AI is already inside.
The question is where.
Find, map and inventory the AI your organization already uses.
People, teams and processes adopt AI tools without going through IT. What you cannot see, you cannot govern. The Shadow AI Assessment reads the telemetry you already have, and ARPIA turns it into an inventory of every AI tool, agent and model API in use, the devices behind it and the systems it reaches, in a live map called Tower Control.

Four ways AI gets in without anyone deciding it.
Free chats
Personal accounts on public assistants, with company information pasted into the prompts.
Agents and scripts
Automations that call model APIs with keys nobody inventoried.
Access to critical data
AI tools connected to email, files and systems of record, outside any control.
No inventory
Nobody knows which AI exists, who uses it, from which device or what data it touches.
Five questions leadership needs to answer.
- Which AI is in use?Tools, copilots, agents and models detected.
- From where?By device, user or unit, depending on the sources available.
- With which account?Company or personal and free, where policy allows us to see it.
- What data does it touch?The systems of record and services each AI actor reaches.
- What risk does it carry?Rated by tool, by device and by system.
The result: a map of your AI ecosystem you can navigate, and a roadmap to govern it.
What we discover.
A basic to intermediate assessment, centered on devices and AI actors.
Web and desktop AI chats
Public assistants and installed apps.
Embedded copilots
Assistants inside the office suite and your SaaS.
Agents and automations
Flows, RPA and scripts that call model APIs.
MCP servers and connectors
Bridges between agents and internal systems.
Local models
Models running directly on computers.
Browser extensions
AI add-ons with access to pages and data.
From raw telemetry to a map of the ecosystem.
- Capture
The logs you already have: proxy, DNS, SSE, identity, EDR or MDM, office suite, SIEM, VPN. Read-only, within an agreed window. Log extracts in ZIP files work too.
- Local preprocessing
Parsers run on a virtual machine inside your network. Data is pseudonymized at the source, and the raw data does not leave.
- ARPIA
Normalizes and correlates everything into one ontology: device, AI actor, model and system of record. Oria classifies with reasoning, and the team validates every finding.
- Tower Control
An interactive graph, the inventory by device, the risk findings with their recommendations, and Oria to ask questions about it.
Once data enters the ontology it is normalized. The only thing that changes from one organization to the next is the capture, which adapts to your sources.
Your AI ecosystem on one screen.
Tower Control is what you receive: not a static report but a live view you can keep querying for 90 days. Click a device to follow its path, filter by risk, and ask Oria what a script sends and where.


Real Tower Control screens from a one-site capture. Device names, addresses, people and the organization are blurred.
We use what your organization already has.
No single source sees everything. The value is in correlating them.
| Source | What it reveals | Usual limit |
|---|---|---|
| Network (proxy, firewall, DNS) | AI domains and APIs contacted, volume and origin | Without TLS inspection, only the domain is visible |
| CASB or SSE | AI SaaS applications and their classification | New tools take time to appear |
| Identity (SSO, OAuth) | AI apps with permissions over email and files | Does not see personal accounts |
| Endpoint (EDR, MDM) | Software, processes, local models and agents | Requires a deployed agent |
| DHCP, NAC, VPN | Mapping an IP to a device and a person | Guest networks and personal devices |
| Systems of record | Which accounts and tools read which data | Requires application logs |
What network data can and cannot tell. Network data identifies the AI provider and the type of use. It does not identify the model, the content of the prompt or whether the account is personal or corporate. That takes SSO logs, the AI providers' consoles, an API gateway such as ARPIA, or TLS inspection or a CASB. The assessment measures what the available sources can see, and the report says how much each source covers.
The data stays where you decide.
ARPIA secure cloud
Local capture and processing on ARPIA's infrastructure, certified SOC 2 Type II and ISO 42001, audited by A-LIGN.
Hybrid
Preprocessing and pseudonymization on a virtual machine of yours. Only data that is already reduced goes to the cloud.
On premise
Capture and processing inside your own infrastructure, with ARPIA Private AI.
The deployment model is chosen with the discovery questionnaire, according to your privacy and data residency policy.
How ARPIA runs on your infrastructure →What your organization receives.
AI inventory
Tools, agents and models, by device.
Interactive graph
Tower Control with the whole ecosystem, with access for 90 days.
Risk assessment
By tool, by device and by system.
Roadmap
Prioritized remediation and an executive presentation.
The report is generated from the ARPIA platform and includes the coverage of each source. Delivery takes two to three weeks from the moment all the data is in.
An assessment that turns into governance.
Governed platform
Roles, audit and full traceability over data and models.
A living ontology
The inventory is not a snapshot: it stays modeled and can be updated.
AI with human control
Oria classifies and proposes. Your team validates each finding.
A base you reuse
The same map serves for continuous monitoring and to govern the AI you approve.
Find out which AI is already inside.
Tell us who you are and we will send you the discovery questionnaire. Your answers set the scope: the sources, the sites and the capture window.
- QuestionnaireYou answer the discovery questionnaire.
- Scope and proposalWe define the scope together and send a proposal.
- PreparationLog samples, volume, the virtual machine and internal approvals.
- StartThe capture window opens.