Agentic Treasury 2026: From Autonomous Treasury Index to Production-Grade Co-Pilots

TL;DR. Agentic treasury co-pilots are moving from pilots to production in 2026 — ISO 20022 data plus tool-calls, with SR 11-7, DORA, and EU AI Act controls around them.

Key takeaways

01. From experiment to production. 2024–2025 was treasury pilots.
02. The architecture: data + ISO + tool calls. A production treasury co-pilot in 2026 has three layers, in this order.
03. Use cases and metrics. Three treasury co-pilot use cases are in production at CIB scale in 2026.
04. Governance, audit, and SR 11-7 fit. Agentic treasury is a model risk problem before it is a productivity story.

A treasury co-pilot in 2026 is not a chatbot bolted onto a cash position screen. It is a bounded agent that reads ISO 20022 statements, projects intraday liquidity, and proposes — or, inside a tight policy envelope, executes — sweeps, FX hedges, and intraday repo. The pattern is the same across Corporate and Investment Banking (CIB) treasuries: agents continuously rebalance liquidity within policy bands, escalating to a human only when a band would be breached or a counterparty limit is close. Practitioner studies put the manual-workload reduction at 30-50% across cash positioning, forecasting, and exception handling, and Capgemini's 2026 banking outlook names agentic AI in treasury as one of the few use cases where 2026 spend is producing measurable cost-to-serve compression.

This article is the natural Part II of the Autonomous Treasury Index. The index defines the destination — programmable liquidity, tokenised deposits, machine-readable policy. Co-pilots are the production unit that gets a CIB treasury there without breaking SR 11-7, DORA, or the EU AI Act.

01. From experiment to production #

2024–2025 was treasury pilots. 2026 is treasury production.

Forrester's Predictions 2026: Banking and Investing is direct: in 2026, a majority of Tier-1 banks will move at least one agentic use case from sandbox into a live, measured production environment, and treasury is one of the first three to clear that bar. Capgemini's banking top-trends 2026 report makes the same call from a different angle — agentic AI investment is moving from horizontal productivity experiments into vertical, function-specific deployments, with corporate treasury, payments operations, and KYC remediation taking most of the new 2026 budget.

What changed? Three things.

First, the data layer. ISO 20022 migration completed for most major currencies in November 2025, so cash and payments data arrives in a structured form an agent can reason over without a brittle screen-scrape layer.

Second, the control plane. MCP standardised how agents call tools, and banks now have a defensible answer to the CRO's question: "what can this agent actually do?" The answer is the registry of MCP tools it is bound to, nothing more.

Third, the regulators stopped being hypothetical. SR 11-7 supervisory guidance was extended to cover non-deterministic models; DORA went live in January 2025; the EU AI Act high-risk classification regime started biting in August 2026.

02. The architecture: data + ISO + tool calls #

A production treasury co-pilot in 2026 has three layers, in this order.

Data. The agent reads ISO 20022 messages — camt.052 (intraday statement), camt.053 (end-of-day), camt.054 (debit/credit notification) — and pacs.008 customer credit transfers as they flow through the bank's payment rails. The agent reads the structured message and reconciles to the general ledger. Elire's 2026 treasury AI playbook frames this as the precondition: if the agent cannot read structured ISO 20022 data, every downstream claim about forecast accuracy is marketing.

Reasoning. A constrained foundation model — typically an internal frontier model with a fine-tuned treasury policy adapter — turns the ISO 20022 reality into a proposed action. The reasoning step never touches a payment rail. It produces a structured tool-call request: "sweep £180m from EUR nostro at BoE-tier counterparty X to GBP RTGS account Y at 14:30 to keep the GBP intraday buffer above the policy floor."

Tool calls. The agent calls MCP-registered tools. Each tool is a typed, audited function: propose_sweep, simulate_fx_hedge, query_limit, submit_pacs008_for_human_approval. The MCP tool is the only path to a real-world effect. SWIFT submission above a configured threshold routes to a human treasurer for sign-off; below threshold, the agent can submit inside a daylight policy band and the action lands in a WORM audit log within the same second.

The discipline is that the model never has database write access, never has direct SWIFT credentials, and never reads unstructured screens. The MCP registry is the security boundary, and OPA policies enforce what each agent identity can call.

03. Use cases and metrics #

Three treasury co-pilot use cases are in production at CIB scale in 2026.

Cash positioning. The agent maintains a live intraday cash position across nostro accounts, anticipates in-flight pacs.008 messages, and proposes sweeps to keep buffers inside policy bands. Reported impact: 35-45% reduction in manual reconciliation time, a measured drop in end-of-day buffer over-funding (which improves Net Interest Margin on idle cash), and intraday RTGS overdraft events trending toward zero in pilots that have completed a full quarter.

