AI agents and workflows

AI that relieves your team instead of replacing it.

We build AI agents and workflows for work that is still done by hand: answering emails, processing documents, preparing reports. We agree with you in advance what the AI may do on its own. We start small, measure what it delivers and only expand once it works.

Discuss your workflow How a pilot works

A woman pins the last card on a board; a teal ribbon links an envelope, a document, a calendar and a tick.

At a glance

Start: A pilot around one clearly defined task
Control: Agreed in advance what AI may do itself and what goes through your team first
Data: In the EU as standard, never used to train models

Demo

Demo with fictional data

What a mailbox workflow looks like

Three new emails in the shared mailbox of a made-up non-profit. On the left what comes in, on the right who does what.

Shared mailbox · info@nonprofit.example

3 new emails

Comes in

Sofie
Routine question

Certificate for the health insurance fund
“Could I get a certificate for my son’s summer camp?”

AI does this itself · reversible

Label: Certificate
→ to the Admin team
Registration looked up and ready

A colleague decides · goes outside

Draft reply with certificate is ready. A colleague reads it over.
Comes in

Jonas
Info missing

Any places left?
“Can my daughter still join next week?”

AI does this itself · reversible

No label
→ to the "To review" folder
Reason: activity not mentioned. AI doesn’t guess.

A colleague decides · goes outside

A colleague picks the label and asks Jonas which activity he means.
Comes in

Mira
Wrong label

Paid twice?
“I think I paid the registration fee twice.”

AI does this itself · reversible

~~Registration~~ → Payment
↺ Reverted: a colleague moved the email to Accounts.
That correction is logged.

A colleague decides · goes outside

Accounts checks the payment and adjusts the draft reply.

Reversible and internal: AI does it itself. Going outside: a colleague first.

We agree on that in the AI fact sheet

Agent or workflow?

For each task we choose the simplest solution that works

An AI workflow always follows the same steps, with the AI doing the reading and writing. An AI agent chooses its own next step, within the agreements we make in advance. Anything that goes outside or decides about a person, your team always approves first.

	Fixed rules	AI workflow	AI agent
What it does	Always the same steps, no AI	Fixed steps. The AI reads, sorts and prepares.	Finds what it needs and picks the next step
Example	A form that always lands in the same folder	Reading invoices and preparing them in your accounting	Answering a question by searching policies and procedures itself
Suited when	What comes in always looks the same	What comes in varies, but the steps don’t	Every question needs a different approach
Who decides	The rules you set in advance	Your team approves what goes outside	The agent only does what can be undone. Your team approves the rest.

Five tasks to start with

AI helps most where someone has to read and interpret today: free text, emails or PDFs. If fixed rules are enough, we build it without AI.

Documents and invoices

Today
Someone copies data from a PDF into the accounts.

With AI
AI reads out the data, your team approves with one click.
Shared mailbox

Today
An hour of sorting and forwarding every morning.

With AI
Emails are labelled and shared out. A reply only goes out once someone has checked it.
Forms and requests

Today
Free text that someone has to interpret by hand.

With AI
AI turns requests into structured data for your systems.
Reporting

Today
Piecing together figures and notes for every meeting.

With AI
A first version is ready; you add to it and decide.
Knowledge assistant

Today
Searching for answers in dozens of documents and files.

With AI
An AI agent that finds the right documents itself and answers with sources, inside your own environment.

Chat tool or custom build?

A chat window is not a process

Separate AI tools are handy for exploring. For recurring work you want something predictable, checkable and managed.

A woman turns away from a board of loose chat bubbles towards a board with a tidy series of steps.

	Separate chat tool	Agent or workflow with Ethic
Where the work happens	In a separate window, copying and pasting	In your own systems, where the work already is
What data AI sees	Whatever someone happens to paste into the chat	Only the data that is needed, agreed in advance
Who catches mistakes	The user, with no fixed agreement	A fixed review step, with thresholds and logging
What stays in your hands	Prompts and results in someone’s chat history	A managed workflow with an owner in your team

Our ground rules

Built responsibly, from day one

Two people lay the first stone of a foundation; she checks it with a spirit level.

Human review

AI may act on its own when the work stays internal and can be undone. A colleague reviews external communication and decisions about people first.

Every step recorded

For each step the logs show the input, the sources used, the outcome, the checks and who approved.

Data protected

European hosting as standard. Never training on your data.

The right model

As small as possible, as large as needed. That saves energy.

How a pilot works

One task, five steps, an honest decision

Only once the pilot proves it works do we expand to other tasks.

1

Choose a task
Together we pick one task that comes up often and is clearly defined.
2

Data and risks
We look at examples, privacy and what could go wrong.
3

Prototype
A first version in your own environment, with a review step for your team.
4

Measure
How much time does it save? How often does someone have to correct it?
5

Decide
Roll out, adjust or stop. Based on figures, not gut feeling.

