AI agents & integration
AI agents that work
inside your systems.
An AI agent reads a request, looks things up in your systems and prepares the next step. You decide what it may carry out on its own.
Discuss your projectFree initial conversation · Directly with Daniel Gläser
- 01Request
- 02Plan + tool
- 03Approval
- 04Action in system
What an agent is
A model, tools and rules.
An AI agent combines a language model, tools and rules. The model understands the task and plans the steps. The tools give it access to your systems, such as a customer search, a calendar or creating a case. The rules decide which tool it may use when, and when a person takes over.
Unlike a chatbot, an agent does more than answer questions: it carries out steps. Technically this works through function calling. The model picks a defined function and supplies the parameters. Your software makes the call and checks permissions and inputs along the way. The model never touches a database directly.
Tool calls are supported by common cloud models, for example from OpenAI, Anthropic, Google or Mistral, and by many open models that run locally. The right model depends on the task, the type of data and running costs.
Examples
What an agent actually takes on.
Typical workflows with input and outcome. Each can start in read-only mode, producing drafts.
Inbox triage
Input: a customer email with a complaint and a photo. The email agent identifies the issue and its urgency, finds the customer and order in the CRM and saves a draft reply with the relevant details. A colleague reviews the draft and sends it.
Appointment and order intake
Input: a request via form or email, such as ‘servicing for two machines next week’. The agent asks for missing details, checks the calendar and proposes two slots. Nothing is booked until someone confirms.
Look up data, open a case
Input: ‘Where is delivery 4711?’ The agent reads status and delivery date from the ERP or inventory system and answers the question. If the date has slipped, it opens a case with all the facts. Which fields it may write to is defined in advance.
Phone reception with forwarding
An AI phone assistant answers calls when nobody is free. It says up front that the caller is speaking to an AI, records the request and a callback number, and forwards urgent calls to the right person. Each call leaves a short note in your system. Voice data and recordings get a separate data protection review.
Assistant, workflow, agent
Assistant, workflow or agent?
Not every task needs an agent. The simpler option is often the better one.
AI assistant
Answers questions from your knowledge base.
- Reads documents, takes no actions.
- Suits knowledge search and customer support.
- Low risk, quick to start.
Workflow
Fixed steps following fixed rules.
- Every case runs through the same sequence.
- AI handles individual steps, such as classification.
- Easy to test and inexpensive to run.
Agent
Plans its own steps and uses tools.
- For cases that look a little different each time.
- Needs permissions, approvals and logs.
- More testing effort and higher running costs.
Integration
Connected to your software.
Existing software via API
Many CRM, ERP and ticketing systems offer an API. I connect the agent through it, using its own technical account with only the permissions it needs. Where there is no API, we look at alternatives such as read-only database access, an export or a small intermediate service.
MCP servers for internal systems
The Model Context Protocol (MCP) is an open standard that lets AI applications connect to tools and data sources in a consistent way. I build MCP servers for your internal systems, such as an order database or document storage. Authentication and permission checks live on the server, so different AI applications can use the same tools under control.
AI inside your own applications
I develop management systems with AI integration, where suggestions and checks appear directly in the interface. One example of my own work is the iPhone app Repp, which can create a training plan with AI on request.
See ReppNew software built with AI in mind
If an existing system cannot be extended sensibly, a dedicated application may be the better foundation. Agent, data model and interface are then planned together.
Software developmentPermissions & control
What an agent may do is set in advance.
Read, suggest, execute
Read means the agent looks things up and changes nothing. Suggest means it prepares drafts that a person adopts. Execute means it acts on its own, but only for tightly limited tasks. We usually start with read and suggest.
Approvals
Actions with consequences, such as sending to customers, postings or deletions, wait for confirmation. Uncertain or unusual cases are handed to a person along with the agent's reasoning.
Logs
Every tool call is logged with time, input and result. That shows why a case was created and makes errors easy to trace. We also agree who may view the logs and how long they are kept.
Cost control
Language models are usually billed by processed tokens. Token limits per request and per day, a cap on steps per task and a smaller model for simple sub-tasks keep costs predictable. Usage is visible at any time during operation.
Testing with real cases
Before launch we test against a collection of real, anonymised cases from your daily work, including unclear and faulty inputs. The same cases run again after every change to prompt, model or tool.
Data protection and transparency
The model only receives the data needed for the task at hand. Anyone writing to or speaking with an agent is told they are dealing with an AI system. Providers, data flows and storage locations are reviewed for the specific use.
Guide to AI and GDPRLimits
Where agents reach their limits.
Language models can misread a task or invent facts. The more steps an agent takes on its own, the more places something can go wrong. That is why workflows stay short, tools stay narrow and results stay checkable.
Emails and documents can contain hidden instructions, known as prompt injection. An agent must never treat such text as a command. Permissions are therefore enforced in the software, not merely described in the prompt.
I do not leave decisions with serious consequences for people, such as hiring or creditworthiness, to an agent. The EU AI Act sets additional requirements for these areas. AI can organise the paperwork; a person makes the decision.
Working together
A clear path to everyday use.
Choose a workflow
Pick a frequent task with a clear input and a clear outcome, and agree success criteria.
Define tools and permissions
Decide what the agent may read, suggest or execute, and collect real test cases.
Pilot in suggestion mode
The agent works on real cases while a person adopts every result. Quality and costs are measured.
Extend step by step
Release proven steps, review the logs and connect further tools.
Before we begin
Good questions. Clear answers.
How is an AI agent different from a chatbot?
A chatbot answers questions. An agent also uses tools: it looks things up in your systems, prepares drafts or opens cases. That requires defined permissions, approvals and logs.
What does it cost to have an AI agent built?
Cost depends on the number of connected systems, the quality of their APIs, testing effort and ongoing model usage. We start with a bounded pilot for one workflow, with the effort agreed in advance. Running costs can then be estimated from real usage.
Can an AI agent work with my existing software?
Usually yes, if the software offers an API or another controlled way to access data. I check this beforehand using the documentation and a test account. Without an API, exports or an intermediate service are possible, with some limitations.
What is MCP and do I need it?
The Model Context Protocol is an open standard for connecting AI applications to tools and data sources. It pays off when several AI applications should use the same internal systems. For a single agent, a direct connection via function calling is often enough.
Can the agent send emails on its own?
Technically, yes. I only set this up for narrowly defined, explicitly approved cases such as acknowledgements of receipt. Replies with real content remain drafts until a person has checked them.
Which data does the language model see?
Only what the task requires and what the approved tools retrieve. Whether a cloud model under a data processing agreement or a local model fits is decided by the type of data. This is reviewed for the specific use.