The Verdict: Five Voices, One Direction

When I started this series, I had one question in mind: is the Corporate Treasurer a dying breed at the dawn of AI? I was not looking for a comforting answer. I was looking for guidance to arrive at my own response.

The different conversations brought me something more nuanced than either the fear narrative or the vendor optimism typically on offer. Across five conversations – with an AI assurance researcher, a treasury technology architect, a sitting corporate treasurer, an AI – a coherent picture emerged. Not identical views. Not a consensus that papers over real disagreement. But a direction. And it points somewhere specific.

This is what they said, where they agreed, and what it means for the function going forward.

The Five Voices

Andreas Hafver  ·  Principal Researcher, NorwAI / DNV

Andreas came at the question from the angle of trust and assurance. His argument: the debate about whether AI can be trusted is being conducted at the wrong level of abstraction. The question is not whether the model is interpretable internally – it is whether its outputs can be verified externally. His solution is task decomposition: breaking complex treasury workflows into smaller, auditable steps where each output can be checked before the next action is taken. AI runs the calculator. The treasurer – or another layer of the system – verifies the result. His most memorable formulation: treat AI like a brilliant but junior employee. High capability, zero institutional judgment. Supervise accordingly.

Marcus Gullers  ·  SkySparc / OmniFi

Marcus addressed the infrastructure problem that makes everything else theoretical. The gap between what AI can do and what it currently does in most treasuries is not a model problem – it is a data connectivity problem. Pre-trained models like GPT or Claude already understand treasury conceptually. What they lack is your position data, your bank feeds, your actual exposure. MCP – Model Context Protocol – is the standardised layer that closes that gap without requiring a multi-year integration project. OmniFi is already delivering this as a bolt-on, sitting on top of existing TMS infrastructure and making it AI-readable within days, not years. His core message: the technology is not the bottleneck. The data architecture is

Christian Gjerde  ·  Head of Group Treasury and Investor Relations at Elopak

Christian provided the ground truth the rest of the series needed. He is running a complex multinational treasury function with real currency exposure across multiple countries, still largely manually in 2026. His RFP process produced an AI-functionalities that his CFO and CIO unanimously backed exploring. Not because these were proven, but because the window to get ahead of the curve felt real. His framing on fear versus excitement: not anxiety about replacement, but a genuine to-do list that never gets shorter and a belief that AI is finally the tool that changes the ratio of operational burden to strategic contribution. His line on accountability was the sharpest in the series: AI will not blush if it makes a mistake. He will. And that accountability stays human, regardless of how autonomous the system becomes.

Claude  ·  AI, Anthropic – interviewed by the author

The most uncomfortable voice in the series, deliberately. When asked directly whether the treasurer is being eliminated, the answer was unambiguous: the elimination thesis is a vendor fantasy, but the evolution thesis requires intellectual honesty about what actually evolves. Headcount will shrink. The profile required will shift dramatically upward. The professionals most at risk are not the most senior ones – they are the mid-career specialists whose entire value proposition was manual execution expertise in a specific operational domain. On architecture: decoupled, system-agnostic AI layers win over monolithic TMS platforms for complex multinationals. On macro blindspots: AI handles distributional risk well and regime change badly. On skills: the treasurer of 2030 is a domain-fluent systems thinker, not a quantitative technician.

Where They Agreed

1.  The operational layer is AI’s territory – immediately

All four voices converged on this without prompting. Reconciliation, reporting, cash concentration, deviation analysis, intraday positioning – these are tasks that are high-volume, rule-governed, and backward-looking. The cost of an AI error in this domain is recoverable. The benefit of removing human bandwidth from this domain is immediate. Andreas called it ‘running the calculator.’ Christian called it ‘everything that does not directly impact economically.’ Marcus built the infrastructure to enable it. The AI described it as the first corridor of graduated autonomy. Same conclusion, four different framings.

2.  The strategic layer stays human – for now

With equal unanimity, all four drew a line at decisions that are forward-looking, economically material, and contextually complex. FX hedging strategy. Funding decisions. Banking relationship management under stress. Responses to regime change – sanctions, capital controls, central bank interventions that invalidate historical patterns overnight. Christian’s FX market example was the most concrete: analysts and economists with full market access are wrong more often than right on currency direction. An AI trained on the same data will inherit the same failure modes, plus additional fragility to the psychological and geopolitical variables that move markets but leave no statistical trace in advance.

3.  Data is the actual bottleneck, not the model

Marcus made this argument most explicitly, but it was implicit in every other conversation. Andreas’s task decomposition framework requires clean, auditable data at each step. The AI interview identified the trusted data layer as a precondition for agentic deployment, not a feature of it. Christian’s entire Q1 answer was a description of what happens when that layer does not exist: manual processes, capacity constraints, decisions made on data that needed to be assembled before it could be used. The AI models are ready. The data infrastructure, in most treasuries, is not.

4.  Human accountability is non-negotiable

Christian said it most directly: AI will not blush if it makes a mistake. But the accountability framing ran through every conversation. Andreas’s task decomposition is specifically designed to keep a human in the verification loop at each stage. The AI interview described ‘graduated autonomy with hard human gates’ as the only guardrail architecture that actually works. Marcus’s MCP framework is specifically designed as an observable, auditable gateway rather than a black-box pipeline. The convergence here is not sentimental. It is structural: in a function where a single error can move material amounts of capital, the accountability must attach to a person who will feel the consequences.

Where They Diverged

The divergences are at least as instructive as the agreements.

