Managed service providers used to sell hands. The deal rested on trust in those people, backed by vetting, contracts and an audit once a year. That deal was right for work done by people. It does not survive agents. When the work is done by software acting under the provider's access, at a speed and breadth no person could match, trust in the people is no longer the control that matters.
Our position is that any provider running AI inside customer systems owes its customers a new baseline, and that the baseline is engineering, not contract language. Below are eight things a customer should be able to ask any provider for. Each one says what ORCA's managed service does, and what we will show you.
1. A binding per customer
The provider's agents reach a customer's systems only through access the customer granted, to an identity created for that customer alone. The customer sees every grant in its own admin screens and can remove any of them without asking. One service account or one API key serving every customer fails this test, however good the policy around it.
ORCA's managed service works through per-customer federated bindings. Each customer has its own connectors behind its own binding, and a request bound to one customer cannot reach another; a request that cannot prove which customer it belongs to is refused, not served. What our agents may call, and with which arguments, is fixed in a manifest the customer approves by digest. Ask us to show you: the grants in your own portal, the manifest version you approved, and a cross-customer request being refused.
2. Content is data, and hostile content cuts authority
A ticket, an email or a document can contain text written to steer an AI. A governed provider makes sure it never becomes the authority for an action, and enforces that in the system around the model, not in a prompt.
In our managed service, every tool result read from a customer's systems is labelled as customer content. When a read looks hostile, a per-customer taint window opens: for a set period, two hours by default, anything proposed for that customer's systems is raised to a human check, whatever the agent concludes. Ask us to show you: the label on a read, and the taint window that follows a planted instruction.
3. No change without the customer's approver
The provider's AI changes nothing in a customer's systems except inside a change request the customer's own approver authorised, under a deterministic policy, with a rollback path where the change can be reversed and a signed record tying the change to the approval. Governance is a precondition the write path reads: if the approval and the governance record for that class of change do not exist, the write is refused.
That is how ORCA's governed write path is built. For customers it is planned, and until it ships our managed service reads and does not write. A customer gets that boundary in writing. Ask us to show you: the read-only grants your binding holds today, and the approval design your approver will sign against when writes arrive.
4. A record the customer verifies without the provider
Every action the provider's software takes is written to an append-only, hash-chained ledger that names the actor, a person or an agent under its own workload identity, and is signed as it is written. Nobody, the provider included, edits or deletes it.
ORCA's ledger is live for every read and action in our managed service. A customer receives an export of its own chain that verifies from the first entry with our public key alone, without our software and without our help. Today we produce the export on request; customers exporting it themselves, and receiving signed chain heads as they are written, are planned. Until heads reach you as they are signed, a verified export proves the chain is intact and signed by us, not that we never re-signed it. Ask "who did this?" about any action and we answer within one working day, with the ledger reference. Ask us to show you: an export, and a verification from genesis on your own machine.
5. Residency stated per hop
The provider says where data is stored, where each model call is processed, which services process it and whether content is tokenised first, and tells the customer before any of that changes. It never rounds the answer forward.
ORCA stores customer data in Microsoft Azure UK South, in a partition set aside for that customer. Before our product sends content to a model, personal details are replaced with tokens; that is pseudonymisation, so we still treat the content as personal data. Models run through Microsoft Azure AI Foundry, with our resource in Sweden Central and the processing location set by the deployment type. Any other route is used only where the customer's terms name it and the customer agreed. We do not train or fine-tune models on customer data. Ask us to show you: the data-flow record for your binding, generated from the same manifest the system runs on.
6. Disclosure when AI did the work
A customer never has to discover that AI was involved. ORCA tells a customer in writing, before the work starts, that we deliver with AI. Our deliverables carry a line saying they were prepared with AI assistance and who reviewed them. The ledger names the agent, and an agent asked whether it is an AI says yes. Ask us to show you: the line on a deliverable and the agent's name on the record behind it.
7. A written rule set, and a person accountable
A provider should write down what its AI may and may not do with customer data and hold every person and every agent to it. Ours is a handbook that sets out, rule by rule, what our AI may read, where content may go and be stored, what is never allowed, and which rules are enforced in code. A named ORCA director reviews and approves everything delivered to a customer, and is accountable for it. Ask us to show you: the rules that apply to your data.
8. Exit with data and evidence intact
A customer who can leave easily can say yes easily. Exit runs in a fixed order: export first, then access ended, credentials and stored tokens purged, and each removal read back. The customer receives a final signed record, and the ledger is kept for its agreed retention period, because evidence that can be deleted on request is not evidence.
ORCA's offboarding produces the verifiable export, purges the customer's credentials, retires the customer's manifest and writes a final signed event. Every grant can be removed by the customer from its own screens at any time, and we hand over the checklist of what it can remove. Ask us to show you: the offboarding sequence before you sign.
What is live, and what is next
- Live in our managed service: per-customer bindings and the approved manifest; labelled tool results and taint windows; the signed ledger for every read and action; the verifiable export, on request; UK data at rest and tokenisation before model calls; disclosure and the named accountable director; the written rule set; the offboarding export, credential purge and final signed record.
- Planned: writes to customer systems behind the customer's approver; customers exporting the ledger themselves; signed chain heads delivered to the customer as they are written. Until writes ship, the service reads and does not write, so nothing waits on an approval that does not yet exist.
The baseline is the product
Per-customer identities are slower to build than one key, and an approval in front of a write is slower than letting the agent act. That is the point. The providers that meet this baseline are the ones customers can keep saying yes to as their AI does more. The rest will meet their customers' boards asking the questions in our supplier questionnaire.
If you are buying a managed service, ask us for any of the eight above and we will show you. If you run a managed service or a consultancy and want your own service to meet this standard, we design and run the bindings, the taint and approval gates, the ledger and the exit, and we can help you build yours. Talk to us.