Agent-first service management
Your agents propose. Your people decide.
A service desk and CMDB where AI agents do the work and a human countersigns it — and where the audit trail afterwards shows which agent proposed what, who signed, and how long the control took. The configuration question is not what can the agent do. It is where do you want a human in the loop.
// nothing moves until somebody confirms NC-000148 Access reviews not performed for Q3 identified → containment Three prior findings name the same control owner, who left in August. The quarterly reminder was never reassigned, so I propose recording containment and freezing privileged accounts pending review. proposed_by: 'nonconformity-analyst' confidence: 0.82 awaiting: countersignature
Nothing moves until somebody confirms — no field, no SLA clock, no notification.
Three rules hold everywhere in the platform. They are not settings that can be switched off, and they are what make an agent’s work auditable rather than merely fast.
It acts for a named identity and reaches exactly as far as that person does. “What can this agent do” is answered by looking at a person — not at a second permission system that has to agree with the first.
It is recorded before the transition’s actions run. Until somebody confirms, nothing has moved: not a field, not an SLA clock, not a notification. Refusing costs nothing, which is what makes review real.
A countersignature re-checks the signer’s own permission and re-derives the transition from where the ticket is now. Nobody can sign an agent into doing something they could not have done themselves.
Nonconformity and corrective action ships as a working lifecycle, not a template. Each state carries the field the clause asks for, and the state machine refuses to leave without it.
Raised from an audit finding, an incident, a complaint or a monitoring result.
source · clause reference · severity
React to it: control and correct it, deal with the consequences.
cannot leave without: immediate_correction
Review it and determine the causes — and whether it could occur elsewhere.
the agent correlates against prior findings
Implement what is needed to stop it recurring.
cannot leave without: root_cause + similar_occurrences
Check whether the corrective action actually worked. Verification can send it back.
cannot leave without: corrective_action
Closed, with the review of effectiveness on record.
cannot close without: effectiveness_review
Closing without recording whether the fix worked is the most common finding written against clause 10.2. Here it is not discouraged by a process document — it is refused by the state machine. That single rule is most of what this workflow is for.
Every figure comes from records written the moment the work happened. Nothing is recorded for the report — which is exactly why the numbers can be quoted. Pick a control and a period; get the instances, who acted, what the agent proposed, who countersigned it, and how long each took, as a screen and as a file you keep.
For this control, over this period: every instance, its full transition trail, and the agent rationale a human put their name to.
One row per transition, denormalised, named for the control and period — so it still says what it is a year later in an audit folder.
A queue nobody ever rejects anything from looks identical to a queue nobody reads. Only one of those is a working control.
evidence.read grants “see the record, touch nothing” — so an auditor never needs the working queue or the ability to edit the control they are testing.
States, stimuli, transitions and actions are rows you edit at runtime. Incidents, requests, problems, changes, known errors and nonconformities.
Configuration items and typed relationships, with recursive impact traversal — what breaks if this does.
change.approve is its own permission, and “implemented” is unreachable except through “approved”.
Business calendars, pause-on-hold states, a scheduler that warns before breach and escalates on it.
Postgres row-level security under the service layer, so a forgotten filter returns nothing rather than someone else’s rows.
Sign-ins, denials, administrative changes and every agent decision — written on their own transaction so a failing request still records what it tried.
End users raise and follow their own requests and see only what they should — separate from the agent console.
Model credentials are per-organisation and stored by reference, never as a key. Point it at a model you host if ticket content cannot leave your estate.
Same code, same release, same tests — the difference is configuration and where it runs, never a fork. Self-hosted means your ticket data and your prompts never leave your environment, which for a regulated buyer is a substantive answer to a question they are obliged to ask.
SOX is deliberately self-hosted only. If we hold your SOX-relevant ITGCs, this platform sits inside your ITGC scope and your auditor needs assurance over us — a SOC 1 engagement, with us as a subservice organisation in your audit. Self-hosted has none of that: you run it, the controls are yours, and we are a software vendor rather than a service organisation. It is a scoping decision, not a smaller product.
Everything marked shipping is built, tested and running today. Everything marked next is planned and specified, and is not in the product yet. We would rather you knew which was which before a procurement review than after one.
Nonconformity and corrective action, with mandatory-field gates and evidence reporting.
Per-organisation model allowlists, confirmation floors, chain-depth ceilings and a rolling spend cap.
Where an organisation requires two signatures, one person cannot supply both.
OIDC with per-customer app registration, and MFA verified from your identity provider's assertion rather than duplicated here.
Campaigns where a revocation actually revokes — the control SOX and SOC 2 CC6 both ask for.
A conflict matrix evaluated at grant time, with recorded overrides rather than silent exceptions.
Security incident management, vulnerability remediation, supplier review.
Because single sign-on alone does not close the window between removing somebody and their session ending.
The full order, and an explicit list of what is not coming at all, is on the roadmap.
This platform ships workflows that support controls. It does not make anybody compliant, certified or conformant, and nothing in the product says it does. Your conformance depends on your scope, your Statement of Applicability, your risk assessment and your evidence — none of which a vendor owns or can see. A SOC 2 report is an auditor’s opinion about a system description you wrote; SOX effectiveness is management’s assertion. Neither is a property of software, and any tool telling you otherwise is making a statement it has no standing to make.
We hold ourselves to it in code: an automated test fails our build if the shipped description of a control workflow ever acquires the words “compliant”, “certified”, “conformant” or “guarantees”. The line is “encodes the control” — never “makes you compliant”.
We are onboarding a small number of organisations at a time so each one is configured properly — lifecycles, service levels, model credentials and who approves what — rather than handed an empty tenant. Tell us about your desk and we will come back with a date and what setup involves.
Enterprise licensing is self-hosted in your environment, with the licence, support terms and deployment help a regulated estate needs. Choose it below and we will answer on that basis.