The Treasury Infrastructure Ceiling: Why enterprise AI investment is failing in finance, and what it takes to cross it

FinSaAIstra Intelligence | Treasury Infrastructure Series | June 2026

Executive Signal

Corporate treasury has already spent on AI. The bills are in. The results are not. A July 2025 MIT study measured $30–40 billion in enterprise AI investment and found that 95% of organizations report no measurable business return, with less than 5% successfully deploying task-specific projects beyond general-purpose LLMs. Treasury and finance are no exception. The bottleneck is not the model — it is what the model is being asked to run on.

Most treasury environments were built for a different era. The median Treasury Management System is a decade old. Treasury technology spend as a percentage of revenue runs at 0.01%–0.03%, against 6–8% for enterprise IT. A Fortune 500 treasury managing 100–200 legal entities operates on roughly 12 full-time employees, unchanged for ten years.

Source: MIT NANDA, The GenAI Divide: State of AI in Business 2025; J.P. Morgan Payments, Treasury 2.0: Built for a New Era (2026)

The Treasury Infrastructure Ceiling — The structural point at which legacy treasury architecture (disconnected TMS platforms, unstructured data, absent policy engines) prevents an institution from graduating from AI experimentation to AI-governed financial operations, regardless of model investment.

Verified Market Signals

🧭 AI capability is accelerating faster than treasury infrastructure is adapting. Frontier models scored a complexity of ~11.5 on software engineering benchmarks at 50% accuracy in 2026, up from near zero in 2022. At 80% accuracy (still below the production threshold for regulated financial workflows), even the most advanced models cluster near the bottom. Source: METR, Task-Completion Time Horizons of Frontier AI Models 2026.

🧭 $761 billion in working capital is locked in the performance gap between top- and bottom-quartile companies within industries (DSO, DPO, DIO). That gap has widened since 2023. Treasury holds the instruments to unlock it — cross-functional cash visibility, scenario modeling, supply chain finance access, counterparty risk intelligence — but the constraint is whether treasury has the infrastructure to act as orchestrator rather than scorekeeper. Source: J.P. Morgan Payments, Treasury 2.0 (2026); Capital IQ, April 2026.

🧭 The 2026–2028 regulatory calendar has no quiet periods: ISO 20022 mandatory by November 2026, Basel IV phasing through 2028, DORA and the EU AI Act in simultaneous rollout, GENIUS Act/CLARITY Act/NACHA fraud monitoring reshaping US digital currency and payments controls, HKMA stablecoin licensing and RTP interoperability moving in Asia-Pacific. Over half of payments firms remain only partially compliant with ISO 20022; 67% of payment errors trace to data quality issues. Source: J.P. Morgan Payments, Treasury 2.0 (2026).

Institutions on real-time payment rails face a compounding obligation: regional model risk management expectations placing explainability/auditability at the center of AI deployment, plus DPDP Act data governance obligations interacting directly with AI training/inference. Agentic treasury on fast payment rails requires governance infrastructure matching rail speed — most institutions haven’t built it.

Structural Shifts

Conversational AI → Agentic Treasury Operations — The first AI wave produced summaries and dashboards. The agentic wave handles multi-step, multi-system workflows autonomously: cash positioning, regulatory intelligence, fraud screening, reconciliation, hedge execution. Different category of infrastructure requirement entirely.

Monolithic TMS → Composable Agent-Native Architecture — TMS remains the plumbing (bank connectivity, SWIFT, payment files, core accounting) but agent-native operations need a layer above: semantic layer with governed definitions, vector-indexed knowledge base, RAG engine, policy engines encoding delegation rules. Most 10+ year-old TMS platforms have none of these.

Working Capital Measurement → Working Capital Orchestration — Only 12% of treasuries consider themselves truly strategic, down from 17% in 2022 despite a decade of tech investment. The next shift requires treasury to move from scorekeeper to active orchestrator of the $761B opportunity.

Reactive Compliance → Regulatory Intelligence as Infrastructure — Multi-jurisdictional regulatory density (ISO 20022, Basel IV, DORA, EU AI Act, digital currency frameworks) converging simultaneously means event-by-event response is a losing strategy. Winners run regulatory intelligence as a continuous operational function.

Systemic Implications

AI in treasury is failing in sequence, not in concept. 80% of organizations have investigated AI for treasury; 60% have piloted; less than 5% have successfully deployed. The 95% failure rate is a sequencing failure — organizations buying AI before building the infrastructure it requires to function.

The control gap is three layers deep, and most institutions have addressed only one:

1. Traditional risks (BEC, account takeover, check/card fraud) — established control frameworks exist.
2. AI-amplified threats (synthetic identity at scale, executive voice-cloning deepfakes, high-velocity model-speed scamming) — qualitatively new attack vectors using the same AI treasury teams are adopting.
3. AI model risks (hallucination producing compliant-sounding wrong outputs, model drift in cash forecasting, training data manipulation, cascading agent failures) — least addressed, and risk tolerance in payments is functionally zero.

76% of companies faced payment fraud attempts in 2025; 83% faced at least one account takeover.

Stablecoins are at scale; the treasury compliance architecture for them is not. Stablecoin market cap reached $322B by May 2026, with $79T in unadjusted transaction volume over the preceding 12 months — but strip out crypto trade settlement and the real-economy commercial payments use case remains nascent. Compliance blockers: KYC/AML for digital currency holders, wallet identification/validation infrastructure, tax treatment clarity, cross-border regulatory uniformity.

The talent pipeline risk is being created now: 43% of companies plan to replace roles with AI, targeting operations, back office, and entry-level positions. Today’s analyst is the 2030 AI governance officer — eliminating entry-level hiring for short-term efficiency builds a leadership deficit that will constrain future AI governance capability.

CXO Action Layer

Board-level The board question has changed from “are we investing in AI?” to “does our infrastructure support the AI we are investing in?” Four dimensions require board-level audit: data unification, policy engines (rules encoded for agent consumption, not stored in documents), agentic delegation (defined boundaries for initiate/flag/escalate), and auditability framework (traceable record with data provenance and authorization chain intact). Until addressed, AI investment in treasury should be classified as capital allocation risk, not technology initiative.

Procurement Reality The evaluation question shifts from “what can your model do?” to “what does your model produce as a defensible audit trail, and who is liable when it fails?” Require decision-level explainability, demonstrated ISO 20022 compatibility (mandatory Nov 2026), and applicable central bank model risk framework compliance. For real-time payment ecosystems, vendor contracts must specify data sovereignty, cross-border transfer constraints, and vendor obligations when an agent-initiated output is challenged in regulatory review.

Architecture Implication Sequence separates the 5% that succeed from the 95% that don’t. Build order: data unification first → policy engine construction → semantic layer → knowledge base (vector-indexed SOPs/regulatory guidance/historical data) → RAG engine and LLM layer → domain agent deployment in controlled pilot (cash forecasting is lowest-risk start) with human-over-the-loop governance active from day one. Starting at the agent layer without the preceding four is the reason 95% of investments stall.

FinSaAIstra Law: A treasury that cannot explain its AI decisions to a regulator has not adopted AI as infrastructure. It has adopted AI as risk.