Cash forecasting. The agent ingests historical ISO 20022 flows, client behavioural signals, and known calendar events (tax dates, dividend dates, bond coupons) and produces a 1-day, 5-day, and 30-day cash forecast with a confidence interval. Mean absolute percentage error on 5-day forecasts has fallen from ~7-9% (regression baseline) to ~3-4% in the better-instrumented CIB treasuries, which materially improves the CFO's funding plan.

Exception handling. The agent triages payment exceptions — failed pacs.008 messages, sanctioned-counterparty hits, mismatched remittance — and proposes the disposition (repair, return, escalate). Triage time has dropped from ~7 minutes (human only) to ~90 seconds (human-in-the-loop), and the human's role moves from data-gathering to decision.

The honest metric is not "tasks automated"; it is "treasurer attention reallocated from data assembly to judgment." That is the number a CFO can defend to the board and a CRO can defend to the regulator.

04. Governance, audit, and SR 11-7 fit #

Agentic treasury is a model risk problem before it is a productivity story.

SR 11-7 and MRM. Under the Federal Reserve's SR 11-7 model risk management guidance, any model materially affecting financial decisions requires documented development, independent validation, and ongoing performance monitoring. A treasury co-pilot is a model under SR 11-7. MRM owns inventory entry, validation owns challenger testing (does the agent's forecast beat the regression baseline on a held-out window?), and production owns drift monitoring. Banks treating the co-pilot as "just tooling" are mis-classifying the risk.

DORA. Regulation (EU) 2022/2554 (DORA) Article 5 makes the board ultimately accountable for ICT risk. Treasury co-pilots are ICT systems supporting a critical function — the board must sign off on the risk framework, the third-party provider concentration, and the exit plan. The kill switch (revoke MCP tool access and roll back to human-only operation within minutes) is a DORA control, not a nice-to-have.

EU AI Act. A treasury co-pilot that influences material financial decisions falls under the high-risk classification, which obliges the bank to maintain a risk-management system, log all operations to OTLP-compliant telemetry, run human oversight, and produce conformity documentation on request. The realistic implementation is full OTLP traces on every agent reasoning step plus WORM-stored tool-call audit logs, and a human reviewer for any action that crosses a policy band.

Supervisory dialogue. The Bank of England (BoE) and the Financial Conduct Authority (FCA) have been explicit through 2025-2026 that they expect to see the inventory, the validation evidence, and the kill switch, in that order. The conversation goes well when the CRO can show all three in one room.

The control plane is the moat. A bank that can show its supervisor the MCP registry, the OPA policy file, the WORM audit log, the OTLP trace stream, and the SR 11-7 validation pack — in one sitting — is ready to run treasury co-pilots in production. A bank that cannot is running an unauthorised pilot.

Conclusion #

The Autonomous Treasury Index defined the destination: programmable liquidity, tokenised deposits, machine-readable policy. This piece is Part II — the production unit that gets a CIB treasury there. The pattern is stable: ISO 20022 data, MCP-bounded tool calls, SR 11-7 governance, DORA accountability, EU AI Act audit. The 2026 treasury co-pilots that survive their first supervisor review have the same shape; the ones that do not are missing the same three things — the MRM validation pack, the kill switch, and the WORM audit log.

The interesting work in 2026 is not the model. It is the control plane around the model, and the CFO conversation that turns 30-50% manual-workload reduction into a defensible cost-to-serve number.

For the upstream context, see the Autonomous Treasury Index and, for the governance frame, the Agentic AI Index for Banks.

Last reviewed 2026-06-29.

Republish this article

Agentic Treasury 2026: From Autonomous Treasury Index to Production-Grade Co-Pilots

Agentic treasury co-pilots are moving from pilots to production in 2026 — ISO 20022 data plus tool-calls, with SR 11-7, DORA, and EU AI Act controls around them.

This article is licensed under Creative Commons Attribution 4.0 International. Republication requires attribution to the canonical URL.

Agentic Treasury 2026: From Autonomous Treasury Index to Production-Grade Co-Pilots

Agentic treasury co-pilots are moving from pilots to production in 2026 — ISO 20022 data plus tool-calls, with SR 11-7, DORA, and EU AI Act controls around them.

Originally published at https://sebastienrousseau.com/2026-06-27-agentic-treasury-co-pilots-autonomous-treasury-production-2026/ by Sebastien Rousseau.
Licensed under CC-BY-4.0.

SEBASTIEN ROUSSEAU FOUNDER · ENGINEER

Agentic Treasury 2026: From Autonomous Treasury Index to Production-Grade Co-Pilots

01. From experiment to production #

02. The architecture: data + ISO + tool calls #

03. Use cases and metrics #

04. Governance, audit, and SR 11-7 fit #

Conclusion #

Format for Medium

Format for Mastodon

Copy formatted for LinkedIn

BibTeX

RIS

Vancouver

Chicago

APA

Republish this article