Frequently asked questions about Xentral AI-ERP

All of Xentral's infrastructure and AI infrastructure is hosted on European servers and is subject to European data protection regulations.

Yes. Xentral maintains dedicated customer instances: every customer gets their own data silo. It is architecturally impossible for one tenant's data to flow into another tenant's answers. The connection to external LLMs runs exclusively through secure APIs that allow the model providers neither to store your data nor to train on it.

No black boxes. All AI activity is designed to be traceable and auditable.

Right now, every action requires a human in the loop: an employee reviews and confirms each suggestion from the agent before any data is written or emails are sent.

No. The AI's permissions always mirror the general system permissions of the respective user.

Xentral has implemented security measures and guardrails to detect and fend off prompt injection attacks.

Xentral uses a best-of-breed approach with multiple models: Anthropic (Claude), Google (Gemini), and OpenAI (GPT-4). The most suitable model is selected for each ERP task. Open-source and EU models are also part of the potential pool.

The orchestrator is a higher-level AI agent that understands incoming requests and routes them to the right sub-agent (e.g., returns, incoming invoices, payment reconciliation). It shows the sub-agent's result for confirmation before any action is carried out.

Yes. With the upcoming MCP server, there's nothing standing in the way in principle.

Yes, it's actively in development.

Yes.

Yes. That's exactly what you'll be able to use the upcoming Claude integration and the MCP server for.

We're working on enabling channels other than email for requests coming into the Agent Hub.

Yes, the system scales with you and works with both large and small data volumes.

No. It's Xentral's own AI, built specifically for ERP tasks and integrated into Xentral.

Currently only via email forwarding to your personal Agent Hub address. More options are planned for the future, such as Zendesk tickets and additional interfaces.

Yes. As long as emails are forwarded to the Agent Hub address, it works without any issues.

No. There's a single address that all emails are sent to. The orchestrator automatically routes them to the right sub-agent.

Yes. In the Agent Hub's knowledge base, you can store different projects along with all their details. The AI can handle this without needing separate Agent Hubs or email addresses.

Incoming messages are filtered for spam upfront.

As of now, each instance gets one email address for forwarding. Individual tasks can be assigned to specific employees using labels.

Yes, the communication is stored on the relevant customer record.

The Agent Hub is a completely new product and independent of the existing ticketing system. It can be used as a replacement.

As long as emails are forwarded to the Agent Hub address, that shouldn't be a problem.

Yes, the specialized product-questions agent is designed for exactly this. In addition, product data sheets, website content, and other documents can be fed in via the Learning Loop.

Yes, the product-questions agent can recognize images and read attachments.

Yes. The agent reads various attachments (PDFs, images, CSV, etc.) and can automatically create orders from them.

Yes. The system can also identify items by name or other details such as color, size, or quantity.

That's a typical use case. The orchestrator recognizes the different topics and distributes them to the responsible sub-agents.

During the beta, we're testing whether intelligent order suggestions based on stored product information and sales history are possible.

Yes, stock levels can be checked via the Co-Pilot (chat), provided the relevant permissions are enabled.

Yes, the Co-Pilot has access to the Analytics module. The relevant user permissions are required.

Yes.

Yes. The agent supports all languages covered by common models like ChatGPT, so a very broad range for everyday needs.

Yes, the tone can be adjusted, for example via the Learning Loop.

Yes, that's easy to do via the Learning Loop. For example: “Always translate into French and use the formal form of address.”

Yes. A product question from someone who isn't a customer yet can be answered just the same.

Purchasing agents are in development but not yet included in the current beta.

That's a good use case for the AI-assisted CSV importer (in development). For product descriptions, there are already Flows in the Flow library that create AI-optimized copy.

This came up a lot in the alpha phase, more than we expected. It's on the roadmap, but there's no firm timeline yet. Check out our roadmap to see what we're working on and follow the progress in real time.

The Learning Loop lets you give the AI agent company-specific knowledge. For example: always give customer XY a 2% early-payment discount. This information is automatically taken into account the next time around.

FAQ pages, PDFs, product data sheets, website content, and other documents.

Yes, there's a guided onboarding the first time you log in to the Agent Hub.

Yes, that's possible via the Learning Loop.

There's no base fee. You only pay for actual usage (the AI tokens you consume). The model is based on three components: a service package (Standard/Growth/Premium), individual custom agents (as an add-on or included in the Scale package), and overage (pay-as-you-go for usage beyond your included quota). The specific costs will take shape over the course of the beta phase and will be based on average usage behavior.

All standard agents are available with no base fee. The included token quota depends on your service package. Higher packages include a larger allowance.

Token usage is visible live in the Agent Hub dashboard at any time. During the beta, this lets you get a feel for your natural usage without paying for it.

You have to actively agree to purchase additional tokens (overage). No amount is charged automatically.

The beta is planned to run until late August/early September 2026. After that, the general availability phase begins, at which point usage becomes paid.

No. Participating doesn't obligate you to move to the paid version.

Access is rolled out in groups of 20 to 50 users. After signing up, you'll receive a separate confirmation email with all the details and terms of use once it's your turn.

No. If you're already registered, you're on the list and don't need to sign up again.

Yes, taking part in the AI beta is independent of the general Xentral beta user group.

All AI features are continuously documented in the Help Center. Just type “AI” into the search field.

By default, every team member has access to the instance.

Yes. Users without the right permissions will see a message letting them know they need to have access granted first.

This is still an open question we're working through.

Yes. Multiple invoices in a single email get processed just fine.

Yes. Different attachment types are supported and used for processing.

Yes. Agents can pull data from various attachment types and use it to complete the task.

Yes. The delivery information agent pulls the relevant details from Xentral ERP and drafts a detailed response with delivery times, which you can then send to the customer.

Right now, DHL is supported. We're continuously expanding the list.

Since that's a rule-based process, Xentral Flows is the better fit for it.

Flows are for deterministic, rule-based processes. Agents come into play when the system needs to interpret context or meaning, like decoding cryptic or incomplete payments, handling emails with mixed content, or making judgment calls on individual customer decisions. We break down the exact difference and when to use which tool in the Xentral Flow Editor Guide.