There is an obvious irony in asking an AI whether AI will replace the corporate treasurer. I asked anyway — and I wanted it to be specific. Over 20 years in treasury, including interim roles as Head of Cash Management and years as an independent advisor running bank and TMS selections across the Nordics and globally, my expectations to develop an automated, well- calibrated TMS has been challenged.

What follows is an edited transcript of that interview. The questions were designed to probe the uncomfortable edges: real elimination risk, the guardrails problem on fragmented tech stacks, the monolith-versus-middleware architecture debate, where AI gets structurally blindsided by a volatile macro environment, and what the next generation of treasury professionals actually needs to build. The answers are Claude’s own.

“The elimination thesis is a vendor fantasy dressed up in legitimate technology. But the evolution thesis requires intellectual honesty about what specifically evolves — and how fast.”

— Claude, by Anthropic

‍

The Five Questions

Q1  The “Eradication vs. Evolution” Reality Check

Looking at the rapid rise of Agentic AI and API-driven solutions, are we looking at the actual elimination of the corporate treasurer role, or is that a tech-vendor fantasy?

The elimination thesis is, frankly, a vendor fantasy dressed up in legitimate technology. But the evolution thesis requires intellectual honesty about what specifically evolves, and how fast.

The core distinction: AI can automate execution within defined parameters far better than any human. Cash positioning, intraday sweeps, threshold-based FX hedging, three-way matching — these are high-volume, rule-governed, latency-sensitive tasks where agentic AI genuinely outperforms a human. The treasurer who manually runs these processes in 2030 will look like the CFO who still consolidates on Excel in 2010.

What does not disappear is the judgment layer. The “day in the life” of a 2030 Treasurer in a mature, AI-enabled treasury looks something like this: they arrive to a dashboard that has already positioned cash across 40 entities overnight, flagged three anomalies for review, pre-executed a rolling FX hedge within the board-approved policy band, and escalated one counterparty limit exception that requires a human decision. The treasurer spends the first hour not doing treasury operations, but supervising and calibrating them.

The rest of the day shifts toward what has always been the highest-value treasurer activity but was crowded out by operational noise: capital structure thinking, banking relationship strategy, board-level financial risk communication, and scenario planning for events with no historical precedent.

Key point: The role is not eliminated. But the number of roles shrinks significantly, and the profile required shifts dramatically upward. That is the uncomfortable truth the “evolution” framing tends to soften.

‍

Q2  The Core Friction of “Agentic AI” vs. Legacy Architecture

If AI’s true value lies in moving from insight to action, how can treasurers safely deploy AI on top of fragmented tech stacks without introducing catastrophic operational or fraud risks?

This is where vendor marketing and operational reality diverge most sharply. The fundamental problem is that agentic AI requires a trusted data foundation as a precondition, not as a nice-to-have. The gap between a predictive model and a reliable action layer is the gap between a clean, normalized, reconciled data API and the actual fragmented reality of a multinational with five ERPs, eleven banking relationships, three legacy TMSs, and a bank connectivity layer that still partially runs on SWIFT MT messages with manual intervention.

The fraud and operational risk vectors when you deploy agents on top of that fragmented stack are not theoretical. They include: stale or conflicting position data causing the agent to initiate a sweep that overdrafts a real account; a poisoned data feed that the agent acts on before a human reviews it; and payment initiation agents that can be manipulated through the data layer rather than the approval layer, bypassing controls designed for human workflows.

The guardrail architecture that works

The principle I would call graduated autonomy with hard human gates: the AI acts fully autonomously within a narrow, pre-approved action corridor — intraday sweeps within the same legal entity, below a defined threshold, within a currency already open for automated dealing. Outside that corridor, it generates a recommendation with reasoning and queues it for human approval.

The critical technical requirement is an immutable audit trail sitting outside the agent’s own write access. This is not a technology problem; it is an architecture governance problem, and it is exactly the kind of thing that gets deprioritized in implementation projects under budget pressure.

Where human supervision begins: Any action crossing legal entity boundaries (tax and transfer pricing implications), any modification of counterparty relationships or credit lines, any FX execution above the pre-approved notional threshold, and any situation where the AI’s confidence on its own data inputs falls below a defined threshold.

Q3  The System-Agnostic Challenge

Do you see the future of treasury tech belonging to monolithic, all-in-one TMS platforms with built-in AI, or to decoupled, system-agnostic AI layers that sit on top of existing multi-ERP/multi-TMS landscapes?

The monolithic, all-in-one TMS with embedded proprietary AI serves the vendor’s interests far better than it serves a complex multinational’s interests. The genuine value of AI in treasury is its ability to synthesize signals across the entire financial data landscape — bank feeds, ERP payables, FX market data, interest rate curves, counterparty credit spreads, internal cash flow forecasts. A proprietary AI layer embedded inside one TMS vendor’s ecosystem can only ingest and act on data that flows through that vendor’s pipes.

