Security and data
How Aidecis protects your data and your authority
This page is for procurement, security and risk teams. It sets out how Aidecis is built, which data it reads and records, where that data is hosted, who processes it, and what you keep if you leave. You can read it without a sales call.
Aidecis is the decision intelligence layer for supply chain. Aidecis reads from your systems and never writes into them. Your people, systems or bots execute the approved action, and Aidecis reads each execution back and reconciles it against the approval. SAP stays your single source of truth.
Read the architectureArchitecture
Four parts, one boundary.
Your systems stay the systems of record. The gateway in your environment calls out to the Aidecis control plane; nothing connects into your network.
- Control plane
Holds your policies, delegation of authority, approval routing and the decision record. It is one shared platform for all clients; each client is a separate tenant, and row-level security in the database scopes every request to that one tenant. The policy kernel is deterministic code. No AI decides what is allowed.
- Gateway and grants, on your side
Software that runs on a virtual machine or container in your environment. It reads approved sources and holds no write rights in your systems. For each approved action, Aidecis issues a signed, single-use grant bound to the exact route, object, action, amount, target and policy version. With enforced control, your own integration checks that grant before it executes: a tampered, expired, replayed or broadened grant does not run, and a retry cannot create a second action.
- Model gateway, hosted in the EU
Language models draft briefs and explain flags. They run behind one model gateway on EU-owned infrastructure. It keeps no prompts or answers beyond the request, keeps tenants apart, holds no credentials and has no network path beyond its approved route. A model never decides, computes authority or an amount, approves or acts.
- Hash-chained decision record
Every step, review, approval, grant, execution and outcome is written to an append-only record in which each event carries the hash of the one before. Corrections are new events. A changed, removed or reordered event breaks the chain.
Fail-safe by default: when policy, identity, evidence or the gateway cannot be trusted, or the control plane cannot be reached, Aidecis issues no grant, and where your integration checks grants, the consequential action holds. Your own systems keep running and your people decide as before.
Data flows
What Aidecis reads, records and shows a model.
Access is declared per decision type in a data contract, and it grows in stages that you approve: files, read-only, live evidence, then governed execution. The first slice of every pilot is read-only.
- ReadOnly what the data contract declares
Approved sources and fields, at the freshness the contract sets. Sensitive fields are excluded by default. Every fact keeps its source system, source record and time, so you can see where it came from.
- ExecuteBy you, read back by Aidecis
Aidecis never writes into your systems. Your people, systems or bots execute the approved action, and Aidecis reads each execution back and reconciles its amount, quantity and object against the approval. A difference is held and settled by the deciding seat, never by the proposer. The decision record points to the SAP document.
- ModelMinimal fields, nothing kept
A model receives only the fields a step needs, without your tenant identity or any credentials. Source text is treated as data, never as instructions. Nothing is stored after the request, and your data is never used to train or fine-tune a model unless you explicitly opt in. The Aidecis model, in development, is trained on Aidecis's own decision library, synthetic cases and fictional demo data, never on your content such as names, suppliers, materials, amounts, documents or free text. Your written opt-in covers only de-identified decision-process records (decision type, the amount as a share of the mandate limit, seat tier, route, timings, rework and outcome class), and you see what left.
Your administrator consents to scopes, purposes, processors, regions, model providers and any learning participation. Retention, deletion, export and legal-hold rules are set per data class and per contract. Connector access is reviewed on personnel change, incident, contract end and at each recertification.
Deployment options
The gateway always runs with you. You choose where the rest runs.
Aidecis supplies software only. Whichever option you choose, the policy kernel, evidence rules and containment rules are identical.
- DefaultAidecis-hosted
The control plane and the model gateway are hosted by Aidecis in the EU, with an open-weight model on EU-owned infrastructure. No third-party AI vendor is in the path. Nothing extra for you to host.
- Inside your boundaryClient-hosted model
The model runs on a host that you provide and manage, connected to the gateway. No model data leaves your network.
- Only if you opt inExternal AI provider
An external provider in an EU region, under contractual no-training terms, receiving minimal fields only. Used only when you explicitly choose it.
- Priced separatelyDedicated control plane
The control plane for your organisation only, on EU-owned infrastructure chosen with you or in your own environment, operated by Aidecis.
Aidecis can also run rules-led, with no language model at all.
