Jump to another pillar
Vertical agentic orchestration · treasury
An outside read of the category: the distinction most buyers cannot yet make, and which recurring treasury decisions are genuine candidates for judgement rather than a threshold.
Part one
| Chat | Copilot | Agent | |
|---|---|---|---|
| What it is | Another way to ask questions of the same data | Chat inside the workspace, aware of the screen, drafting the next step | Has a goal, runs on a schedule, produces the output without being asked |
| Who starts it | The human, every time | The human, every time | Nobody. A trigger does |
| Where the work sits | Entirely with the human | Mostly with the human | With the system, inside rules the customer sets |
| What it removes | A query here and there | Friction from work the human is still doing | The work |
Treasury differs from every other orchestration domain in one respect that governs the whole design: the actions are largely irreversible and priced in cash. A released payment is gone, an executed sweep has a value date, and the sign-off chain runs outside treasury entirely — CFO, external auditor, sometimes an audit committee. So how far the agent is allowed to act on its own is a control rather than a preference, and the system’s ability to decline to act and say why matters more than its ability to act.
Three layers, and only one of them is agentic
Bank connections, overnight statement ingestion, multi-entity and multi-currency consolidation, reconciliation, the dashboard. Considerable engineering, and largely finished.
Test — is the answer a fact about what happened? Then it is data, not judgement.
Alerting when a balance drops below a buffer or an exposure passes a limit. Running a stated scenario and reporting the impact by week and entity. Where the rule can be written down, a rules engine applies it perfectly.
Test — can it be expressed as if this, then flag that? Then it is a monitor. Do not reason about it.
A monitor tells you the buffer is broken. It does not tell you whether to sweep, draw the facility, defer a payment run or do nothing — because those are all policy-compliant, they cost different amounts, and the right one depends on cut-offs, tax, forecast confidence and who you would rather not call.
Test — do two good treasurers disagree, and can both defend the answer? Only then is there something to teach.
Part two
The daily position and how to fund it
Overnight statements land and one entity sits below its minimum operating balance.
Sweep from an approved intercompany pair · draw the facility · defer a discretionary payment run within supplier terms · do nothing, it is inside tolerance
Group surplus exceeds what the policy requires held.
Place it at the horizon the forecast confidence supports · pay down the drawn facility · hold it because a known outflow lands inside the term
A statement is missing, or unreconciled items sit above the threshold.
Refuse to run and say which entities are missing · run on partial data and flag it · wait for the feed. The first is the right answer, and the hardest to demo.
The forecast and reality diverged
The gap needs decomposing to driver, entity and counterparty.
Treat it as a timing shift · treat it as a structural change and propose a named assumption change · flag it as noise and leave the model alone
Collections behaviour has moved, not just slipped once.
Re-forecast on observed behaviour · escalate to collections · adjust the terms at renewal · absorb it in the buffer
A stress set is needed: collections slippage, largest-customer delay, a rate move on the revolver.
Which scenarios, at what severity, over what horizon — and which ones drawn from the group’s own history that nobody thought to ask for
Policy is breached, or about to be
Concentration passes the per-institution limit in the policy.
Move it and accept the cost · request a temporary limit waiver · net it against a facility at the same institution
Net exposure moves past the point the policy says to act.
Hedge now at the quoted rate · hedge a portion · net internally first · escalate, because the policy is ambiguous here
A payment or a funding action has aged past the policy window with nobody acting.
Escalate to the next approver · release under a standing delegation · let it lapse and re-raise · call the person
Part three
Three questions an auditor asks
The delegation-of-authority matrix, extended to the agent. It names the entities, data and actions the agent may touch. It writes down what is prohibited instead of leaving it out. It also sets approval tiers by amount and separates the person who configures a run from the person who approves it. The situation card says how far a strategy could be delegated. The framework says how far this customer has delegated it. The agent acts on the lower of the two.
Test — can you point to the clause that permitted the action, in the version in force that day?
A dashboard shows what happened. An audit needs a linked record of the whole decision. That record holds the data and its as-of time, the options weighed and why each lost, and the recommendation as the approver saw it. It also holds the approver and their reason, and the bank’s confirmation. The record is append-only and kept for the statutory period, by the customer, not by an observability vendor’s retention window.
Test — could an auditor rebuild the decision from the record without asking anyone?
When a treasurer changes an amount or rejects a recommendation, that is audit evidence. It is also the most useful signal the system receives. Code checks run on every decision. A model-based reviewer samples what code cannot check. Each corrected case joins the set the next prompt or model must pass before it ships.
Test — does a correction made today change what the agent is tested on tomorrow?
How ATO answers them — typed steps in a graph
Each decision runs as a fixed graph of typed steps: a guardrail, a calculation, a tool call, a model call, a human approval. A model chooses at one step only. The sequence, the arithmetic and the authority checks are code. Each step writes one structured record, so the reasoning is kept as data rather than as a transcript.
Qualify
Which situation, and is it judgement or already solved
Data gate
Every balance fresh, every entity present
Options
Each strategy’s preconditions, met or not, and its amounts
Select
One strategy, and why each other lost. No figure of its own
Authority
Lower of card and framework ceiling; approvers by amount
Stage
Instruction assembled and validated, awaiting release
Approve
The agent prepares; a person releases
Release
Written to the treasury system as approved
Confirm
Bank booking matched to the approved payload
■ not in the library, or not detectable in this stack
■ stale or partial data: refuse and name the entity
—
—
■ outside scope, or recommend only: escalate
■ a person already moved the cash: halt
■ approver configured the run: reject
—
■ booking differs from approval: flag
Decision ledger append-only · each row hashed with the one before · each row names the step that wrote it
SITUATION_
DATA_
(on a stop)
—
RECOMMENDATION_
AUTHZ_
INSTRUCTION_
APPROVAL_
ACTION_
EXECUTION_