AI agent workshop:Your team builds an agentit can also test
In one day from a real workflow to a tested prototype, with clear approvals and an honest answer to the question of whether it is worth pursuing.
Get in touch- What is an AI agent workshop, and how does it differ from a training course?
- How does an AI agent workshop run over one day?
- Which tasks suit the first agent in an AI agent workshop?
- How do you know the agent from your AI agent workshop is good enough?
- Which data belongs in an AI agent workshop, and what helps against prompt injection?
- What should an AI agent workshop clarify about the AI Act and Article 50?
- What does an AI agent workshop cost?
- Where does the prototype go after the AI agent workshop?
- Frequently asked questions
- How to prepare your team for an AI agent workshop
- Where the information on this page comes from

Let's talk about your project.
First we check whether the project fits your business model. Then you get a proposal with phases and effort.
Enquire about an AI agent workshop or call: +49 151 1576 5566An AI agent workshop is usually a hands-on day in which a team takes one of its own workflows apart, builds an agent for it on test data and checks the result. By the evening you should have three things: a prototype, test cases with criteria agreed in advance, and clear rules on what a person signs off. Ask every provider, including us, about exactly that last part: How do you know the agent is good enough, which data may enter the room, and who is responsible for it the day after? We build and use workflows like these in our own agency work, including where they get things wrong, so this article is less about tools and more about what decides whether the day pays off.
- A good AI agent workshop works on a real workflow from your team, using test data rather than real correspondence.
- On day one a fixed workflow with a model step is often enough. A full agent only pays off when the steps cannot be set in advance.
- Test cases are written before anything is built, otherwise the team ends up judging an impression.
- Prices for an in-house day vary widely: the German pages we read list figures from around 1,900 to over 6,000 euros, mostly net and often plus travel.
On this page
- What is an AI agent workshop, and how does it differ from a training course?
- How does an AI agent workshop run over one day?
- Which tasks suit the first agent in an AI agent workshop?
- How do you know the agent from your AI agent workshop is good enough?
- Which data belongs in an AI agent workshop, and what helps against prompt injection?
- What should an AI agent workshop clarify about the AI Act and Article 50?
- What does an AI agent workshop cost?
- Where does the prototype go after the AI agent workshop?
- Frequently asked questions
- How to prepare your team for an AI agent workshop
- Where the information on this page comes from
What is an AI agent workshop, and how does it differ from a training course?
An AI agent workshop takes on exactly one workflow, whereas an AI training course covers how a team works with AI in general, which tasks suit it and what is allowed. The workflow is broken into steps, built and tested until everyone involved can see where the agent helps and where it fails.
What an AI agent actually is becomes clear quickly through a distinction drawn by the AI company Anthropic. In a workflow, predefined code paths orchestrate the model and tools; in an agent, the model directs its own process and choice of tools. Anthropic explicitly advises starting with the simplest solution, which may mean not building an agent at all. That is good news for a workshop, because some teams leave with a small fixed workflow and are better off than with a large agent nobody wants to supervise.
The day works best with a mixed group, by the way. The header image is AI-generated and does not show one of our workshops. People who run the process every day know the exceptions, and developers see where interfaces and permissions will cause trouble. With only developers in the room you often get clean engineering for a task that does not really exist in that form.
How does an AI agent workshop run over one day?
In an AI agent workshop, one full day is enough for a prototype but not for production, and everyone should know that from the start. This is how we would structure the day, with the timings fitted to the venue and its breaks.
The block before lunch matters most. Before anybody builds, one page states where the workflow starts, which information it needs, where a person decides and which cases the agent has to pass. Teams that shortcut this tend to spend the afternoon arguing about what the result should have been.
| Block | What happens | Result |
|---|---|---|
| Opening | The team's workflows, expectations, approved AI services | One workflow chosen for the day |
| Foundations | Assistant, workflow and agent, each shown on running examples | Shared vocabulary |
| Breaking it down | Steps, data sources, decisions, risks per step | Workflow sketch and first test cases |
| Building | Mixed groups implement the workflow on test data | Working prototype |
| Testing | Run test cases, hunt for errors, simulate an attack | Hit rate and error list |
| Safeguards | Approvals, permissions, logging, stop switch, legal position | Rules for the workflow |
| Closing | Continue, change or stop, with names and dates | Decision and next steps |
Which tasks suit the first agent in an AI agent workshop?
