A bank that puts agents into lending, fraud and customer service has two things to show. The calls those agents make stay inside the rules the bank answers to. Each decision a process reached can be taken apart afterwards. Gate holds the first at the call. Flow keeps the second as a replayable case record.
At the call: Gate and the banking packs
A control pack is a directory of rules and mappings that kirto-gate pack add copies into a configuration. Each rule cites the control identifiers it is for, and every decision it makes carries them onto the signed audit record. The packs that exist today are Saudi, and the one written for banks is SAMA’s.
The SAMA Cyber Security Framework pack reads subdomain 3.4.3, Cloud Computing. A hosted model is a cloud service, so a hybrid or public provider is refused unless it declares SAMA’s approval, and one outside Saudi Arabia is refused unless it declares the explicit approval for that. Four more considerations are mapped to the mechanism and the evidence: review and audit rights over the provider, asset classification, the use of approved cryptography, and protected, centrally analysed security logs.
A Member Organization also falls under the Personal Data Protection Law. The PDPL pack replaces national ids, contact details, IBANs and card numbers whose checksum held before a request leaves, and records every request that carried personal data against the disclosure, transfer and record-keeping articles.
| Pack | Controls | Rules | Mapped |
|---|---|---|---|
| SAMA Cyber Security Framework | 6 | 2 | 4 |
| Saudi Personal Data Protection Law | 5 | 4 | 0 |
| NDMO data classification | 10 | 4 | 3 |
| SDAIA AI ethics and generative AI guidelines | 6 | 1 | 5 |
A mapped control is one no rule can enforce. The pack names the mechanism that addresses it and the evidence it leaves, and attests nothing. kirto-gate controls lists each control a configuration cites with the rules that name it, and flags the ones nothing enforces. Start from a control pack walks through adding one.
In the process: Flow
Credit decisioning. The underwriting flow is a BPMN process. Its eligibility rules are a DMN decision evaluated inline by a business rule task. An agent step that drafts the assessment is one task among the others. Each case records the process name and version it ran, every input that arrived and the agent step’s output, so a decision months old replays against the definition that made it.
Fraud investigation. An investigation is a CMMN case. Evidence arrives as events in any order, and the case plan decides which task becomes available next. An analyst’s approval of an escalation is a human task, and the decision is a human.gate row on the case.
Approval of an agent’s actions. An agent step can name tool namespaces whose calls need approval, and the approver groups that may give it. The approval comes after the tool runs and decides whether the loop continues. An action that must not happen without a person’s yes belongs in a human task placed before it in the process.
Model risk
The reviewer’s questions have their own page.
What SR 11-7 and SR 26-2, which replaced it in April 2026, ask of a model in production, and how a Flow replay and the Gate record answer each ask, is set out on Model risk.
The evidence
Gate’s audit trail is a hash chain per node, signed with the node’s ed25519 key, and shipped to the bank’s SIEM. kirto-gate audit verify checks an export without trusting the node that wrote it. See Ship the audit trail to your SIEM. Flow’s case record is read from the case history API, and kirto-flow-replay exits 0 when the recorded chain re-derives clean. See the Flow docs.
What this does not do
- Packs outside Saudi Arabia. The four packs are Saudi. No pack exists for PCI DSS, DORA or a US banking rule, and a bank writes its own in the pack format.
- Certification. A pack is a starting point and certifies nothing. Its control titles are paraphrases to check against the text.
- The approval itself. The SAMA pack reads that a provider declares an approval. It does not check that the approval exists, is current or covers the service.
- Sanctions, KYC and fraud scoring. Neither product screens names or scores transactions. A process calls the system that does, and records its result.
- Model validation. Neither product validates or back-tests a model. The model risk page lists what they leave to the validator.