On how fast the transition happens

Andreas and Marcus, coming from technology and research backgrounds, were more optimistic about speed. The infrastructure exists. The models are capable. The transition is a deployment problem, not a capability problem. Christian, sitting in the operational reality of a treasury that is still largely manual in 2026, offered an implicit corrective: the distance between ‘the technology exists’ and ‘the technology is working in my treasury’ is measured in years of data cleanup, change management, and careful testing. ‘Gradually and slowly’ was his phrase. That is not pessimism. It is operational honesty.

On the monolith versus the open layer

Marcus argued clearly for a decoupled, system-agnostic architecture – AI as an orchestration layer above existing infrastructure, not embedded inside any single vendor’s ecosystem. The AI interview reached the same conclusion for different reasons: a multinational that bakes its intelligence layer into one TMS vendor’s proprietary stack loses the ability to swap components without rebuilding from scratch. Christian’s RFP found a third direction: AI-native platforms that are neither legacy TMS nor bolt-on middleware, but greenfield build designed from the ground up around the intelligence layer. For a treasury starting without legacy constraints, that may be the most coherent path. For a treasury already running on established infrastructure, the Marcus / OmniFi approach is the more realistic near-term option.

On the generational question

Christian was the only voice who addressed this with personal candour. He described two camps among seasoned peers: those who are anxiously curious about job security, and those who are constructively curious about what AI enables. He sees both reactions in himself. The AI interview was more clinical about it: the mid-career operational specialist is the most structurally at risk. Andreas’s framing was more optimistic: the function evolves, the practitioner who evolves with it becomes more valuable, not less. These are not necessarily contradictory. The outcome depends on which camp the individual chooses.

What This Means in Practice

The question I started with – is the treasurer a dying breed? – has a clear answer by now. No. But the answer comes with a condition that deserves to be stated plainly.

The treasurer who survives and thrives in the AI era is not the one who understands AI best. It is the one who understands treasury best, and then learns to govern AI as the tool that amplifies that understanding. The sequence matters. Christian’s advice to the next generation was the most direct articulation of this: build the fundamentals first. Do not let AI become a cushion for not learning yourself. Do not let it do the job for you before you understand the job.

The four areas where AI is already delivering in treasury today, and where any function not piloting these in 2026 is operating at a structural disadvantage:

• Cash reconciliation. Pattern-matching that identifies mismatches and root causes in minutes, not days, and learns from each correction cycle.
• Cash forecasting. LLMs interrogating variance drivers across ERPs, bank feeds, and business commentary simultaneously, replacing the manual investigation cycle.
• Exposure monitoring. AI agents tracking FX and commodity exposure thresholds in real time and flagging mitigation options before the treasurer opens their inbox.
• Treasury briefings. Dedicated agents synthesising position, forecast, hedge rationale, and cost of debt into a single briefing – on demand, not after a reporting cycle.

And the areas where the human remains not just useful but essential: anything that requires a counterparty relationship, a political judgment, a contextual read on a market that has just shifted its regime, or a conversation with a board that needs to understand not just what the numbers say but what they mean.

A Final Note

I started this series with a reference to “nature is metal” – the raw footage of predator and prey, and the principle of survival of the fittest. I stand by the framing, but I want to be precise about what it actually means here. And before I close, I want to leave you with one historical footnote worth sitting with.

The species that survives is not necessarily the strongest. It is the one most responsive to change. The treasurer who treats AI as a threat to be waited out will be displaced by the treasurer who treats it as infrastructure to be mastered. The function that resists automation at the operational layer will continue to spend its best people on reconciliation and reporting while competitors redirect that capacity toward strategy and risk.

In the mid-19th century, whale oil was the crude oil of its era. It powered the lamps of homes and factories across Europe and North America, lubricated the machines of the Industrial Revolution, and generated fleets of hundreds of ships out of ports like New Bedford and Nantucket. At its peak it represented roughly 5% of US GDP. It was not a niche business. It was foundational infrastructure for the world as it existed then.

Then kerosene arrived. Not overnight – gradually. And the industry that had dominated illumination for a century did not disappear in a moment; it faded, as those who could not adapt found their expertise increasingly irrelevant to a world that had quietly moved on.

I am not predicting that the treasurer disappears the way the whaler did. The analogy has limits. But I am suggesting that the treasury professional who gets too comfortable – who assumes that because the function has always been necessary, their particular version of it always will be – is making the same mistake those captains made in the 1860s.

The value of what treasury does is permanent – or at least, it is as permanent as the underlying markets that make it necessary. As long as currencies fluctuate, interest rates diverge, commodities swing, and capital needs to move across borders, there will be a function whose job it is to navigate that complexity. That is a reasonable assumption. Though I will admit: if AI ever solves volatility overnight, we will all have bigger things to worry about than job titles.

What is not permanent is the way that work gets done. And that is the part that is changing now. The Augmented Treasurer is not a prediction. It is an invitation – to adapt, to evolve, and to be the one steering the technology rather than waiting to see what it does to you.

The four voices in this series – from research, from technology, from the model itself, and from the operational front line – each arrived at this conclusion from a different angle. The direction is clear.

Human + AI = The Augmented Treasurer

“The treasurer is not a dying breed. The treasurer is becoming something more capable, more strategic, and frankly more interesting than the role has ever been. But only for those who choose to evolve.”
– Rafael Dominguez, SkySparc

‍

This is the concluding piece in the SkySparc thought-leadership series “The Augmented Treasurer,” which brought together perspectives from academia, treasury technology, independent advisory, artificial intelligence, and corporate practice. The full series is available at skysparc.com/insights-and-articles.

Rafael Dominguez has spent over 20 years in corporate treasury, including interim positions as Head of Cash Management and years as an independent advisor at SkySparc, selecting and implementing treasury management