The best task for an AI agent workshop comes up often, has several steps, and its errors can still be caught by a person before anything goes out. Reading steps are harmless. It gets delicate the moment the agent writes, sends or books. The table shows how we rate typical workflows, which is our judgement from working with such systems.
Ask early whether the target system has an API. If it does, use it. An agent that operates screens is meant for software without an interface, and it is slower and harder to secure.
| Workflow | Fit for the day | Where the person stays |
|---|---|---|
| Categorising incoming emails | Very good, a fixed workflow is often enough | Checks samples and all uncertain cases |
| Drafting replies from past emails and documents | Very good, also shows the limits of document search | Reads and sends personally |
| Updating contact details in a spreadsheet | Good, if the agent only proposes changes | Approves changes |
| Preparing an event, gathering documents and deadlines | Good, many steps and little risk | Decides on dates and commitments |
| Preparing a travel request | Good up to the finished draft | Books, or approves each booking |
| Screening applicants or allocating tasks by behaviour | Not for the workshop, may count as high-risk | Stays entirely with people |
How do you know the agent from your AI agent workshop is good enough?
The agent from your AI agent workshop is good enough when it meets a standard you set before it ran. That sounds obvious. Even so, on the four provider pages we read we found no mention of test cases set in advance, which of course says nothing about what happens in the workshop itself. To begin with, 20 to 30 realistic cases will do, recreated or verified as anonymised, explicitly including the awkward ones: the unclear email, the missing detail, the case in which the agent should do nothing at all. For each one, the correct outcome is written down in advance.
Anthropic's guide to evaluating agents explains why this matters. Without such checks, teams react to complaints and cannot tell real regressions from noise. Since language models can answer differently to the same input, each case should run several times. One successful run in a demo proves very little.
Set an error threshold too. A misfiled email is tolerable if a person sees the uncertain cases; a wrong commitment to a customer usually is not. That threshold determines what the agent may do alone and when it stops. Measuring cost at the same time, per task rather than per token, saves a nasty surprise later, as we show in our piece on LLM costs.
Which data belongs in an AI agent workshop, and what helps against prompt injection?
Real correspondence stays out of the AI agent workshop. Practice happens on test data or properly anonymised samples, and before the day the team clarifies which AI services are approved and where the model runs. For organisations with their own or locally hosted models, the workshop should be able to run on those, otherwise people practise with something they cannot use afterwards.
One feature of many agent workflows is often overlooked. An agent that reads emails or documents also reads whatever a stranger has written into them. Germany's Federal Office for Information Security (BSI) calls such indirect prompt injections an intrinsic weakness of application-integrated language models, and OWASP ranks prompt injection first on its list of risks for AI applications. It also says that document search does not fully solve the problem. Since December 2025 OWASP has published a separate list for agentic applications.
That is why a small attack belongs in the day: a test email telling the agent to pass on data or change something. After that, even the last sceptic understands why we keep permissions narrow, approve tools one at a time and, for now, have a person check everything outgoing.
What should an AI agent workshop clarify about the AI Act and Article 50?
An AI agent workshop should clarify three points above all, and all three have moved recently. Article 4 was softened in 2026 by the Digital Omnibus. According to Germany's Federal Network Agency, no particular level of AI literacy has to be guaranteed for individuals; what is required is measures. A workshop is therefore a sensible measure you can document, not a compulsory course with a certificate (more in our piece on the AI literacy obligation).
Second, Article 50, which has applied since 2 August 2026. The European Commission's non-binding guidelines of 5 November 2026 name agents directly, including those making bookings or managing correspondence. Such agents should disclose that they are AI and on whose behalf they act. No notice is needed for a text that a person, as the main interlocutor, has properly reviewed and sent, although the mere possibility of review is not enough. Towards the person instructing it, the agent should still show that it is AI unless that is obvious, and anyone developing and deploying in-house may count as a provider themselves.
