What a buyer can verify without sending an email

The services are contracted through the group operating entity incorporated in Ireland, with local service schedules in the United States, the United Kingdom, Germany, France, Brazil, Singapore and Australia. The entity list and the schedule that applies to each country sits on the residency page.

Ninety days either side. Service continues unchanged for the whole notice period; the estate is frozen rather than degraded, and the exit plan starts on the day notice is received rather than at the next renewal.

Yes to the first, not confirmed on the second. A $25 million cyber liability policy is in force to March 2027. Whether an incorrect decision taken autonomously by an agent falls inside the policy has been raised with the broker and has not been answered in writing. Treat agentic acts as uncovered until it is.

Yes. Type II covering security, availability and confidentiality for the twelve months to 31 January 2026, issued by an independent firm. The report is released under non-disclosure on request; the scope, exclusions and opinion are stated on the trust center without one.

The certificate is current to October 2027. The annual surveillance audit was due on 4 May 2026 and has not been held; it is 46 days overdue and rescheduled for 28 July. A certificate stays valid through a late surveillance audit up to a limit, and we are inside that limit, but the honest position is that the certification is current and its supervision is late.

An independent test of the cockpit application and its APIs completed on 12 March 2026. The agent runtime was explicitly out of its scope. A second test covering the runtime, prompt handling and tool invocation is under way, and it is being run by our own red team rather than an independent firm, so it will not carry independent assurance when it reports.

Twenty-four adversarial attempts have been run against the estate across fifteen techniques, producing nineteen findings. Every attempt, what it reached and whether the finding is closed is published on the adversarial page, including the attempts that succeeded.

Not currently. The Cyber Essentials Plus certificate expired on 28 February 2026 and has not been renewed. Recertification is scheduled but not booked. If this is a mandatory requirement of your process we do not meet it today.

No, and it is out of scope. The estate does not store, process or transmit cardholder data at any station. If your process requires a PCI attestation regardless of scope, we cannot provide one.

Identity is federated through a single provider with enforced multi-factor authentication. Agents hold service credentials scoped to the authority granted to them, and 309 authority grants are enumerated with the value ceiling and second-signature rule attached to each. The duties page shows every pair of duties that must not be held together and where the estate currently breaches that.

For the hire-to-retire tower, yes — completed and signed on 19 February 2026. For order-to-cash the assessment is still in progress while the agents in that tower are already running in production. That sequence is the wrong way round and it is recorded as such rather than presented as work in flight.

The full sub-processor register is published on the trust center with the purpose, the record classes each one touches, the processing locations, the data processing agreement reference and the last assessment outcome. Thirty days of written notice before any addition, with a right to object.

Processing locations are published per record class on the residency page. Transfers out of the EU rely on standard contractual clauses supported by a transfer impact assessment, and that assessment expires in 40 days. Two sub-processors added since it was written are not covered by it.

No. Every provider endpoint in the registry is contracted with training excluded and zero retention. The self-hosted models never leave our tenancy at all, and special-category personal data is routed only to those. Each entry in the model registry states its data boundary and its training position individually.

No. It has not been pursued and no timetable exists. The privacy programme runs against the data protection regulation directly rather than against a certified management system.

Retention is set per record class and published. Deletion propagates through the operational stores. It does not propagate through the nightly extracts written to our own object storage, which currently have no retention policy configured, so a satisfied request today can leave a copy behind in an export.

Not yet. ISO/IEC 42001 is at stage 2 audit, booked for 14 September 2026. Until that audit passes there is no independent assurance over the AI management system, and no amount of internal documentation substitutes for it.

Twelve models are registered. Each carries its provider, version, modality, hosting location, data boundary, training position, contract type, the towers and agent count that call it, its lifecycle status and its named substitute. Three are self-hosted open weights, two are retired and kept visible, one is under a published deprecation notice.

Every model has a named substitute except three. Eight of the nine named substitutions have been tested against the golden sets; one has not. The untested one is the embedding model, and replacing it means re-embedding the entire retrieval corpus — work that has never been scoped or costed. One model is under a hard end-of-life date 135 days away with 168 agents still calling it and no named owner for the cutover.

No. Every evaluation in the estate runs against golden sets written by the team that built the agents. A passing suite shows an agent behaves the way we specified; it does not show that the specification is right. No model risk management review has been performed by anybody outside that team.

Agents move through five stages against evidence gates, and demotion is armed automatically when a gate stops holding. The full stage history for every managed agent, including every promotion and every demotion with the evaluation run behind it, is published. 178 agents are under lifecycle management and 1,183 are not, which is stated on the page rather than left to be discovered.

Six months of monthly availability for four components against their targets, with incident counts and major-outage minutes. It is self-reported from our own monitoring. There is no independent monitor and no third party has verified any of these figures.

The continuity exercise aligned to ISO 22301 ran on 12 May 2026 across three scenarios — model provider outage, evidence store failure and full region loss. Separately, sixteen recovery drills are scheduled and eleven have been run; two of those missed their recovery objective and three had no independent observer. Every drill, including the two misses, is published.

Six of every ten inference dollars go to one model provider. Three of the five workloads with that provider have a tested route to a second provider. The embedding workload does not, and that is the concentration that matters rather than the headline share.

Fourteen asset classes in CSV, JSON, JSONL, Parquet, Markdown and PDF — all open, none proprietary. Four have a live route or a nightly job; ten are produced by hand from a database query. The full extract is planned at fourteen days and that assumes the hand-produced ones go right first time.

Two of fourteen asset classes have been read outside the product — the evidence packs and the trust register, both through live routes an auditor has used. The other twelve have never been imported into anything. No parallel run and no restore verification has ever been performed. This is the weakest point in the exit plan and the plan depends on it.

The run-book is generated from the same tables the product reads, so it cannot go stale, and it has never been handed to anyone outside the team that wrote it. Deletion produces a signed certificate listing every store; backups age out over 35 days rather than being deleted immediately, and the certificate says so.

Reported weekly for the estate, for each of the twelve towers and for each of the twelve models, with a five-week trend. It is derived from call volume by arithmetic, not metered. No provider in the registry discloses metered energy per request, so the ranking between towers is meaningful and the absolute kilowatt hours are an estimate.

No. None of the three model providers exposes per-request energy, and the self-hosted models run on shared accelerators where consumption is not attributed per tenant. Anyone claiming a metered per-transaction figure for hosted inference today is estimating it the same way we are and calling it something else.

Eighteen employee and customer measures are published, each paired with the throughput measure it sits beside. Seven of them have moved the wrong way while throughput improved, including the share of customers who reached a person when they asked, which fell from 69 percent to 44.2.