Your data has already left the building.
Right now, without asking, your people are pasting contracts, customer lists, pricing and source code into consumer AI to get their work done. It's fast, it's useful — and it's a daily data-exfiltration event with no audit, no access control and no way to take it back. You can't ban it; prohibition just drives it underground. You can substitute it. Dimbo is the sanctioned alternative — the same everyday AI, pulled back inside your perimeter and governed by construction.
Only 13.5% of EU enterprises used AI at all in 2024 (Eurostat) — yet three in four of their employees already use it privately. The adoption isn't missing. The governance is.
Shadow AI is already inside your walls.
This isn't a future risk to plan for. It's happening in every department today — and because it happens on personal accounts, on public models, you have no record that it happened at all.
The prompt is the leak
A sales rep pastes a signed contract to "summarize the terms." An engineer drops proprietary source into a debugger. Sales pricing, customer PII, supplier formulations — copied into a public model, stored on someone else's servers, gone. The Samsung engineers who leaked source code into ChatGPT in 2023 weren't reckless; they were just trying to work faster.
You can't govern what you can't see
It happens on personal logins, on consumer apps, off your network. There is no log of what was shared, with which model, by whom. When a DPO asks "what customer data has left the company?", the honest answer today is: nobody knows. That's not a policy gap — it's a blind spot the size of your workforce.
Four regimes at once
For an EU firm, a single careless paste can collide with the GDPR (personal data to a third country), the EU AI Act (human-oversight obligations), NIS2 (security duties) and Schrems II + the US CLOUD Act (data sovereignty). Consumer AI is convenient precisely because it ignores all four.
You can't prohibit your way out of this.
A policy memo that says "don't use ChatGPT" doesn't remove the incentive — the work is still faster with AI, so people just do it more quietly, on their phones, off the record. Prohibition doesn't reduce shadow AI; it removes your last shred of visibility into it. The only strategy that works is substitution: give people an AI that is better to reach for — one that already knows the company, answers in their language, and happens to be governed. Take the reason to go outside the walls away, and the traffic comes home on its own.
"The remedy isn't a rule. It's a better default."
The same everyday AI — governed by construction.
Dimbo gives your team a role-scoped assistant that answers from live company knowledge, on your infrastructure. Every property that makes shadow AI dangerous is inverted here — not by policy, but by architecture.
The model runs on your box
A benchmarked local LLM — gemma-class on a single RTX 3090, at reference-parity with the cloud model — plus local vision and local Whisper transcription. The default deployment is air-gapped: nothing leaves the building. When an EU-hosted or PII-gated cloud tier is ever used instead, that's a choice you make per install, not a hidden dependency. No CLOUD Act reach, no Schrems II fragility.
Anonymized before any external model
On the rare path where text does reach an external model, it passes through the Presidio privacy gateway first — names, IBANs, cards, phone numbers, credentials masked before a single token leaves. Local deployment keeps everything inside the perimeter to begin with. The raw data stays in your database, un-anonymized, under your control.
An operator never sees the cap table
The assistant is role-scoped: access control is applied at the moment knowledge is retrieved, not bolted on after. A line operator asking "how do I do X" gets operator knowledge; they never see the board minutes, the pricing model or the cap table. Every employee gets the expert colleague — filtered to exactly what their role is allowed to know.
The autonomy ladder is the EU AI Act mechanism
Every AI action climbs a per-process autonomy ladder — and it only earns autonomy through a measured track record you approve. Human-in-the-loop is the construction, not a setting: every level change is audited and emitted as an event, promotion is always your decision, and a master kill-switch is one flip away. AI Act human-oversight obligations are met by the product's core mechanism, not a compliance bolt-on.
What changes when the AI is yours.
"Summarize this customer contract and draft a reply." The task is identical. Where it runs, what it leaks, and whether you can prove any of it — is not.
| Shadow AI · consumer tools | Dimbo · governed | |
|---|---|---|
| Where the model runs | ✕ Someone else's cloud, outside the EU | ✓ On your box, air-gapped by default |
| What happens to PII | ✕ Pasted raw, stored on their servers | ✓ Never leaves locally; Presidio-masked if ever external |
| Access control | ✕ None — anyone sees anything they paste | ✓ RBAC at retrieval, per role |
| Audit trail | ✕ No record it ever happened | ✓ Every action logged, provenance on every edge |
| Human oversight | ✕ Autonomous, invisible, ungoverned | ✓ Autonomy ladder — human-in-the-loop by construction |
| Regulatory posture | ✕ Collides with GDPR · AI Act · NIS2 · Schrems II | ✓ GDPR-by-design · AI Act oversight-by-design |
Every AI action is a proposal you approve.
The assistant doesn't act behind your back. It proposes, with evidence attached, and waits — until the process has earned the autonomy you granted it. This is the oversight the EU AI Act asks for, rendered as a button.
The same query that would have leaked the contract to a public model is answered inside the walls, filtered to the role, logged in full — and still handed to a human to sign off.
Real properties, not badges we don't hold.
Everything below is a property of the architecture, live today. We do not claim certifications we haven't earned.
How the governed perimeter is built
The expected annual cost of a shadow-AI-linked incident for a mid-market firm sits around €25,000/year in avoidable exposure (Dimbo value model §2.5) — before the unquantified competitive leakage of pricing and formulations walking out the door. Substitution is the cheapest insurance you can buy.
The full argument, sourced.
Shadow AI: Your Data Has Already Left the Building
The scale, why prohibition fails, the regulatory collision, and the governed alternative — anchored to the EU AI Act, GDPR, the Work Trend Index and named real incidents.
Where the model runs is your choice
On-prem appliance, EU-hosted, or PII-gated cloud — same platform, same intelligence. A benchmarked local model makes air-gapped the default, not a premium tier.
The role-scoped assistant in context
How the assistant reads from the shared knowledge store, respects RBAC at retrieval, and answers in the operator's own language — inside the perimeter.
Bring the AI your team already uses back inside the walls.
Start with the free Deadline Audit — it runs on your data, on your infrastructure, and never leaves the building. The surest way to see the sovereignty story is to watch it stay home.