What a pilot covers

Small enough to keep track of

One task

One task that comes up often, with a clear start and finish. Other tasks only come after that.

An agreed benchmark

We agree up front what we measure, such as time per task or how often someone corrects, and which mistakes are acceptable.

A review moment

At a moment chosen in advance, we set the results against the benchmark. Everyone who works with the workflow gets an explanation beforehand.

No surprises afterwards

After a first conversation you get a proposal with a clear scope and a fixed price. Anything outside it, we plan separately.

AI fact sheet

Every agent and workflow comes with an AI fact sheet

What does the AI do, what data does it see and who keeps an eye on it? For every agent and workflow we build, that goes on one clear sheet that your team and your DPO understand straight away. It is also where we agree in advance what AI may do itself and what is checked by a colleague first.

A man reads a sheet with a short, clear list.

Example

HR policy knowledge assistant

Version 1.2 · March 2026

What it does

Answers staff questions about leave, working from home and expenses, always with a link to the source document.
Which data

Only reads the HR policy folder on SharePoint. Personnel files and pay data stay out of reach.
Model and location

A small language model, hosted in the EU. Your data is not used to train models.
Who checks

If the sources don’t back up an answer well enough, the assistant gives no answer and refers to HR. HR checks a sample every month.
What we log

Every question and every answer, kept for 90 days and visible only to HR.
Owner and stopping

Owner: the HR coordinator. You can stop at any time: the integration is withdrawn and the logs are handed over.

After the pilot

What we build should still work in three years

A successful test is only the beginning. Suppliers change their software, models disappear and your team changes. That is why we agree in advance what happens after launch.

A woman plants a cutting in a planter next to other young plants.

You’ll know when something goes wrong

Every agent and workflow has checks and error alerts. If Microsoft or Odoo changes something, we usually spot it before your team does.

Switch models without rebuilding

If a smaller, more efficient or European model comes along, you can switch without rebuilding the workflow. We do test it again afterwards.

Everything is in your name

Code, prompts, documentation and access are in your name. If you carry on with your own team or another partner, we hand everything over properly.

Maintenance in a fixed agreement

Updates, monitoring and small changes are covered by a monthly agreement. New workflows are planned separately.

Zenjoy has been maintaining platforms and integrations for years after launch. They grow step by step with new systems, without starting over each time.

See Zenjoy’s work

Direct contact with the builders

No account manager in between: the people who design your project also build it.

Everything in your name

Code, documentation and access are in your name. If you carry on with your own team or another partner, we hand everything over.

The same team stays

After launch, the same people provide maintenance and support. If a supplier changes something, they make the adjustments.

See Zenjoy’s work

Frequently asked questions

What people often ask us

Question not listed? Just ask us.

What is the difference between an AI agent and an AI workflow?

An AI workflow always follows the same steps: the AI reads, sorts and prepares. An AI agent picks its own next step, for example which document to open to answer a question. For each task we choose the simplest solution that works. Often that’s a workflow: it’s more predictable and cheaper.

Can an AI agent make decisions on its own?

Only about things that can be undone and stay inside your organisation, such as labelling an email or preparing a draft. Anything that goes outside or decides about a person is approved by a colleague first. We set this out for each agent in the AI fact sheet.

Does AI train on our data?

No, not without an explicit agreement. We choose the model and hosting based on how sensitive your data is, and record where everything ends up.

Can AI make mistakes?

Yes, no model is always right. That is why we build in a review step. Anything that goes outside or decides about a person is approved by a colleague first. The logs show what happened. We agree in advance how many mistakes are acceptable and what happens then.

Why not just use Microsoft Copilot?

For personal help with emails, meetings and documents, Copilot is often enough, and then we say so. An agent or workflow is something else: a fixed task across several systems, with a review step and logging. There you choose the model and where it runs.

And tools like n8n or Zapier?

They are handy for getting started quickly, and sometimes we recommend them too. For a workflow that has to run for years, every extra step in between is another subscription and another place where something can break. For each workflow we choose what is easiest for you to maintain.

Do we need to know which model we want?

No, we choose that together. We take a model that is as small as possible and as large as needed, and that runs in the EU. Which model it is and where it runs goes in the AI fact sheet.

How long before we see something?

A pilot is built around one workflow, so you see results quickly. The proposal after the first conversation includes a concrete schedule, with a review point where you decide: roll out, adjust or stop.

Which task costs your team hours every week?

Tell us. In a first conversation we work out together what makes sense: an AI agent, a workflow or something simpler.

Discuss your workflow Start with a 1-hour Pitstop

No obligation · 30 minutes · online, with someone who builds, tests and fine-tunes it