Most companies we talk to want an AI assistant. Very few want one that can see everything, decide on its own and send their employees’ data to a model somewhere outside their control. That tension shaped Guido, the assistant we built into our own platforms: ComplyCore, elearning.edu.pl and EasyACC.
Guido helps managers and HR teams with the work that fills their days: checking who needs which training, assigning courses, preparing groups, updating staff records and summarising progress. People ask in plain language. Guido answers, or prepares the change. A person decides.
Start with the job, not the model
We did not start by asking what a large language model can do. We started by listing what a training manager does every week and which of those tasks are repetitive, rule-based and easy to check. Those became Guido’s tasks. Anything outside that list is outside Guido’s job.
This matters more than it sounds. A general-purpose chatbot invites general-purpose questions, and general-purpose answers are hard to verify. An assistant with a defined job can be tested against that job, and its users quickly learn what to expect from it.
The user’s permissions are Guido’s permissions
Guido never has more access than the person using it. When a shift leader asks a question, Guido sees what that shift leader is allowed to see. When an HR administrator asks for a change, Guido can prepare only the changes that administrator is allowed to make. If a request goes beyond the user’s rights, Guido says so instead of finding a way around it.
It sounds obvious, but it is the most common shortcut in AI features: one service account with broad access, and a model trusted to behave. We chose the opposite. The boundaries live in the platform, not in the prompt.
Nothing changes without “Approve”
Guido can answer questions directly, but it never saves a change by itself. Every action — an assignment, a new group, an update to a staff record — is shown as a card that summarises what will happen. The change is saved only when the user presses “Approve”.
The card does two jobs. It keeps a human in control of every decision, which is what the EU AI Act expects of systems like this. And it makes mistakes cheap: if Guido misunderstood the request, the user sees it before anything happens and simply edits or rejects the proposal.
Every call leaves a trace
Each request to Guido is recorded in an audit log: who asked, what Guido proposed and what was approved. When an auditor or a customer asks why something changed, the answer is already in the system. The cost of each answer is shown next to it, so AI usage never becomes a surprise on the invoice.
No personal data in the prompt
Training records are full of personal data: names, employee numbers, dates, sometimes health-related permits. Before a request reaches the AI model, we strip personal data from it. The model works on the structure of the task, not on people’s identities.
For our customers this answers the first question their data protection officer asks. Personal data stays in the platform, hosted in the EU, and the model never needs to see it to be useful.
Many models, one assistant
Guido runs on Amazon Bedrock. That gives us access to several AI models and lets us choose one per task: a fast, inexpensive model for simple lookups, a stronger one where reasoning matters. Models run in an EU region.
It also protects the product from the pace of the AI market. When a better model appears, we can evaluate it and switch without rebuilding the assistant or retraining users. The boundaries, the approval flow and the audit log stay exactly the same.
Learning the routine
Many tasks repeat: the same induction for every new starter, the same reminder at the end of each month. Guido learns which actions a user performs most often and offers to repeat them, so a routine that took a dozen clicks takes one sentence. Repeated actions still go through approval.
Support, not replacement
Guido is not designed to replace anyone. It takes the repetitive part of the work off people’s plates, so they have more time for the parts that need judgement — talking to employees, planning, handling exceptions. The aim is more productive and more comfortable work, not fewer people.
What this means for your business
If you are considering an AI assistant for your own platform or internal tools, the questions that matter are the ones we asked: which tasks should it do, with whose permissions, who approves its actions, which data may leave your systems, and how will you know what happened. Get those right and the choice of model becomes a detail you can change later.
We build assistants like Guido for client systems on the same principles. If you would like to see Guido in action or discuss an assistant for your own processes, contact us.