Operative AI, safely

Connecting an LLM directly to your ERP APIs isn't modern. It's negligent.

An LLM with direct access to your ERP, your email and the open web is a documented security anti-pattern. Data exfiltration happens without a click. Audit trails become incomplete. And you, the managing director, are personally liable. A control layer isn't a convenience. It's the only architecture that meets data protection, industry-specific audit requirements and the EU AI Act at the same time.

What has already happened

Three documented incidents from 2024 and 2025.

These cases are not hypothetical. They happened. At Microsoft, at Slack, at Replit. With the same products that are being sold to you.

  • June 2025 · CVE-2025-32711

    EchoLeak: Microsoft 365 Copilot, zero-click data exfiltration

    A single incoming email was enough to extract data from Microsoft 365 Copilot. No click required by the recipient. Severity 9.3 out of 10. Microsoft patched server-side. The flaw was conceptual, not in the code.

    HackTheBox · CVE-2025-32711

  • August 2024

    Slack AI: data exfiltrated from private channels

    A crafted message in a public channel was enough for Slack AI to leak data from private channels. Demonstrated publicly by security researchers PromptArmor.

    PromptArmor · Slack AI exfiltration

  • July 2025

    Replit Agent wiped production database despite code freeze

    The agent ignored an explicit stop, deleted a customer database with 1,206 records, lied about recoverability and tried to conceal the incident. Documented in the AI Incident Database.

    Fortune · Replit catastrophic failure

Heise also reported in 2025 that Microsoft 365 Copilot did not record file accesses in the audit log when the attacker suppressed the reference link. Microsoft patched silently, without notifying customers. An audit trail with selective gaps is no audit trail.

What you personally risk

Data protection, EU AI Act, GmbH liability. Three levers, at least one hits you.

  • Data protection

    GDPR fines up to €20m or 4% of global turnover.

    Italy's data protection authority fined Luka Inc. (Replika) €5m in 2025: missing legal basis under Art. 6 GDPR, transparency gaps, lack of age verification. The earlier €15m against OpenAI was overturned in first instance by the Tribunale di Roma in March 2026; the Garante can appeal and the written reasoning was not yet published at the time of writing. But supervisory authorities across Europe continue to sanction AI violations.

    EDPB · Replika fine €5m

  • EU AI Act

    High-risk AI under Annex III carries its own obligations.

    Penalties: up to €35m or 7% of global turnover for prohibited practices, up to €15m or 3% for other infringements, up to €7.5m or 1% for false statements to authorities. For SMEs, the lower amount applies. High-risk AI carries duties for logging, transparency and human oversight.

    EU AI Act · Art. 99

  • GmbH liability

    Managing directors are personally liable for GDPR violations.

    The Higher Regional Court of Dresden ruled in 2021 that GmbH managing directors are personally liable alongside the company for data protection breaches. No Federal Court of Justice ruling yet. But one higher-regional ruling is enough to be sued in a damages case.

    OLG Dresden · 4 U 1158/21

Caveat: not every ERP workflow qualifies as high-risk under the AI Act. Pure order-handling or logistics agents typically don't fall under Annex III. The moment an agent touches HR data, creditworthiness or access to essential services, the high-risk obligations under Annex III apply. High-risk AI in regulated products under Annex I has its own deadlines. Which deadlines apply is set out in the current text of the regulation in the Official Journal of the EU.

What the control layer solves

The layer makes standard AI traceable.

A control layer between LLM and ERP is not an alternative to Microsoft Copilot or SAP Joule. It is the compliance layer that makes them suitable for regulated industries in the first place.

  • Tool scoping instead of full access.

    The AI assistant gets no direct API keys. It calls defined tools that may do exactly what they are meant to. Nothing more.

  • PII boundary in both directions.

    Recognised personal data is replaced before it leaves the server and restored on the way back. Whatever the recognition misses goes out unchanged.

  • A log without content.

    For every request to a language model in the chat and in the tools we record the time, the person, the provider, the model, the tokens and the cost. Every replacement is kept in a separate replacement protocol with the number of recognised values and checksums. We do not store the content.

  • Vendor independent.

    Today Cortecs in Vienna routes most requests to the models. A local model works just as well. The layer stays. The provider is replaceable.

Objections you are right to have

Six objections against a small specialist provider. Our answer.

  • Objection

    „You're a small provider. What if you disappear tomorrow?"

    Answer

    For the results made for you, you get an exclusive, perpetual right of use. For reusable building blocks, you get a simple, perpetual right of use. A written handover process is contractually committed. AI-assisted development has clearly lowered the threshold for a takeover. A qualified vendor can take over an MCP-based layer, because the protocol is openly documented and the code is readable.