Third, the upper boundary. High-risk obligations under Annex III apply from 2 December 2027, and they cover things like filtering job applications or allocating tasks based on individual behaviour. In Germany, section 90 of the Works Constitution Act also requires employers to inform the works council about work processes, expressly including the use of artificial intelligence, and it is worth checking whether co-determination under section 87(1) no. 6 applies, particularly if a log records who approved what. This is our reading and not legal advice, so your organisation's responsible office should look at it beforehand.
What does an AI agent workshop cost?
For an AI agent workshop there are no market statistics, only what providers publish. On the German pages we read on 5 November 2026, prices run from 3,200 euros with a standard agenda to 4,900 euros with your own process, both net and plus travel on site. Others list 3,490 and 4,500 euros net, partly for general AI workshops, one provider starts at 6,000 euros, and another estimates the market at 1,900 to 4,300 euros. Open seats and webinars run at roughly 500 to 1,100 euros per person (some gross, some net), but they are only loosely comparable because nobody works on your own workflow there.
When comparing, ask above all what the price includes. A preparatory call and preparation for your workflow belong in it, as does the question of whether you keep the prototype and test cases. Ongoing model costs and later integration with production systems are usually not included. Separately, Gartner expects over 40 percent of agentic AI projects to be cancelled by the end of 2027, partly because of rising costs, unclear value and inadequate risk controls.
| Format | Published price | What it is good for |
|---|---|---|
| In-house workshop, one day | Listed at around 1,900 to over 6,000 €, mostly net, often plus travel | Prototype and decision within your team |
| Open course or webinar | Roughly 500 to 1,100 € per person, some gross, some net | Getting started as an individual |
| Including set-up of own infrastructure | Considerably more, one provider lists 9,000 € plus VAT | Technical teams intending to operate it |
Where does the prototype go after the AI agent workshop?
The day after the AI agent workshop, the prototype is evidence and not yet a service. To turn it into one, it needs a person responsible for it, a stop switch someone is actually allowed to use, and a log showing which model did what with which input and who approved it. A usable field list is in our article on logging requirements.
A limited trial makes sense: one team, one mailbox or one event, two to four weeks, the same test cases after every change. If the hit rate holds, widen the permissions step by step. If it does not, the workshop was still worthwhile because the team now understands the workflow far better. Anyone who then wants a permanent AI agent system starts from a much better position.
Frequently asked questions
How long does an AI agent workshop take?
A full day is a realistic frame for a prototype with tests. Three to four hour formats exist as well, and those are more about orientation and picking use cases than building.
How many people should attend an AI agent workshop?
Providers mostly mention groups of up to about eight or twelve. To make sure every group really builds, we prefer small mixed teams of business and technical staff.
Does our team need programming skills?
Not everyone. Business staff describe the workflow, exceptions and test cases; developers handle interfaces, permissions and evaluation. Both should work on the same workflow, or you end up with two results that do not fit together.
Can we use the tools and models we already have?
That should be the rule. It is settled in the preparatory call, including access and the internet connection at the venue. A plan for a service going down is part of it.
Is the workshop available in English?
Yes, we also offer the whole day in English, including the materials. For international teams it is often the one language in which everyone can contribute equally.
Does an AI agent workshop replace AI training under Article 4?
It can be part of your measures and can be documented. For everyone else in the organisation using AI, broader training usually still makes sense because the workshop covers only one workflow.
How to prepare your team for an AI agent workshop
- 01
Pick a workflow
Choose a task that comes up often and whose errors can be caught before anything goes out. Write down where it starts and where it ends.
- 02
Collect examples
Put together 20 to 30 realistic cases, recreated or verified as anonymised, including difficult ones, and note what would be correct for each.
- 03
Clarify the framework
Ask data protection, IT security and any staff representatives which services and data are allowed. Check the internet connection at the venue as well.
- 04
Name who is responsible
Decide in advance who will choose whether to continue after the day and who will look after the prototype.
Our impression from our own workflows: an agent does not become good by working in the demo. It becomes good when your team knows when it is wrong and what happens then.
