AI agents
AI that doesn't just answer. It takes on tasks.
An AI agent gathers information, checks cases and prepares actions. Your team approves what matters.
- Human approval
- Minimal access rights
- Every action logged
Who this service is for
For processes that occur often, can be described clearly and involve data from several systems or unstructured content such as emails and documents. For rare, inconsistent processes the effort is usually not worthwhile.
What we deliver
- One agent for a selected process with defined tasks, limits and approval points
- Connection to the agreed systems with separate read and write permissions
- Access to documents and data you approve, answers with source references
- Action log: read, suggested, approved, by whom
- Test cases from your day-to-day work and a pilot with a small group
- Documentation of data flows and the service providers involved
Costs
- One-off
- No flat rate. One-off costs cover design, system integration, tests and the pilot; the quote follows after the intro call.
- Ongoing
- Ongoing: model and service usage based on consumption, licenses, hosting and, if applicable, support and adjustments. Shown in the quote separately from the setup.
- Depends on
- Number of connected systems, State of the data and interfaces, Volume of processed cases, Number of approval points, Scope of support
What we need from you6
- Selection and description of the process, plus a contact person from the department
- Access to the systems concerned (CRM, ticketing, email, storage), preferably via interfaces
- Approval of which data the agent may use
- Decision on which steps it handles on its own and which need approval
- Real example cases for testing
- Employees who give approvals and feedback during the pilot
Requirements and limits5
- Mistakes cannot be ruled out; for critical results, a person checks.
- Without a usable interface or clean data, an agent will not get far.
- It does not replace expert decisions; it prepares them.
- In uncertain or unforeseen cases, it stops and hands over to a person.
- The data protection assessment is the responsibility of your data protection adviser.
The difference
A chatbot answers. An agent does the groundwork.
AI agents for businesses are programs that carry out several steps in sequence: gathering information from emails, documents and CRM, preparing a case, creating a draft. You define beforehand what an agent may read and change, and at critical points it waits for a person's approval.
Chatbot
Answers questions in conversation
- question
- Answer
AI agent
Works through a task step by step
- 1
Task
A trigger or a task starts the process.
- 2
Check information
Emails, documents and systems that you approve.
- 3
Prepare a decision
Classify, summarize, suggest options.
- 4
Action by rules
Create a draft or entry, within its permissions.
- 5
Result
Result with log; critical steps wait for approval.
A human decides at critical points
SimplifiedSimplified illustration. Actual systems differ in scope of functions.
01
Knowledge is scattered everywhere
Some of it is in the mailbox, some on the drive, the rest in the CRM. Just gathering it often takes longer than the actual work, and anyone who doesn't know the project won't find it at all.
02
Groundwork eats up the working day
Preparing quotes, creating tickets, updating data, summarising conversations. It has to be done, but it is rarely what you hired your people for.
03
A chatbot is not enough
It answers questions, but it does not get anything done within your processes. In the end someone copies the answer into the right system by hand anyway.
04
The worry about losing control
If an AI sends emails or changes data on its own, it must be clear what it may do and who oversees it.
Example
Handling a customer request
From inbox to approval: this is how an agent could prepare an incoming request. Click through the steps.
Maintenance contract request
Hello, we would like a quote for the maintenance of two systems starting next month. Could you send us something?
2Identify the request
3Check information
4Structure data
5Prepare next step
6Employee approves
ExampleFictitious example for illustration, not a real customer request.
Which process of yours comes up all the time?
In the intro call we check whether it suits an agent and what that would require.
Capabilities
What an agent can take on
Research in your data
The agent searches documents and systems that you approve and answers with source references. That way, any information can be verified.
Recognise and classify incoming items
Invoice, order, complaint or enquiry: the agent recognises the type of document, extracts the key details and files it in the right place.
Prepare tickets and cases
A long customer email is summarised, fields are pre-filled, and the case goes to the responsible team. There, someone sees right away what it is about.
Quote drafts based on your rules
Building blocks, prices and terms come from your templates. The draft is ready, and you review and approve it before anything leaves the company.
Suggest CRM entries
After a conversation, the agent suggests notes, status and follow-up. A person confirms the change, and only then is it written.
Hand over when unsure
If a case does not fit the rules or the agent is not sure, it stops and hands over to a person with all the information gathered so far.
Architecture
How an agent system is built
Which systems an agent can reach depends on the available interfaces and access rights. Often what you already use is enough. Where no interface exists, we tell you in the intro call.
Trigger
What starts the agent
- Incoming email
- New ticket
- Form
- Schedule
- Request in team chat
Knowledge sources
What it may read
- Document storage and knowledge bases
- Email and calendar
- Spreadsheets and databases
read-only
Agent logic
Task, limits, sequence
- Understand the request
- Plan steps
- Compare sources
- Stop when uncertain
Tools
What it may trigger
- CRM and ticketing systems
- Your own software via APIs
writes, by rules
Security and approvals
Rules enforced by technical means
- Minimal permissions
- Approval for critical steps
- Log of every action
- Abort rules
- Handover to a person
Result
What is ready at the end
- Draft or proposal
- Case created
- Summary with sources
- Log entry
Use cases
Where agents help in everyday work
Who it's for
- Sales and service teams with many recurring processes
- Companies with a lot of paperwork and document handling
- Teams whose knowledge is spread across mailboxes, drives and systems
- Managing directors who want to use AI without giving up control
- 01
Sales assistance in mechanical engineering
Before the customer meeting, the agent compiles the email history, open quotes and latest notes. After the meeting, it suggests the meeting note and the follow-up.
Agent preparesHuman decides
- 02
Internal knowledge assistant at a tax advisory firm
What was the rule for travel expenses, what does the client handbook say? The agent answers from the internal documents and names the source.
