Governed STP savingsSTP explainedWhy MidshipsWhy governedOther toolsKeycloak savingsRenewal plannerAI risk readiness
Governed STP

Why not use an ungoverned agent?

An agent with access to your systems can do the work. The question is whether you can let it finish the work without a person checking. Two things decide that: whether you can trust it, and what your regulator expects.

Trustwithout bounded authority, a person reviews every case
Savinglimited to cheaper reviews; the straight through rate does not move
Regulationoversight and records an ungoverned agent cannot evidence
1. Trust

If you cannot trust it, a person still has to sign off.

An ungoverned agent has permission: access to your systems. It does not have authority: a defined purpose, limits and someone who answers for what it does. Nothing checks an action against your policy before it runs, so the organisation cannot accept its outcomes unchecked. Someone has to be accountable, which means a person reviews every case.

Ungoverned agent
Governed agent
What it has
Access: permission to call your systems.
An appointment: purpose, limits, an approver and a named owner.
Before each action
Nothing checks that the action serves its approved purpose, within limits.
Every action is checked against the authority delegated to it. Anything beyond goes to a person.
Afterwards
Logs show what happened, not why or under whose authority.
Every decision is recorded with its reasoning, the policy applied and the person accountable.
What a person does
Reviews every case before the outcome stands, because someone has to answer for it.
Decides only what your policy routes to them, with the case already prepared.
Straight through rate
Unchanged. No case completes without a person.
Rises. Cases within authority complete with no person involved.
2. What it is worth

The saving stops at cheaper reviews.

Take the calculator’s example: 100,000 cases a month, 50% straight through today, $100 to handle an exception by hand, and a review of an agent-prepared case at half that cost.

Ungoverned agent
Governed agent
The 50,000 exceptions a month
The agent prepares all of them. A person reviews all of them.
25,000 complete within the agent’s authority. A person decides the other 25,000, already prepared.
Straight through rate
50%, as today.
75%.
Saving a year, against all manual
About $30 million, all from cheaper reviews.
About $45 million, from work that no longer needs a person and cheaper reviews of the rest.
Liability
Stays with you, and depends on every review catching every problem.
Contained: actions outside authority are stopped before they run, and every decision has a named person accountable.

The gap is about $15 million a year on this one process. That is the value of authority, before counting the regulatory exposure below. Illustrative figures, before the cost of the solution and not a quotation. Run it on your own numbers →

3. Regulation

Regulators expect bounded authority, oversight and records.

The regimes we track converge on the same core: know your AI, bound its authority, keep a person accountable and record every decision. An ungoverned agent cannot show the last three, so the institution answers for whatever it does.

Regime
What it expects
If you are found wanting
EU AI Act
For high-risk uses, including credit scoring of individuals and life and health insurance pricing: competent human oversight and automatic logs, kept by deployers for at least six months. Applies from 2 December 2027.
Fines of up to EUR 15 million or 3% of worldwide annual turnover, whichever is higher, for most breaches.
MAS Guidelines on AI Risk Management (Singapore)
Human oversight proportionate to risk across the AI lifecycle; the institution stays accountable, including for third-party AI. Sections 3 and 4 apply from 7 October 2027.
Supervisory action.
APRA (Australia)
Controls over agentic and autonomous workflows, strong privileged access management and human involvement in high-risk decisions. Applies now, within existing prudential standards.
Supervisory action and, where appropriate, enforcement.

Not every exception queue is a high-risk use under the EU AI Act. The supervisors above still expect these controls wherever AI acts on customers or money, and a fine is only part of the cost: remediation programmes and supervisory restrictions follow findings too. The four regimes compared in full →

Sources: MAS media release, 7 Oct 2026 · European Commission AI Act page · European Parliament on the Digital Omnibus · AI Act Article 99 · AI Act Annex III · APRA release, 30 Apr 2026. Not legal advice.

Give your agents authority, not just access.

See what governed agents are worth on your process, or talk to us about the queue that hurts most. Icebreaker, our runtime governance engine, is patent pending.

Open the calculatorScope a proof of value