Agents that work, and ask before they act.
An AI agent does not just answer questions, it completes multi-step tasks: preparing emails, looking up data, maintaining content. With us, it works with strictly defined tools, and anything that has an external effect is approved by a person.
- Fixed tools instead of a blank cheque
- Approval before external effect
- Every step logged
- Monthly cost limit
Example: customer enquiry by email
What agents do for our customers.
We have already built these agents. For you, we tailor them to your processes and systems.
Email assistant
Reads incoming emails from Microsoft 365 or Exchange, classifies them and writes a draft reply. Every reply is approved by a person, automatic sending is ruled out.
Information from production and merchandise management
Answers questions such as "What is the status of order 4711?" through a fixed list of queries. No free SQL, no write permissions, and the data comes from a replica, not from the production system.
Website assistant with tools
Answers guest questions from the content of the website and uses tools such as booking and tourism data or weather for this. As a ready-made product: AlpineChat.
Review queue instead of flying blind
If the agent is unsure about a classification, the case ends up in a review list for your team instead of the agent guessing.
How we keep agents within limits.
Fixed tools
An agent can only do what we give it as a tool: a specific query, a specific action. Everything else is technically impossible for it.
Approval before external effect
Emails, changes in systems, publications: whatever has an external effect, the agent only prepares. It is sent after approval.
Prompt injection protection
Agents read third-party content such as emails and documents. Instructions hidden in them are filtered, and the tool limits still apply.
Log
Every step, every tool and every approval is logged. You can trace at any time what the agent has done.
Cost limit
A monthly budget for model usage is checked before every call. Once it is reached, the agent stops instead of letting the bill grow.
Data protection
Personal data is masked before every model call. On request, sensitive agents run only with local models.
Let it prepare first, then trust it with more.
Define the task
In the AI potential workshop, we look for a task with clear steps and measurable effort, such as preparing replies to enquiries.
Pilot with approval
The agent prepares, your team checks every step. We measure how much time this saves and how often corrections are needed.
Expansion
What proves itself gets further tools or processes. The approval for external effect remains.
More from webx cortex.
AI knowledge management
Ask for company knowledge instead of searching for it, with source references and permissions.
AI funding Tyrol
Automation and AI integration are eligible for funding, the application must be submitted before the project starts.
Frequently asked questions
What is an AI agent?
An AI agent is a language model that does not just answer, but is allowed to use tools: query data, draft texts, prepare actions. This way it completes tasks with several steps. Which tools it has, we define.
What is an agent allowed to do on its own?
Read, look up, summarise and prepare. Anything that has an external effect, that is sending emails, changing data or publishing, is approved by a person with us.
What happens when the agent is unsure?
Uncertain cases end up in a review list for your team. The agent does not guess, and it does not invent tools it does not have.
What does operation cost?
We offer the implementation as a project. The ongoing model costs depend on usage and are shown separately. A monthly budget that is checked before every call prevents surprises.
Which task should an agent take off your hands?
Describe a process that costs you time today. We will tell you whether an agent can sensibly prepare it. The initial consultation is free of charge.
or by email to info@web-crossing.com · To the AI overview