Subprocessors and data residency
Hosted in the EU, on European providers.
The application and its database run in the Netherlands, model inference in Finland, and encrypted backups in Germany and Spain. Hosting, model inference and backups run on European-owned providers rather than the EU regions of US hyperscalers.
| Supplier | Purpose | Country |
|---|---|---|
| TransIP | Application hosting: the control plane and its database; website hosting and DNS | Netherlands |
| Hetzner | Model inference for the model gateway | Finland |
| Hetzner | Encrypted off-site backups; internal client administration | Germany |
| Internxt | Long-term encrypted backup copies | Spain |
| Business email; optional Google sign-in | As stated in the subprocessor schedule | |
| Your identity provider | Single sign-on and user provisioning, if you connect one | Your choice |
Backups are encrypted before they leave the server, with a key held offline. Source control, code review and internal chat tools receive no client data. The current subprocessor schedule, with processing locations, is part of the data processing agreement in the vendor-trust pack.
Control coverage
What Aidecis enforces, and what it observes.
Enforced before execution where your own integration checks the Aidecis grant; observed and reconciled after execution for registered actors outside it.
- Before the actionEnforced control
Policy and authority are checked before Aidecis issues a single-use grant. With enforced control, your own integration checks that grant before it executes: without a valid grant, the action does not run.
- After the actionObserved control
For registered actors that write in their own systems with their own credentials, such as a vendor's agent or your own algorithm. Aidecis reads their writes from the systems' change records, checks them against the actor's role and limits, reconciles them into the decision record and flags a breach within the agreed time.
Observed control never prevents an action; it shows it. An observed decision is never labelled as authorised before the fact. You choose which actors move from observed to enforced.
Security practice
Controls you can test.
- Identity and segregation of duties
Single sign-on with your identity provider over OIDC, multi-factor authentication through that provider, and SCIM provisioning. Nobody approves their own proposal, and the sponsor of an agent cannot approve that agent's proposals.
- Agents with limited keys
Each agent has its own key, stored only as a hash, limited to reading and proposing, and valid for at most 90 days. Switching off the agent or its sponsor stops the next request.
- Containment of language models
A model runs as a separate process with no credentials, no write tools and no network access beyond its approved endpoint. An instruction planted in an e-mail, a supplier message or an ERP text field cannot cause an action, because Aidecis holds no write rights in your systems, and an enforced action runs only with a grant no model can create.
- Engineering and operations
Every change goes through a reviewed pull request. Dependencies, secrets, container images and code are scanned automatically on every change and every week. Backups run nightly, are encrypted off-site and are restore-tested.
Continuity and exit
Your governance does not depend on a young vendor.
You own your decisions, dossiers, run records, evaluation sets and accepted policies. Every subscription carries a continuity package.
- Full export in open formats
Your administrator exports everything your environment holds, in JSON and CSV, at any time and at the end of the contract, without asking Aidecis.
- Open-source ledger verifier
A verifier under the MIT licence checks an export against its hash chains and file hashes on your own machine, with no Aidecis system and no internet connection. A copy comes with every export. Read its source code on GitHub.
- Source-code escrow
A contract term: the source code of the control kernel and the gateway is held by a named escrow agent and released on insolvency or end of support, with a guide to build and run it.
- Exit assistance and deletion
A contract term: an agreed number of days at the published day rate to move your policies and records to the platform you choose. Once you confirm you have your export, Aidecis deletes your environment's data.
Certification status
No certification claimed. Documentation on request.
Aidecis does not hold a security or AI certification and does not claim one.
Our security and AI-governance documentation is mapped to ISO/IEC 42001 and to your deployer duties under the EU AI Act. It is available on request in the vendor-trust pack, with the threat model, the security baseline, the incident-response and restore runbooks, the data processing agreement and the subprocessor schedule.
Security is improved continuously under a documented process on the ISO/IEC 27001 model: a risk register reviewed monthly, scanning on every change to code, dependencies and containers, security acceptance tests run as a regression set, a log of findings and fixes with dates, and a monthly management review. The process and its log are in the vendor-trust pack.
Vendor-trust pack
Give your security team the documents they need.
Ask for the vendor-trust pack: security, privacy and AI-governance documentation, our prices and an architecture walkthrough.

Contact