Where the AI runs matters

Where will the data live? Where will the AI run? And who operates all of it?

Those three questions open most of our conversations with IT directors at mid-to-large industrial companies. Not as a formality — it’s the point where the conversation about AI applied to operations either starts, or never gets anywhere.

The AI market’s default answer is a single one: “in our cloud, with our model, we operate it.” For many use cases, that works. For an industrial company with process secrets, contracts with confidentiality clauses, health audits or end customers who ask where their data is, a single answer stalls the conversation. Not because the company doesn’t want AI — because the offer doesn’t respect the perimeter it took years to build.

The “cloud or nothing” dichotomy doesn’t work

Anyone who has worked in technology long enough knows the boundary:

• Managed public cloud — economies of scale, continuous updates, no servers to buy. But the data lives outside the company’s perimeter, and protecting the trade secret comes down to a contract.
• Traditional on-premises stack — full control of the environment, no external dependency. But it needs an in-house operations team and tends to move slower than its cloud counterpart.

Industry runs a good part of its critical systems in the second category — ERP, MES, QMS, inside the perimeter. For a real reason: the process secret, the contract margin, the procedure calibrated over years of operation. That doesn’t leave the building for convenience.

That’s why, at F7 KORE, the question “where” has no answer of ours. It has two options, and the customer chooses.

Option 1 — SaaS on Google Cloud in São Paulo

F7 KORE as a service, hosted on Google Cloud, São Paulo region. Platform data and processing in Brazil, with operations handled by F7: updates, monitoring, backup. The company accesses it through the web, Teams and Outlook, with no infrastructure of its own and no dedicated operations team.

The platform is multi-company, and the question every CISO asks comes right away: “what stops company A’s data from showing up for company B?”

The answer is architectural, not a promise:

• Per-company isolation in the database. Separation is enforced by row-level security, in the data layer itself: every query only sees the rows of the company of whoever is querying. It doesn’t depend on the application code remembering to filter — nor on an instruction to the AI model.
• Per-person roles and permissions. Within the company, each user sees and does what their role allows. The AI agent acts with the permissions of whoever asked, and permissions are resolved before any data reaches the model.
• Corporate single sign-on. Access through the company’s own identity provider.
• Backup and operations under F7’s responsibility, with the same audit trail the customer sees.

This design serves companies that want fast deployment and don’t want to build their own operations — without giving up isolation.

Option 2 — Fully on-premises

Here, everything stays on the company’s infrastructure: the platform, the data, the audit trail and the AI. A server in the company’s rack or its private cloud — the perimeter belongs to the customer.

• The whole platform in the customer’s environment — database, workflows, ERP integration, search, AI agent.
• Trail anchored in the customer’s own environment. Audit trail anchors go to immutable storage with locked retention, inside the company’s infrastructure.
• The same platform as the SaaS. Not a simplified version; the same product, with the environment swapped.

It’s the model for companies that can’t let the trade secret leave the building — whether because of a contract with an end customer, an audit requirement or the board’s own decision.

And the AI? That’s the customer’s choice too

Where the data lives is half the question. The other half is where the AI runs — and it’s independent of the first.

F7 KORE doesn’t come married to a model. The company chooses:

• A cloud provider, under the contract the company itself signed.
• A private endpoint, dedicated to the company.
• An open model running on its own server, with nothing leaving the perimeter.

And it can switch later without rebuilding the operation: workflows, permissions, trail and integrations stay the same.

Two things hold for any model chosen. First: permissions are resolved up front. The model never receives data the person who asked wasn’t allowed to see. Second: only the excerpt that matters goes to the model. Vector search finds the relevant piece of the standard, the minutes or the contract in the corpus, instead of dumping entire documents into the context. Fewer tokens, fewer hallucinations, predictable cost — and less data in motion, wherever the model runs.

What stays the same across models

Choosing where to run changes the address. Governance stays the same:

• Tamper-evident trail. Every action — by a person or by the agent — goes into a cryptographic chain. Any change is detectable, and any auditor can verify it without relying on F7.
• Customer-controlled keys (BYOK/HYOK) and legal hold to preserve what is under investigation or dispute.
• Prepared for Brazil’s AI bill (PL 2338/2023) and for article 20 of the LGPD, with a record of AI use by person, data and purpose; supports internal and third-party audits.

InterSystems IRIS as the foundation

Both models run on the same foundation: InterSystems IRIS, the engine behind critical systems in finance and healthcare.

The choice isn’t a matter of style. IRIS brings together in a single core what, assembled piece by piece, becomes a point of failure:

• A multi-model transactional database — relational, document and object in the same place, with the volume and latency of a critical system.
• Vector search and SQL on the same base, under the same access rule.
• Workflow and orchestration inside the same process — no collection of services talking over the network.
• Native integration with other systems, which is where the platform comes from.

On-premises, this matters because the company runs a robust platform without building a whole team for database, infrastructure and AI. In SaaS, because isolation between companies lives in the database, not in a patch on top of it.

The choice is the customer’s

The conversation ends where it should: where the AI runs matters. We know that. That’s why the choice is yours, not ours.

Companies with stricter sovereignty requirements choose on-premises, with an open model on their own server. Companies that want fast deployment and have no operations team choose SaaS in São Paulo, with the AI provider they already have a contract with. The platform is the same; the trail is the same; the permissions are the same.

What changes is where the data sleeps.

F7 KORE is built by F7 KORE SOFTWARE LTDA, an InterSystems partner — in cases such as B2B sales and regulated chemicals. Security and sovereignty in detail on the security page; deployment questions on the FAQ page.

If you run a mid-to-large industrial operation with a real data sovereignty requirement and want to understand which model fits your operation — talk to us.