The architecture that genuinely serves a complex multinational is a decoupled data orchestration layer — a treasury data fabric — that normalizes and reconciles data from all sources into a trusted information layer, then exposes that layer to AI models through well-defined APIs. The AI reasoning and execution layer sits above this fabric, system-agnostic, and the TMS becomes one of several execution targets rather than the intelligence hub.

The MCP model is exactly the right conceptual framing: the AI talks to tools through standardized interfaces, rather than being baked into any single tool’s proprietary logic. This gives the multinational the ability to swap individual components — change their primary bank, upgrade their ERP, adopt a new risk platform — without rebuilding their AI layer from scratch.

The honest answer: The decoupled, system-agnostic architecture is structurally the right answer and where the market is heading. But the timeline is gated by the slowest link in the banking API ecosystem — not by the readiness of the AI models themselves.

Q4  Volatility, Fragmented Economics, and the Human Element

In a world where macroeconomic anomalies and unexpected political actions are the new normal, where does AI blindside a treasury?

AI systems are genuinely excellent at distributional risk — events that have happened before, in forms that left traces in training data, at frequencies that allow statistical inference. Seasonal liquidity patterns, commodity price correlation to FX moves, interest rate transmission through credit spreads: a well-trained model handles these better than most human practitioners.

Where AI gets structurally blindsided is regime change — when the underlying rules of the game shift in ways that make historical patterns actively misleading. We are in a period characterized by central bank policy divergence with no clean historical analog; geopolitical fragmentation rewiring supply chains and currency corridors faster than any model can retrain; and political interventions — tariff announcements, sanctions, capital controls — that arrive as discrete shocks with no statistical precursor.

Specific blindside cases

The model assumes USD liquidity will remain available in a particular corridor because it always has been. A sanctions regime change makes it structurally unavailable overnight. The AI has no mechanism to anticipate this; a treasurer with banking relationships at the correspondent bank level gets a call the afternoon before the announcement becomes public.

The model optimizes FX hedging based on implied volatility surfaces. A central bank intervention breaks the implied/realized vol relationship in a currency pair. The model’s hedge ratios are now based on a distributional assumption deliberately invalidated by policy action.

Where no algorithm succeeds: Bilateral relationship negotiation with banks under stress; cross-border transactions requiring navigation of capital control regimes through local banking relationships; and board-level communication of financial risk in plain language. An AI can generate the numbers. It cannot read the room.

Q5  The Skillset Shift for the Next Generation

What does the ideal profile of a treasury professional look like when quantitative skills and prompt engineering become embedded into finance functions?

The profile that survives and thrives is not the quantitative technician, and it is not the pure relationship manager. It is the domain-fluent systems thinker — someone who understands treasury deeply enough to define what the AI should optimize for, technically literate enough to interrogate whether it is actually doing that, and commercially credible enough to represent the function at the executive and banking relationship level.

The five skills to build now

•       Deep process fluency before automation. Understand cash positioning, reconciliation, and FX execution as manual processes before automating them. The professionals who will be dangerous with AI are those who know exactly what the AI is supposed to be doing and can immediately identify when it is wrong.

•       Data architecture literacy. Not programming, but the ability to read a data model, trace a reconciliation break to its source, and specify what a clean data feed should look like. This skill bridges treasury domain knowledge and technical teams — it is currently extremely rare and extremely valuable.

•       Structured critical evaluation of AI outputs. Knowing how to stress-test a forecast, identify when a model’s confidence interval is suspiciously narrow, and design the exception-handling protocols that determine when the AI escalates to humans. Prompt engineering is a component, but downstream of the judgment about what questions to ask.

•       Relationship capital at the right level. In a world where transactional banking is increasingly automated, the relationship that matters is with the structuring and credit teams who can move when you need an emergency liquidity facility, or the regulatory contacts who give early intelligence on capital control changes. This is built over years and cannot be substituted.

•       Strategic communication. The ability to translate AI-generated analysis into board-level narrative. An AI can produce the numbers; the treasurer must own the story.

“Your job is not to compete with the AI at tasks the AI does better than humans. Your job is to be the person who decides what the AI should be doing, verifies that it is doing it correctly, and handles every situation the AI was not designed for.”

— Claude, by Anthropic

‍

Editorial Note

This interview was conducted in May 2026 by Rafael Dominguez as part of the SkySparc thought-leadership series “The Augmented Treasurer.” The interviewee is Claude Sonnet 4.6, developed by Anthropic. The questions were designed to probe the genuinely difficult edges of the AI-in-treasury debate — not to produce a promotional narrative. The answers are the model’s own, shaped by the interviewer’s questions but not edited for comfort.

Rafael 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 systems across Nordic and global clients.