneurial firms: “Small companies know where their data is. They know its value. They distinguish themselves by creating a data moat. In big enterprises, that story got lost.”
Why explainability is now the central question
Ask Garrett why data lineage and governance matter more in an AI context than they did before, and he responds in a single word: explainability.
“You see this on trading floors. If an AI system supports a trade, recommends an action, or flags a risk, you can’t just accept the output. You need to know why it reached that conclusion. The more complex the decision, and the more regulated the environment, the more explainability is needed.”
This is not an abstract concern. For firms using AI-assisted decision-making in regulated environments — whether in energy trading, risk management, or financial services — the ability to reconstruct why a system made a decision is a compliance requirement, not a preference.
If an algorithm informed a trader’s decision, and that trade later came under scrutiny, the firm would need to explain the basis for the recommendation. “You need to explain every step: what data was used, what changed, and why the system produced that output. If you don’t have explainability, and you don’t have a historical record of the data the system used, you can’t answer the question.”
The challenge compounds as models evolve. When a frontier model is updated, its parameters shift. When data used to train or inform that model changes, its outputs change. “You have to be able to audit how decisions were made, especially when the model, the data, or the parameters change. What changed, why did it change, and did the output change as a result? Then you have humans trying to understand all of that across big data, large models, and complex systems. It’s hard, complex work, likely beyond human scale.”
The last two years produced a wave of AI pilots across almost every sector Cititec covers. Garrett’s assessment of what most of them actually demonstrated is unflinching.
“When companies try and build AI solutions on top of poorly understood or poorly governed data, you get impressive pilots because of the curated context and limited scale. You can then end up with very dangerous production environments.”
The reason pilots succeed where production fails is almost always the same: pilots run on curated data. The data has been cleaned, structured, and bounded specifically to make the use case work. Move into production, where documents arrive in unexpected formats, where data comes from multiple sources with different lineage, and the foundations give way.
“Garbage in, confidently wrong answers out.”
He is also critical of the model-selection complacency that tends to set in among pilots once the model appears to be working. “What model were they using? Were they using an older GPT-4-class model? Is it already a dinosaur in their context? Should they move to a newer version or switch fully? And if they change the model, do they retest everything?”
Token costs, often invisible at pilot scale, become significant at production scale, and few organisations have built the FinOps discipline to track them.
His summary is direct: “Many pilots look impressive, but they are not yet strategically ambitious.”
What separates a capability from a pilot
The question Cititec put to Garrett — what separates an AI pilot that looks impressive from one that genuinely changes how a business operates — produced one of his sharper formulations.
“A pilot becomes a capability when it survives contact with the real business.”
He draws a distinction between capacity and capability that is worth unpacking. Cloud infrastructure offers near-unlimited capacity. Capability (ie. performing a task reliably, repeatedly, and in a way that’s provable against defined standards) is another thing completely. “Capability has to be there, provable, comparable, and show what value it’s adding.”
This is where he connects AI adoption to the broader question of human roles in an evolving organisation. “Am I easily replaced, after I’ve matured my capabilities in that space for a long time, by an agentic agent? Potentially… You need to look at the agent’s track record. Did it get things right 80% of the time? That may not be enough. It needs to meet the accuracy, reliability, and auditability threshold required for the business process. And if you compare that to a human with the same capability — is that human 100% right? So you have a certain confidence range you’re willing to accept.”
The honest answer, he suggests, is that most organisations are not currently testing this at all. Capabilities are being claimed; they are rarely measured.
What separates a capability from a pilot
Garrett’s characterisation of the gap between organisations making genuine progress and those that are not reads as a single line.
“The winners connect the dots. The laggards collect pilots.”
In his framing, mature organisations are not running AI as a side experiment. They have connected data governance, technology decisions, business processes, and change management into a coherent operating model. Their teams function as teams rather than competing factions, and risk and compliance are involved from the beginning.
A useful example is JPMorgan Chase, which has moved beyond isolated AI experimentation by tying data and AI strategy directly into its operating model. In 2026, Reuters reported that the bank appointed Guy Halamish as COO of its commercial and investment banking division, with responsibility for overseeing data and AI strategy. Under this structure, chief data and analytics officers across payments, global banking, securities services, and markets report to both Halamish and their respective business heads. The stated focus is not simply more pilots, but improving data quality, strengthening governance, preparing infrastructure for AI agents, and transforming operational processes such as credit and client onboarding.
That is the distinction Garrett is pointing to: mature organisations connect AI to business ownership, data quality, governance, and measurable operational change. Laggards collect pilots, often mistaking activity for progress. Garrett has seen the pattern repeatedly: a head of AI or CIO delivers a stream of pilots, but without provable metrics, clear ownership, or evidence that capability is being built. The result is higher costs, greater complexity, and little operational change.
“Organisations do not think about it holistically,” he says. “They don’t audit their systems to determine whether they’re meeting compliance and governance requirements. What are they actually delivering?”
The 2026 message: organisational trust
In 2025, Garrett’s message to enterprise leaders was to master their data. His message for 2026 is different in emphasis, though it builds on the same foundation.
“Mastering data is still essential. But it’s no longer the full story. The next step is building organisational trust.”
He describes a vision of AI systems that can be interrogated. Where a user can not just ask for an answer, but for the citations, the data sources, and the reasoning claim that produced the answer. “I want a conversation with AI where I trust what comes out of its mouth. Where did the data come from? What’s the risk of using this data? It tells me what the story was.”
This is not, he is clear, about trusting AI in the abstract. It is about building the organisational conditions under which AI-generated outputs can be verified, explained, and defended to regulators, auditors, and the people accountable for decisions made on the basis of those outputs.
The data lineage work, the governance frameworks, and the explainability requirements all work together to enable trust. And trust, in Garrett’s view, is what the next phase of AI adoption will actually be built on.
David Garrett spoke to Cititec for Commodities Tech Insider, part of our Voices of Energy platform. Voices of Energy brings together senior practitioners from energy, commodities, and data-intensive sectors to explore how technology is reshaping markets, organisations, and talent.
Looking for more insights?
Get exclusive insights from industry leaders, stay up-to-date with the latest news, and explore the cutting-edge tech shaping the sector by subscribing to our newsletter, Commodities Tech Insider.
