FAQ
The questions IT leads actually ask.
Straight answers, including the ones that don't favor us.
Are open-weight models good enough compared to GPT-5- or Claude-class models?
For focused business-process agents — parsing an IDoc error, reconciling a batch record, answering from your own documents — yes. A scoped agent doesn't need frontier reasoning; it needs your data and a tight task. If your workload genuinely requires frontier-model reasoning and can't accept the trade-off, we're the wrong partner — and we'll tell you that in the fit call.
What hardware do we need?
Less than most expect. Compact models run on a single GPU; a 70B-class model at 4-bit quantization needs roughly 35–40 GB of VRAM. We size the exact requirement during scoping — and if you have no on-prem hardware, a GPU instance inside your own cloud tenant works the same way.
Where does our data go?
In a deployment: nowhere. There is no aplexio backend between you and the answer — the model weights are static files on your infrastructure, with no phone-home and no telemetry. We never receive, store, or train on your data. The free version is different: it runs in Claude Code with Anthropic's cloud model (see the free-trial question).
Can the agents post in SAP?
Yes, where you decide they should. Our agents can read, recommend, and execute through interfaces you sanction. You set permissions and autonomy per workflow — observe, advise, act with approval, or act on their own where the process justifies it. Each agent gets its own technical user with least-privilege roles, changes go through your normal transport and approval process, and every action is logged. The free SAP Logistics Copilot stays read-only: it works on exports you paste.
What about GDPR — do we need a data processing agreement (DPA)?
You remain the controller throughout; your data is processed by you, on your systems. Where we touch personal data at all during an engagement, we act as a processor under a DPA (Art. 28 GDPR) concluded for the engagement. After handover, aplexio is out of the processing picture entirely.
Do we need to involve the works council?
In Germany, often yes — under § 87 BetrVG, systems that can monitor employee performance are subject to co-determination. We scope agents to work on process data, not people data, and we provide the technical documentation your works-council conversation needs. Plan the conversation early; it's faster than retrofitting.
Are you ISO 27001 or SOC 2 certified?
No, and we won't pretend otherwise. Our answer is architectural: because no customer data ever transits aplexio systems, our certification status matters far less than it would with a SaaS vendor. What you should audit is the deployment in your environment — and we hand over the documentation to do exactly that.
What does the free trial involve?
The free SAP Logistics Copilot runs in Claude Code: the skill pack runs on your machine and Anthropic's cloud model does the reasoning. Try it on the bundled sample IDocs or a redacted export from your system, not live production data — before any contract. You judge the output on your own cases, not on our demo. The two-week paid pilot also runs in Claude Code, on an anonymised export; in production the agent runs in your environment on open-weight models.
What realistically fits into 2–6 weeks?
One scoped use case: we build the agent, deploy it into your environment, integrate it with the system it works against, and hand over a running setup with documentation. Not a platform program, not an AI strategy deck — one workflow that works.
Which models do you deploy?
Open-weight models — Llama, Qwen, DeepSeek, GLM, Mistral — selected per use case, hardware budget, and license terms. We're vendor-independent: the model is a component we pick for the job, not a partnership we have to justify.
A question we didn't answer?
Ask it in a 30-minute fit call — you'll get a clear yes, a clear no, or a pointer.