Turn internal data into finished presentations without sending anything to an outside AI.
Every team that builds presentations from proprietary data faces the same problem: the AI tools that speed up the work require uploading that intellectual property (IP) to an external service. A global automotive research institute avoided this by running the AI on its own servers. Its IP stays on its own network, and it still recovered about 30% of expert time.
Talk to us →Already proven in production.
Turning proprietary data into presentation-ready output is slow and manual. Most AI tools that could accelerate the process require uploading this intellectual property to an external, internet-based service. For a regulated or IP-sensitive business, that is not an option.
One of the world’s largest automotive research centres faced exactly this. Its senior researchers were losing roughly 30% of their day to the work, but no internal data was permitted to reach any external commercial AI.
rinf.tech built a system that runs on the institute’s own servers, with no internet connection, turning restricted data into finished presentations without sending anything outside. It recovered about a third of expert time in the first six months, confirmed by an internal audit showing that no data left the institute’s own network.
Expert time recovered. Researchers previously lost this time to manual retrieval and formatting. The institute measured the recovery across its research workflows over the first six months after deployment.
Answer accuracy, out of 1.0. When human reviewers check a sample of the AI’s answers, they now rarely find a factual error. The score comes from RAGAS, a standard measure of how accurately an AI answers from its source documents.
Data sent outside the network. No data sent to any cloud or external server; all processing runs on the institute’s own hardware. Verified by an internal audit of all outbound network traffic.
The manual work of turning data into decks, and why AI couldn’t help
Senior researchers were losing roughly 30% of the working day to the manual work of turning internal data into cross-referenced, formatted presentations.
Public AI tools could help with the drafting, but only at a cost. Before anything could be sent, every piece of proprietary data had to be stripped out by hand, so the tool never saw the institute’s IP. The tool would then draft from that stripped-down data, and the real data had to be put back into the draft by hand afterwards. Each output still needed full checking on top. Between stripping the data out, putting it back, and checking the result, the tools that were meant to save time only moved the work around, and total effort did not actually fall.
That routine existed only because the raw data could not leave the institute: all internal research data is classified as restricted IP, and access to unredacted material is prohibited under UN R155 supply-chain security regulations. Any workable answer would have to run where the data already lived, inside constraints the institute could not change.
Building the AI on the institute’s own servers, with no internet connection
The system answers questions directly from the institute’s internal documents and automates the retrieval and formatting work behind every deliverable. It was delivered as a native PowerPoint add-in, inside the tool researchers already used, so the capability arrived where the decks were already being built and nothing about how people worked had to change. Researchers work from the raw data directly, with no anonymising beforehand and no stitching the IP back in afterwards. The system keeps no state between sessions, clearing each user’s data when the document closes.
What set the work apart was the delivery method: it treated the institute’s security boundary as the first requirement, mapped before any code was written. Even the build happened without touching the institute’s data, running against synthetic test data instead.
# 01 — Boundary audit
An audit of the system boundaries and security constraints — data-classification rules, internal network segmentation, and document-retention policies in place at the institute since 2018. Development began only after the client security team signed the audit off.
# 02 — Isolation architecture
The system was built to operate within the existing security boundary. No outbound connections were permitted, and model execution ran on institute-owned hardware: no external services, no API calls, no telemetry. Isolation was the foundation of the design, built in before anything else.
# 03 — Rollback architecture
Each phase included a standardised, defined rollback procedure, tested in staging before deployment, with triggers agreed in advance with the institute’s governance team. No rollback was required during deployment — but its existence was a precondition of client sign-off at each gate.
Working with rinf.tech was the first time we implemented an AI system where total effort actually went down instead of up. They didn’t ask us to change our security model. They didn’t ask us to move data. They built the solution to work inside the constraints we already had. Respecting those boundaries is the core priority in research.— Project Lead, Automotive Research Institute
Effort went down instead of up.
About 30% of expert time returned to research
In practice: roughly a third of senior researchers’ working day went back to actual research, without added staff or new verification steps.
Answers accurate enough to trust
In practice: independent reviewers, checking the answers blind, now rarely catch a factual error (a score of 0.9 out of 1.0). The manual checking that used to erase the time savings largely falls away, because the output no longer needs policing line by line.
No data sent outside, confirmed by audit
In practice: the institute gets the productivity of a modern AI while all of its data stays on its own servers, within the network boundary UN R155 requires, keeping full compliance while gaining the speed. An internal audit of all outbound network traffic confirmed that no data was sent to any external or cloud system. Deployment completed in Q1 2026, with no rollback required.
The same problem across regulated industries
The constraint here belonged to one institute, but the pattern does not. Teams that build presentations from data they cannot upload to an outside service face the same choice between speed and control. The same approach resolves it: an AI running on the organisation’s own servers turns the data into presentation-ready output, with nothing sent to any external or cloud system.
Pharmaceutical & life-sciences R&D
Trial results, compound data and regulatory dossiers that confidentiality rules keep off external services.
Defence & aerospace
Classified or export-controlled material that cannot go to a commercial cloud.
Financial services
Material non-public information and client data governed by strict handling and disclosure rules.
Legal & professional services
Privileged documents and matter files where a single exposure carries real liability.
Engineering & manufacturing
Proprietary designs, specifications and process IP that competitors would pay to see.
One of the world’s largest automotive research centres
Part of a $170B global automotive manufacturer, the institute runs a large body of internal research classified as restricted IP under UN R155 supply-chain security regulations. The engagement covered internal knowledge retrieval and research-deliverable automation, with deployment completed in Q1 2026.
- Industry
- Automotive R&D
- Integration
- Native PowerPoint add-in
- Deployment
- Q1 2026
- Constraints
- No data sent outside · no external AI · no workflow change
- Delivery duration
- 3 months
- Solution
- On-premise AI for document Q&A and deck generation
See how spec-based AI delivery works inside your constraints
If IP or compliance rules have kept AI out of your most sensitive work, the same approach applies: an AI that runs on your own hardware and turns your data into finished work, with nothing sent outside. We can walk through what that would look like in your environment.
Talk to rinf.tech →