AI agents for businesses
AI agents that work with your existing systems
A customer asks when their spare part will arrive. Your service team needs to find the order, check what the warehouse knows, and leave the next person a clear record.
I build AI agents that can look up information, suggest a reply, or update a record after someone approves the change.
You set the limits on what the agent can read and change. I’m Marcus Biel. I lead the project, write code alongside your developers and bring in specialists when the work calls for them.
Give your agent a specific job
Start with the result you want. That might be a service request ready for review, an updated record or a task passed to the right person with the information they need.
The agent gets access to selected functions in your systems. It can use these tools to look up customers, check order status or prepare the next step in a case.
Within that scope, the agent can use the request and its findings so far to choose what else to look up and which tool to use.
Together we define what counts as complete and when a team member needs to take over.
Follow a service request through to the updated ticket
A customer wants to know when a spare part will arrive. The agent can check the order and draft a note for the service team.
In this browser-only example, check the agent’s draft, approve the exact wording and status change, then read the ticket again. That last step shows what the ticket system actually saved. All records are fictional.
Customer request · Ticket S-2048
“When will my spare part arrive?”
- Order A-821 · read-only access to the order system
- The part is ready to ship. The carrier hasn’t picked it up, and there’s no confirmed delivery date.
Last recorded ticket details
- Ticket status
- New
- Internal note
- No note has been added yet.
Agent’s proposal
Review the change to S-2048
Change the status from New to Under review and add the following internal note to the ticket.
Your approval applies to this status and this version of the internal note. The agent can update the ticket. It has no permission to change the order or email the customer.
From incoming request to verified change
The agent reads order data. In this example, its permission to make changes covers just two fields on the service ticket.
The history of S-2048 tells you who did what. The agent read A-821 and drafted the note. You, acting as service lead, approved its wording. The ticket system didn’t acknowledge the write. Reading the ticket again is how you verify the saved note and status.
-
Classify the request
The agent reads the incoming request and checks whether it’s a case it’s allowed to handle.
If it can’t tell what the customer needs, it asks a follow-up question or passes the case to a colleague.
-
Retrieve the information needed
Depending on the request, the agent retrieves relevant customer and order data or searches the approved product documentation.
The tools and their permissions determine which queries the agent can run.
-
Prepare a proposal
Using the information it found, the agent proposes how to proceed. That might include a summary, a draft reply and the intended changes to the ticket.
The reviewer can see which information the agent used and which actions are waiting for approval.
-
Review and approve the proposal
A reviewer can correct details, edit the draft, and approve the changes they see.
Approval applies to the specific changes and content shown.
-
Carry out the approved actions and show the result
After approval, the application updates the ticket through its existing API. The agreed scope may also include sending the approved reply through your service system.
Your team sees which systems confirmed a change and which actions still need attention.
How AI and business rules work together
I identify the parts of a workflow that need interpretation and the parts governed by fixed rules.
AI can interpret a free-text request, bring together relevant information or propose a response.
The application checks required fields, permitted status changes, user permissions and approvals.
That lets the agent help with the wording or interpretation while the application enforces the rules for changing a record.
More on process automationGive the agent the permissions its task requires
The agent gets access only to the tools and data its task requires.
Before I connect the agent, I confirm with your team what it’s allowed to do.
- Data access. Which customers, cases and documents may the agent use?
- Permitted changes. What information can it add or change?
- Approvals. Which actions need someone’s approval?
- Scope. When is the task complete, and when should a team member take over?
- Usage limits. How long may a task run, and how many calls may it make?
The application enforces these limits while the agent works, including checking that an action applies to the right customer, user and record.
See what actually happened
Staff need to be able to tell how far each case has progressed.
The activity log shows which sources the agent used, what it proposed, who approved it and what each connected system reported.
I design for timeouts, too. The workflow checks whether the original change went through before anyone tries again. If the outcome is still unclear, a person takes over.
Your team can see what has happened and what still needs attention.
Start with one defined workflow
A useful first project focuses on one kind of case, with a specific set of data sources and actions.
For example, I can start with one type of request, connect your ticket system and supply the customer information it needs.
I work through five questions with your team.
- What starts the workflow?
- What should happen by the end?
- Which tools and data can the agent use?
- Which decisions stay with your team?
- How can someone check the result?
You receive a working agent in the agreed environment, the necessary integrations, an evaluation of the test results and technical documentation.
I help you decide which actions are ready for the pilot and which to try in a test system first.
Plan a pilot with an AI agentDid the change actually happen?
I test cases your team recognizes from its own work.
For example, after a ticket update I check the target system to verify that the right change reached the right ticket. I also review the information the agent used and whether it respected the required approvals.
Test cases also include incomplete requests, missing data, conflicting information and interrupted system calls.
The pilot measures
- how often the agent completes the task
- where a person has to correct it or take over
- whether it respects permissions and approvals
- how long a case takes
- what it would cost at the volume you expect
With those figures, you can decide whether to roll it out or add another kind of request.
Integrate with your existing systems
Agents need access to the applications involved in the work. These might include CRM, ERP, ticketing, document management or custom business applications.
I review the available APIs and build the connections, including data mapping, permissions and the handling of acknowledgments.
More about APIs and system integrationWhat it costs to build and run an agent
The work depends on the task, the systems the agent needs to use, the changes it’s allowed to make, and how to verify them.
To estimate operating costs, I consider case volume, model choice, the number of system calls and the infrastructure required.
I work with your team to assign responsibility for operations, changes, and cases the agent couldn’t finish.
Who leads the work?
I’m Marcus Biel, a software architect and developer. I’ve been building Java applications since 2001.
I’ll design and build the agent with your team. If we need specialist help with an integration, I can bring it in.
We’ll work out who approves each action and who looks after the workflow once it’s running.
More about working with Marcus
Where could an AI agent help your team?
Tell me about one recurring task, the systems your team uses, and what a successful outcome looks like.
We can use that to decide which steps an agent could handle in a first release.
Prefer to talk first? Call +49 171 1151013 or message me on LinkedIn.