The problem · in depth
AI agents are powerful. Ungoverned, they’re a liability.
Agentic AI is crossing from suggesting to executing. Most agent stacks have no layer between an agent’s decision and the real world — and no way to prove what happened afterward. Three structural gaps make ungoverned agents a liability.
Problem 01
Actions without accountability
Agents have crossed from suggesting to executing — and execution is where liability begins.
For years, AI sat behind a human. It drafted, recommended, and predicted, but a person clicked the button. Agentic systems remove that person: they now issue refunds, move funds, change customer records, revoke access, and trigger downstream systems on their own.
The instant an action touches the real world it becomes accountable — to a customer, a regulator, or a court. Yet most agent stacks have no layer between the decision and execution. There is no checkpoint that can say “this one needs a human,” and no record that can later explain who, or what, authorised it.
The failure mode is rarely dramatic. It is a single over-broad action — a bulk update applied to the wrong segment, a deletion run at scale, a payment to an unverified account — that becomes a reportable event because nothing stopped it and nothing recorded it.
- Agents act directly on money, records, and customers.
- No enforced checkpoint between an agent’s decision and execution.
- No attributable record of who or what approved each action.
How Crelis closes it
Crelis sits in front of execution. Each consequential action you route through it is evaluated against policy before it runs, and the decision — allow, require human approval, escalate, or block — is sealed in a tamper-evident record. Allowed actions receive a short-lived, cryptographically signed runtime execution visa; the rest wait for an authorized reviewer or do not run.
Problem 02
Confidence is not certainty
A model acts with the same authority at 99% confidence and at 60%.
Models emit a confidence score, but they act with identical authority whether that score is high or low. A 60%-confident decision to wire a large sum executes exactly like a 99%-confident password reset. The system has no notion of stakes.
Risk is not just probability — it is probability multiplied by consequence. A confident action on a trivial task is safe; a slightly-less-confident action on an irreversible, high-value, or regulated task is not. Nothing in a raw model pipeline weighs that consequence or knows when to stop and ask.
The result is a flat risk surface: each action is treated as equally safe to automate. The rare, expensive mistakes hide inside the long tail of low-certainty, high-stakes decisions that never should have run unattended.
- Confidence scores don’t map to real-world consequence.
- High-stakes, low-certainty actions execute like routine ones.
- No threshold that escalates when the cost of being wrong is high.
How Crelis closes it
Crelis evaluates each consequential action deterministically — task type, proposed action, amount, industry and channel, customer tier, verifications — against your policy. When the stakes exceed what policy allows, the outcome is require human approval and the execution visa is withheld until an authorized reviewer signs off; an unrecognised action escalates to a human instead of executing.
Problem 03
Compliance can’t see inside
Risk and audit teams are asked to approve AI programs they cannot inspect.
Compliance, risk, and audit functions are accountable for AI behaviour, yet the systems they must sign off on are opaque. They cannot inspect why an action was allowed, replay how a decision was reached, or attribute an outcome to a specific policy and approver.
Under frameworks like the EU AI Act, MAS guidelines, and internal model-risk policies, “the model decided” is not a defensible answer. Auditors ask for evidence: what was requested, what governed it, what the risk was, who approved it, and proof the record hasn’t been altered.
Without that evidence trail, governance becomes a matter of trust rather than proof — and trust does not survive an audit, an incident review, or a regulator’s request for the decision history.
- No way to inspect, replay, or attribute AI decisions.
- “The model decided” fails EU AI Act / MAS / model-risk scrutiny.
- Governance rests on assurances instead of evidence.
How Crelis closes it
Crelis is the system of record for each consequential action you route through it. Each decision — its inputs, the policy that governed it, the recorded risk context, the visa issued or withheld, and the authorized reviewer who approved it — is written to a hash-chained, tamper-evident audit trail your compliance team can hand to auditors and regulators.
This is the gap Crelis governs
An independent authority layer between AI and the real world — evaluating each consequential action, applying policy, issuing the execution visa, and recording the evidence.