Blog · AI agents · 5 Nov 2026
Blog
AI agents

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
Enquiry
Nikolai Schöbel und Jeremias Burger, Co-Founder Scalableloops

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 5566
Blog · AI agents

An 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.

In brief
  • 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.
Published 5 Nov 2026Nikolai Schöbel and Jeremias Burger9 min read
Nikolai SchöbelJeremias Burger

Nikolai Schöbel and Jeremias Burger

Co-founders of Scalableloops GmbH. Nikolai Schöbel leads online marketing and AI strategy, Jeremias Burger the AI architecture. Both build AI systems and train teams on them in their own agency work.

On this page
  1. What is an AI agent workshop, and how does it differ from a training course?
  2. How does an AI agent workshop run over one day?
  3. Which tasks suit the first agent in an AI agent workshop?
  4. How do you know the agent from your AI agent workshop is good enough?
  5. Which data belongs in an AI agent workshop, and what helps against prompt injection?
  6. What should an AI agent workshop clarify about the AI Act and Article 50?
  7. What does an AI agent workshop cost?
  8. Where does the prototype go after the AI agent workshop?
  9. Frequently asked questions
  10. How to prepare your team for an AI agent workshop
  11. Where the information on this page comes from
Basics

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.

Agenda

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.

BlockWhat happensResult
OpeningThe team's workflows, expectations, approved AI servicesOne workflow chosen for the day
FoundationsAssistant, workflow and agent, each shown on running examplesShared vocabulary
Breaking it downSteps, data sources, decisions, risks per stepWorkflow sketch and first test cases
BuildingMixed groups implement the workflow on test dataWorking prototype
TestingRun test cases, hunt for errors, simulate an attackHit rate and error list
SafeguardsApprovals, permissions, logging, stop switch, legal positionRules for the workflow
ClosingContinue, change or stop, with names and datesDecision and next steps
Choosing

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.

WorkflowFit for the dayWhere the person stays
Categorising incoming emailsVery good, a fixed workflow is often enoughChecks samples and all uncertain cases
Drafting replies from past emails and documentsVery good, also shows the limits of document searchReads and sends personally
Updating contact details in a spreadsheetGood, if the agent only proposes changesApproves changes
Preparing an event, gathering documents and deadlinesGood, many steps and little riskDecides on dates and commitments
Preparing a travel requestGood up to the finished draftBooks, or approves each booking
Screening applicants or allocating tasks by behaviourNot for the workshop, may count as high-riskStays entirely with people
Acceptance

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.

Security

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.

Law

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.

Cost

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.

FormatPublished priceWhat it is good for
In-house workshop, one dayListed at around 1,900 to over 6,000 €, mostly net, often plus travelPrototype and decision within your team
Open course or webinarRoughly 500 to 1,100 € per person, some gross, some netGetting started as an individual
Including set-up of own infrastructureConsiderably more, one provider lists 9,000 € plus VATTechnical teams intending to operate it
Afterwards

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

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.

What now

How to prepare your team for an AI agent workshop

  1. 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.

  2. 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.

  3. 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.

  4. 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.

or call: +49 151 1576 5566

Further reading

Projekt-Detail

    Got a project in mind?

    We reply personally. First a use-case check, then an architecture proposal.

    Start your inquiry