AI agentsAgentic AI Explained: What It Means for Your Business
Google spam update recoveryGoogle Spam Update Recovery: Find the Cause Before You Fix
AI agent vs chatbotAI Agent vs Chatbot: The Difference and When Each One Fits
SEO vs GEOSEO vs GEO: What Stays the Same and What Changes
Zero-click searchZero-Click Search: What the Studies Show and What to Do
Google indexingRequest Indexing on Google: What the Button Really Does
AI Act loggingAI Act Logging Requirements: What Applies and When
SEO visibility indexSEO Visibility Index: What It Measures and What It Misses
RAGWhat Is RAG in AI? Retrieval Augmented Generation Explained
Gemini 4 ArgonGemini 4 Argon: What the Model Is Actually Built For
Computer useComputer Use Agent: What It Operates and What It May Do
AI voice agentAI Voice Agent Pricing: Cost per Call and Disclosure
Google spam updateGoogle Spam Update: What Can Your Numbers Really Tell You?
LLM costsLLM Cost Optimization: When a Model Switch Pays Off
Detect AI-written textDetect AI-written text: what AI detectors get wrong
Google AI ModeGoogle AI Mode: What It Means for Your Website
AI agentsWhat Is an AI Agent, and How Is It Different From a Chatbot?
Structured DataStructured Data: What It Really Does for AI Search
AI AssistantAI Assistant for Business: Types, Uses and Data Protection
GEOE-E-A-T: Trust Signals for Google and AI Search
Local AILocal AI for Business: What “Local” Really Means
GEOAI Crawlers in robots.txt: Managing GPTBot and Co.
AI for SMEsAI for SMEs: How to Introduce AI Step by Step
Perplexity SEOPerplexity SEO: How to Get Your Site Cited as a Source
ChatGPT SEOChatGPT SEO: How to Get Your Business Found in ChatGPT
llms.txtllms.txt: What the File Does and When It Pays Off
AI SEOAI SEO: What Changes Compared to Traditional SEO
AI OverviewsGoogle AI Overviews: How Google Picks Its Sources
GEO vs AEO vs LLMOGEO vs AEO vs LLMO: The AI Search Terms Explained
Measure AI VisibilityMeasure AI Visibility: Method, Metrics and Limits
ChatGPT AdsChatGPT Ads: How to Advertise on ChatGPT in Germany
AI Literacy ObligationAI Literacy Obligation: What Article 4 Requires Since 2026
AI DisclosureAI Chatbot Disclosure: Article 50 in Practice
AI Text WatermarkAI Text Watermark: What Claude Marks and What It Doesn't
Where the information on this page comes from
- Anthropic: Building effective agentsaccessed 5 Nov 2026
- Anthropic: Demystifying evals for AI agentsaccessed 5 Nov 2026
- BSI: Indirect prompt injections, intrinsic weaknessaccessed 5 Nov 2026
- OWASP: LLM01 Prompt Injectionaccessed 5 Nov 2026
- OWASP: Top 10 for Agentic Applicationsaccessed 5 Nov 2026
- Bundesnetzagentur: AI literacyaccessed 5 Nov 2026
- European Commission: Guidelines on Article 50, C(2026) 7702accessed 5 Nov 2026
- AI Act Service Desk: AI agents in the AI Actaccessed 5 Nov 2026
- Regulation (EU) 2026/1744 (Digital Omnibus)accessed 5 Nov 2026
- Section 90 Works Constitution Act (BetrVG)accessed 5 Nov 2026
- Gartner: Over 40 percent of agentic AI projects canceled by end of 2027accessed 5 Nov 2026
- souveraign: 3,200, 4,900 and 9,000 € plus VAT, travel on site extraaccessed 5 Nov 2026
- Never Code Alone: KI-Workshop, 4,500 € netaccessed 5 Nov 2026
- Jean Hinz: KI-Workshop, 3,490 €, market estimate 1,900 to 4,300 € plus VATaccessed 5 Nov 2026
- Lucid Labs: KI-Workshop, fixed price from 6,000 €accessed 5 Nov 2026


