AI proposes by default. Where you grant autonomy, a person sets the limits and escalation is always one step away. Every step is logged.

This statement discloses how RMM Labs Ltd uses artificial-intelligence features inside Xcellerate OPS and on rmmlabs.io. It is published in fulfilment of the transparency duties applicable under Regulation (EU) 2024/1689 (the EU AI Act) and to ensure that Xcellerate OPS customers and their end users understand how AI is, and is not, used in respect of their data.

AI proposes. A named person decides — unless that person deliberately grants autonomy.

By default, Xcellerate OPS's AI reads, classifies, summarises, compares, drafts and proposes; it does not autonomously send a customer reply, apply a triage decision, route a ticket or accept a risk. Every action that produces a legal or similar significant effect routes to a named, authenticated technician, who applies, edits or dismisses it before it becomes the record of truth. There are two deliberate, disclosed exceptions, both off by default. The first is the optional AI Gatekeeper, which may reply to a requester only to ask for missing information; it makes no binding decision and is clearly, machine-readably marked as AI. The second is AI colleagues: virtual team members a workspace can hire and, capability by capability, allow to act on their own. Autonomy is never the starting state — the feature is hidden until switched on, a newly hired colleague can do nothing but search the knowledge base, and each capability is set to off, propose or autonomous one at a time. Even once granted, escalation to a human stays permanently on and cannot be switched off, and anything touching money, quotes or change documentation can never be made autonomous.

• AI never sends a customer reply automatically unless a workspace has deliberately hired an AI colleague and granted it that channel.
• AI never applies a triage or routing decision without a technician confirming it.
• Internal notes are never sent to the AI.
• Secrets-vault entries, IT documentation and cost or profitability figures are never included in AI prompts; an AI colleague can list credential entry names but never read their values.
• AI does not execute transactions or instructions outside Xcellerate OPS.
• The AI Gatekeeper only asks requesters for missing information; it never resolves a ticket, sends a solution, or makes a routing or priority decision.
• An AI colleague never touches administration — it cannot authenticate, is never an admin, team lead or approver, and can never change settings, users, teams, roles, permissions, integrations, billing or API keys.
• An AI colleague never allocates work to a human and never evaluates one; escalating returns a ticket to the queue rather than assigning a named person.
• Autonomy is never offered for money, quotes or ITIL change documentation; publishing knowledge articles, sending invoices or quotes, and approving changes are absent from what a colleague can do at all.

• AI features run on the platform operator’s configured providers (Anthropic Claude or an OpenAI-compatible endpoint) and, for translation, a dedicated translation provider.
• Provider API keys are stored encrypted and are never echoed back in the interface.
• Customer content is not used to train foundation models, and foundation-model parameters are not modified on the basis of customer content.
• Every AI action bills a transparent per-technician credit ledger, and technicians can be excluded from all AI assistance with a per-user toggle.
• Before any prompt leaves the platform, a deterministic, always-on layer replaces personal and secret data — email addresses, phone numbers, IP and MAC addresses, IBANs, card and VAT numbers, and API or bearer tokens — with stable placeholders that are restored in the answer. No AI is involved in this scrubbing, so it cannot leak.

• AI-generated triage suggestions, summaries and drafts may contain errors and must be reviewed.
• AI outputs are reproducible for a given input and model version; customers may request a logged copy of the relevant inputs and outputs in respect of a specific decision.
• AI is not used to detect minors or other special-category data subjects unless the customer explicitly enables such a feature and meets the relevant GDPR Art. 9 obligations.
• An autonomous AI colleague can make mistakes; its reach is bounded by a deny-closed capability gate, per-run step, credit and wall-clock caps, a workspace 'pause all' switch and the platform-wide kill-switch, and every run is recorded step by step for review and one-click recall.

RMM Labs Ltd does not make automated decisions producing legal effects on natural persons on the basis of Xcellerate OPS data. Where a customer uses Xcellerate OPS outputs in a context that itself produces a legal or similarly significant effect on a data subject, the customer is the controller and remains responsible for GDPR Art. 22 compliance.

Xcellerate OPS integrates third-party general-purpose AI models (Anthropic Claude or an OpenAI-compatible endpoint) into product features and places them on the EU market under our name. RMM Labs Ltd is therefore a provider of these AI systems under Regulation (EU) 2024/1689, as amended by the Digital Omnibus on AI (Regulation (EU) 2026/1744); our customers are deployers and their end users are affected persons. We are not a general-purpose AI model provider — those model obligations sit with the model makers.

• From 2 August 2026, the Act's transparency duties under Art. 50 apply: people are told when they interact with AI, and AI-generated content is disclosed. The optional AI Gatekeeper carries a fixed, human-written transparency disclaimer in all 25 supported languages, and every autonomous AI-colleague message is disclosed as AI.
• AI-generated messages that reach a person are machine-readably marked and visibly labelled as AI-generated, ahead of the 2 December 2026 deadline for machine-readable marking under Art. 50(2). Autonomous AI-colleague replies go out through the same send path as staff replies, so this marking applies to them automatically.
• No AI feature in Xcellerate OPS is currently classified as high-risk. Team routing stays outside the high-risk category because the person is chosen by a deterministic, rule-based strategy — never by AI profiling of individuals — and the QA check stays advisory and is never aggregated per technician. AI colleagues are limited-risk, not Annex III high-risk: by design and enforced in code, they never allocate work to a human or evaluate one, always offer escalation, are capped on money and change actions, and cannot reach administration; a risk-classification memo records why. The high-risk regime for stand-alone Annex III systems applies from 2 December 2027.
• For our deployer-customers, the in-product AI Act compliance center provides the AI registry, action log, oversight reporting, worker information, AI-literacy tracking and FRIA/DPIA material needed to meet their own duties under Arts. 4, 13, 14, 26 and 27.

Every AI action is recorded in a unified AI action log — the feature, provider and model, the initiating user or system, the related record, redaction counts by type, credits billed, the outcome and the human decision applied, dismissed or edited, each with a timestamp and request identifier. Prompt and response contents are never stored. Each AI colleague additionally keeps a per-run ledger — every tool it called, what came back, and every refusal with its reason — reviewable on the colleague's oversight page, where any work it holds can be recalled in one click. Logs are retained for a configurable period of at least six months and are exportable as CSV or JSON and via a read-only API, in addition to the schedule in the security statement.

Material changes to model usage or to the human-approval boundary are notified per the change-notice period on the trust center index page.