An AI employee for every department.

Nine departments, one shared operator data layer

Start with one human leader per department and add specialized AI team members as the operation grows. Select a department to see its agents.

AI assistants your team uses every day

Each assistant answers a request and hands the result to a person to review and save.

Five execution levels, set per agent and per action

Every agent starts in shadow mode. An action moves up a level only with evidence from shadow mode and the sign-off of its human owner.

1. Level 1 · AdvisoryRead data and produce analysis, answers, alerts and recommendations. Creates nothing on the platform.
2. Level 2 · DraftingCreate inert drafts and internal items: campaigns and journeys in draft status, inactive bonus templates, unpublished page versions, internal notes, reply drafts, documents, and handoffs or tasks addressed to a named human. Nothing reaches a player, partner or the public, and no setting changes, until a human publishes, sends or activates it.
3. Level 3 · Approval requiredStage one concrete platform action as a proposal carrying the full payload. An authorized human approves or rejects it; on approval the action runs through the same validators as a human action. Where the platform applies a two-person rule, the approver must be a different person from the requester. A proposal whose target changed after it was made is stale and must be proposed again (the same check the risk-rule proposal queue uses).
4. Level 4 · Limited autonomous executionExecute allow-listed, low-risk, reversible or idempotent actions within hard caps (volume, value, frequency), with a kill switch, immediate notification of the human owner and after-the-fact review. Unlocked per action only after shadow-mode evidence and sign-off by the human owner and the tenant super admin. Never used for money movement, verification, AML, responsible gaming, legal, pricing or access decisions.
5. Level 5 · RestrictedNothing beyond preparing evidence and summaries for the authorized human. The agent cannot initiate, execute or approve the action, and the platform route refuses agent tokens (deny by default, the risk desk's pattern).

Decisions no agent makes

Agents prepare evidence and drafts, and an authorized person decides. Agents assist the MLRO, compliance officer, legal counsel and responsible-gaming professionals; they never replace them.

Agents working together, with a human decision at the end

Agents from different departments hand work to each other through the shared data layer, and every workflow ends with an authorized person.

Guardrails that apply to every agent

• Tenant isolation. Each agent instance belongs to one tenant. Its token is pinned to that tenant and to the brands it is granted; services refuse cross-tenant requests. No shared memory, prompts or fine-tuning across tenants. No agent ever accesses another tenant's data.
• Least privilege, deny by default. Every agent has its own token class with a per-route policy (generalizing the risk desk's route policy): reads are listed, writes are listed per agent, and routes for restricted decisions refuse every agent token.
• Humans decide restricted matters. Decisions R1 to R15 are human decisions. Agents assist the MLRO, compliance officer, counsel and RG professional; they never replace them.
• Player protection first. Never contact self-excluded, cooling-off, RG-flagged or opted-out players; never target harm markers with offers; never encourage play, chasing losses or borrowing. Platform suppression is enforced and re-checked by the agent.
• Money safety. Agents move no money. Value is issued only through existing idempotent engines, inside approved caps, after approval. No payouts, withdrawal decisions or balance adjustments by an agent.
• Data minimization. Pseudonymous player ids by default; direct identifiers only when the task needs them; never send personal data to tools or endpoints outside the approved list; redact personal data in logs and summaries.
• Truthfulness. Every number cites the query or record it came from; unknown means 'unknown', never an estimate presented as fact (the SEO helper already refuses to invent search volumes).
• Audit everything. Every run, tool call, personal-data read, draft, proposal and action is logged (see baseline_audit_requirements).
• Kill switch and ladder. Off / shadow / suggest / auto per agent and per action; the tenant super admin and the human owner can switch an agent off instantly.
• Budget caps. Hard monthly token and per-run caps per agent; at the cap the agent stops or falls back to shadow and notifies its owner.
• No unapproved external communication. Nothing goes to players, partners, vendors, regulators, investors or the public without approval, unless that exact intent is on the agent's level-4 allow-list.
• Market rules. Respect each brand's restricted countries, age rules, advertising rules and license conditions; when unclear, stop and escalate.
• No self-modification. Agents cannot change their own permissions, level, budget, prompts, tools or guardrails, or those of another agent.
• Model provider terms. Provider-agnostic model adapter; zero data retention and no training on tenant data; pinned model versions per agent release.

Your data stays yours

Pay for the agents you use, add more as you grow

Your package sets how many agents work for you. The more agents you add, the higher the package.