It's an AI — so how is the confidentiality of our information guaranteed?
It's the first question every customer asks, and the answer is architectural: identity, permission and audit on every interaction — including Kris's. This page covers where your data lives, what leaves, what doesn't, and what stays on record.
The short answer
Kris is not a chat with access to everything. It acts inside the platform the way a user does: under the identity of whoever asked, limited to that person's permissions, and with every step recorded in a history that cannot be edited.
In practice: Kris cannot see what you cannot see. If an analyst has no access to the legal contract, neither does Kris acting on their behalf. That isn't a rule written into the AI's instructions — it's the same permission check the screen performs, applied at the data layer.
Where your data lives
Two models, and the choice is yours:
The same layers of isolation, permission and audit apply in both models.
- On your server (on-premises). The entire platform runs inside your company. Documents, forms, history and indexes stay on your infrastructure. If the internet goes down, the operation keeps running.
- In the cloud, managed by F7. Each customer gets their own environment, separate from the others, with data at rest in Brazil (region set in the contract). Encrypted in transit; secrets encrypted at rest.
What goes to the AI — and what doesn't
This is the boundary that matters, and we state it openly: KORE uses Claude, by Anthropic, as its AI engine. That means a slice of context leaves for the model's API at the moment of use. What governs that path:
- Only the context the task needs is sent — not your database, not your history, not your archive. The slice that goes is the same one the user's permissions would allow them to read.
- Content sent through the API is not used to train models. That's Anthropic's standard commercial term, and their operational logs are discarded within days.
- The transfer is under contract. Anthropic's data processing agreement incorporates Standard Contractual Clauses — the recognised mechanism for international transfer — and their subprocessor list is public.
- On-premises, the data doesn't leave: the query does. The database stays in your house; what travels is the slice sent to the model at the moment of the task.
On-premises the database never leaves your datacenter: only that slice travels.
We don't claim the transfer is eliminated — it is governed. Anyone claiming a frontier generative AI runs with nothing leaving the perimeter is describing something else.
Tenant isolation: enforced by the database, not by code
Most multi-tenant platforms filter by customer in the application — every query carries a hand-written "where tenant = X". One query forgetting the filter, one report aggregating without scope, or one diagnostic endpoint left open, and everyone's data leaks at once.
F7 KORE uses native Row-Level Security in InterSystems IRIS. The isolation policy lives in the table definition, not in the application:
- Each authenticated request assumes its tenant's role before touching data.
- It applies to reads and writes — a write outside the correct role is refused by the database itself.
- Even if a defect produces an unfiltered query, the database returns only what belongs to that customer.
- It is auditor-verifiable: the policy is declared, versioned and inspectable without reading application code.
- It covers background work too — integrations and schedulers establish the tenant role before writing.
The policy is declared, versioned and auditor-inspectable — without reading application code.
That's the difference between "we trust no developer forgets the filter" and "the database won't hand over what isn't yours".
Who sees what
Granular role-based permissions, enforced route by route. Document access is restricted by default and shared explicitly — per person or per group. Kris inherits exactly that map: it executes under the permissions of whoever asked, never with privileged access of its own.
The trail behind every action
Every meaningful action — by a person or by Kris — lands in an immutable history: who asked, what was done, when, under which permission, and what changed. This isn't a debug log; it's a trail built to withstand external audit, with its own retention and legal-hold handling.
The useful side effect: audit evidence — ISO 9001, regulatory good practice, traceability your customer demands — is already assembled when the auditor arrives, instead of becoming a scramble.
Your systems' credentials never leave
When KORE talks to your ERP, the Connect Agent does it — installed inside your perimeter. Usernames and passwords for Protheus, SAP, Sankhya or Senior stay in your datacenter; the end-system credential never crosses the boundary. Architecture detail on the integrations page.
Layers of defence
Every request passes through six layers before it touches data. If one fails, the next still holds:
Kris passes through exactly the same six layers — there is no shortcut for the AI.
What's ours and what's yours
No platform makes a company compliant on its own — be wary of anyone who says otherwise. The split we work to:
- Ours: isolation between customers, access control, the immutable audit, encryption in transit and at rest, the declared AI boundary, and the contract with the AI subprocessor.
- Yours: who holds which role, what each profile may see, how long each piece of information must be retained, and the legal bases for the processing you carry out in your operation.
- A conversation between us: the data processing agreement between your company and F7, with its subprocessor annex — part of the commercial process, not a surprise after signing.
What KORE does is reduce the risk surface and govern the use of AI across the operation: instead of company knowledge flowing through personal tools with no trail, every interaction passes through identity, permission and record.