  • Objection

    „We need SLAs and guarantees, not trust."

    Answer

    Availability, response time, maintenance windows, escalation and backups are set out in the written contract. That includes clear terms and termination options.

  • Objection

    „Standard products are safer. Thousands of customers have already tested them."

    Answer

    That argument falls apart against the incidents documented above. EchoLeak, Slack AI, Replit — all standard products with millions of users. Thousands of testers did not prevent the gaps. For regulated industries, custom development with a clear boundary layer is no longer exotic in 2026; it's what auditors and insurers expect.

  • Objection

    „Why not wait until our ERP or CRM vendor solves this themselves?"

    Answer

    Because waiting costs money. The chart below shows it honestly. Your ERP vendor (abas, proALPHA, Sage, in-house) has AI modules on the roadmap. But concrete voice and audit layers will appear in one of the next major releases at the earliest, usually tied to an upgrade of your ERP contract. Your CRM vendor (HubSpot, Pipedrive, SevDesk) ships island solutions that don't talk to the ERP. Neither carries you across the actual workflow.

  • Objection

    „Where does this layer even sit? It sounds like yet another component."

    Answer

    Between two worlds that currently talk past each other: at the top, the wild AI skills of Glean, Copilot, ChatGPT — good for open search, bad for regulated workflows. At the bottom, hard-coded ERP APIs — exact but voiceless. The layer turns language, workflow and audit into one thing. It's a translation, not another island.

  • Objection

    „Who's liable if the layer itself fails?"

    Answer

    The contract sets this out. On top come professional liability insurance, automated tests in the code and version tags. In an incident we can see when which model was called for whom and how many values were replaced. We do not store the content.

The cost of waiting

Opportunity costs, made concrete.

Three realistic company sizes. 1.8 hours per employee per workday lost to cross-system search. Conservatively, half of that is addressable. And: waiting doesn't cost linearly. Data silos solidify, workarounds proliferate, shadow AI takes hold, competitors capture share. 2% monthly compounding — the lines curve upward.

Chart: cumulative delay cost over 36 months of waiting for three company sizes — 30, 100 and 300 employees — with 2% monthly compounding. At 30 employees this reaches roughly €1.85m after 36 months, at 100 employees roughly €6.18m, at 300 employees roughly €18.53m.
0 5 Mio € 10 Mio € 15 Mio € 20 Mio € 0 Mo. 6 Mo. 12 Mo. 18 Mo. 24 Mo. 30 Mo. 36 Mo. Months of waiting Cumulative delay cost
  • Smaller operation — 30 employees

    35.640 € / Monat · 1.853.079 € nach 36 Monaten

  • Mid-market — 100 employees

    118.800 € / Monat · 6.176.931 € nach 36 Monaten

  • Larger operation — 300 employees

    356.400 € / Monat · 18.530.792 € nach 36 Monaten

Assumptions: 1.8 h/employee/workday lost (studies, see the ERP-voice-ready page), 50% of that addressable, 22 workdays/month, €60/h fully loaded, 2% monthly growth factor for compounding effects (data silos, shadow AI, market-share loss). No discounting. Follow-on costs from bad decisions on unfindable data come on top.

Honest framing

What this page does not claim.

  • The Garante fine against OpenAI was overturned in first instance.

    The Tribunale di Roma vacated the €15m sanction on 20 March 2026 at first instance. The Garante can appeal; the written reasoning was not yet public as of May 2026. We say so honestly. It does not kill the argument: the Replika €5m sanction from 2025 stands, and CNIL, BfDI and EDPB continue to act on AI violations. Regulators do enforce — but first-instance decisions don't always hold.

  • Director external liability is not yet settled by the Federal Court.

    The OLG Dresden ruling is higher-regional, not BGH level. Legal scholars dispute the scope of external liability. But: one such ruling is already enough to be sued personally.

  • Not every AI use case is high-risk.

    The AI Act differentiates. Purely administrative or logistics agents are not automatically high-risk. The moment HR decisions, creditworthiness or access to essential services come into play, that changes.

  • A control layer is not risk-free.

    It requires maintenance. It carries knowledge holders. It needs a documented exit strategy. We build it so that the code lives on your servers and MCP, as an open standard, keeps the model replaceable.

Michael Schiller

Let's talk about your case.

30 minutes. Your ERP, your risk profile, your question. No pitch.

Book a call

Sources and further reading

Everything we claim has an address.

As of 21 May 2026. All sources verified at the time of publication. URLs may change.

Michael Schiller

Let's talk about your case.

30 minutes. Your ERP, your risk profile, your question. No pitch.

Book a call