Agent preparesHuman decides
- 03
Support triage in online retail
Incoming requests are sorted by type and urgency, with order data already looked up. Complaints and special cases go straight to a person.
Agent preparesHuman decides
- 04
Quote preparation at a trade business
A request with measurements and wishes becomes a draft based on your price list. The master craftsperson reviews it, adjusts it and sends it out personally.
Agent preparesHuman decides
Limits and security
Permissions, approvals, log
We define an agent's limits before its first use and enforce them technically, not just in the instructions to the model.
Minimal rights
the agent gets only the access its task requires, read or write per system
Human approval for critical steps such as sending, pricing or changes to customer data
Actions are logged
what was read, what was suggested, what was approved and by whom
Fixed termination rules and a clear path to hand over to a person
Data flows and the service providers involved are documented before the pilot; you clarify data protection details with your data protection adviser
Honestly speaking
What an agent can't do
- Mistakes cannot be ruled out; for critical results, a person checks.
- Without a usable interface or clean data, an agent will not get far.
- It does not replace expert decisions; it prepares them.
- In uncertain or unforeseen cases, it stops and hands over to a person.
- The data protection assessment is the responsibility of your data protection adviser.
Process
From the first process to live operation
We start small, with one process and one pilot, and only expand once it proves itself.
- 1
Choose a process that occurs often and can be described clearly
- 2
Review data, systems and permissions
- 3
Design the agent: tasks, limits, approval points
- 4
Run through test cases from your real day-to-day work
- 5
Pilot with a small group and full logging
- 6
Evaluate continuously and refine the rules
Start small with a pilot
You see early on whether it pays off. We give costs after the intro call, with setup and running costs listed separately.
Result
What you get out of it
Routine work is ready and prepared, instead of being done by hand first
You find faster what knowledge already exists in-house
Processes are traceable because the actions are in the log
Your team keeps the decision and gains time for it
In the pilot you see on a small scale whether the use pays off before you expand
Actual results depend on the starting point, scope and implementation. We do not give guarantees of success.
Concept studies
Concept studies on AI agents
Detailed scenarios for typical starting points and possible approaches. They are not reports on client projects.
- Scenario · not a client project
Quote drafts · 14 min
The quote still sitting on the desk on Friday evening
An AI agent reads the enquiry, asks about whatever is missing and builds the quote draft from your building blocks, but never sends anything without human approval.
Read concept study → - Expert analysis · not a client project
Agent security · 14 min
What an AI agent must never do alone, and how you enforce that
An AI agent is only as safe as the rights you give it: if you separate access, define approval levels and enforce limits technically, you limit the damage an error can do.
Read concept study → - Concept study · not a client project
Choose the pilot process · 13 min
Which process should go to the AI agent first?
The first process decides whether an AI agent gains trust in the business or becomes a permanent stopgap, and it can be chosen cleanly with a few criteria.
Read concept study → - Concept study · not a client project
Inbox triage · 13 min
The complaint was number 31: how an agent presorts the inbox
An agent that classifies incoming emails, ranks them by urgency and suggests replies saves sorting work, as long as it hands over to people when it is uncertain.
Read concept study → - Expert analysis · not a client project
Agent operations · 13 min
The agent did something, and nobody can say why
An agent that acts on its own needs a trace of every step, alerts for typical failure patterns and the ability to roll back a change in an orderly way.
Read concept study → - Scenario · not a client project
Document research · 14 min
The answer is somewhere in the folder, just not where people look
An agent that searches internal documents is only as useful as its source citations, its permission checks and the way it handles contradictions.
Read concept study →
FAQ
Frequently asked questions
What are AI agents for businesses and how do they differ from chatbots?
A chatbot holds a conversation and answers questions. An agent can also carry out several steps in a row and work with your systems: retrieve data, classify a document, create a draft or an entry. This happens within the rules and permissions that we define with you beforehand.
Can an agent act on its own?
Only where you allow it. Non-critical steps such as summarising or classifying can be handled on its own. For sending, pricing or changes to customer data, we clearly recommend human approval. Which steps run freely and which wait is set out in the design and can be changed later.
How secure is access to our systems?
The agent gets its own access with minimal rights, separated into read and write. What it should not see, it technically cannot see. In addition, every action is logged. There is no guarantee against every risk, but permissions and the log make misbehaviour containable and traceable.
How are mistakes handled?
Before the pilot we write test cases from your day-to-day work, log every action and define when the agent aborts and hands over to a person. Mistakes can never be ruled out entirely. That is why a person checks critical results, and during operation we review conspicuous cases together.
What data does the agent use and where is it stored?
Only sources that you explicitly approve. Which data flows where and which service providers are involved, we document before the pilot. Whether and how personal data may be processed is a question for your data protection adviser. We prepare the information for that in an understandable form.
Which systems can be connected?
Anything with a usable interface: CRM, ticketing system, email, calendar, document storage or your own software. For older programs without an API we check alternatives or tell you if the effort is not worthwhile. What is possible in your case, we clarify in the intro call.
When does a pilot pay off, and how long does it take?
When a process comes up often, can be described clearly, and the data is available. The timeline depends on systems and scope, so we only give one after the intro call. Costs are not flat either: the quote lists the one-time setup and the ongoing costs (model usage, licenses, support) separately. In the pilot, a small group works with the agent, and you see on a small scale whether the effort pays off.
How does this differ from classic automation?
Automation follows fixed rules: if A, then B. That is reliable, but rigid. An agent can handle unstructured content such as emails or documents and decides, within the scope of its task, which step makes sense next. Often the two are combined: fixed workflows for the standard business, an agent for reading and classifying.
Which tasks could an AI agent take over in your company?
Describe a recurring process to us. We will tell you honestly whether an agent is a fit.