AI agents
AI that does more than answer — it acts in your systems.
An AI agent gets work done: it pulls data, runs steps in your systems (ERP, CRM, helpdesk) and closes the loop — within set boundaries, with human approval where the stakes are high. We build it on AWS — your data and permissions stay with you.
RAG vs agent
RAG answers a question. An agent takes the next step.
RAG gives an answer from a source — and that is where its work ends. In a company the answer is only the beginning: someone has to issue a document, update a CRM ticket, check a status in the ERP, send an email. Today a person does it by hand, copying data between systems.
An AI agent closes that gap: it plans the steps, uses the tools and integrations you grant it, and carries the task end to end. But the autonomy works only within boundaries — with per-role permissions, logs, and human approval before the agent does anything irreversible.
How we build agents
Autonomy within boundaries, not a "black box".
An agent that runs in production needs the same safeguards as a production RAG — except here the stake is action, not just an answer.
1. Bounded workflows
The agent gets a concrete, defined scope of tasks — not "do whatever". It is known up front what it can do and what it will never touch.
2. Tools and integrations
The agent acts through controlled tools and connections to your systems (ERP, CRM, DMS, helpdesk). Every action passes through a layer you control.
3. Permissions and human approval
Access is per role — the agent uses only what that role is allowed to. For high-stakes actions it stops and waits for human approval (human-in-the-loop).
4. Logs, evaluations and fallback
We log every agent action, measure quality and cost, and when the agent is unsure it hands the case to a human instead of guessing. We keep quality up over time.
Architecture
What a production AI agent on AWS looks like
Every layer runs in your AWS account (EU region) — from planning to actions in your systems.
- Goal / task
- Planning (LLM, Amazon Bedrock)
- Tool layer
- Integrations (ERP/CRM/DMS)
- Human approval
- Execution
- Logs & evals
At every layer: per-role permissions, audit logs and human-approval checkpoints — what AI that takes action in a regulated company requires.
For regulated companies
An agent that acts — under control and with an audit trail.
When AI not only answers but takes actions, control matters more, not less. The same mechanisms that keep quality up also provide oversight and accountability.
Action boundaries
The agent acts only within a defined scope, through granted tools. Beyond the boundaries it has no access.
Per-role permissions
The agent uses only the data and actions that a given role is allowed to — exactly like an employee.
Audit logs
We log every action and its basis — leaving a trail of what the agent did, when and why.
Human-in-the-loop
Irreversible or high-stakes actions require human approval before the agent carries them out.
Compliance: AI Act and GDPR
Logs, evaluations and oversight points support AI Act and GDPR obligations; the risk classification is closed by an AI Act audit.
Need answers grounded in your sources first? See RAG / knowledge bases, or our full services.
Use cases
Where an agent closes the loop
Example processes where an agent acts within set boundaries — with human oversight where it is needed.
Ticket handling
The agent triages a ticket, pulls data from the CRM and the knowledge base, proposes or performs the next step — and routes unusual cases to a human.
Internal helpdesk
An employee asks, the agent finds the answer in the procedures and — with approval — performs a routine action: opens a ticket, updates a status, drafts a reply.
Back-office processes
The agent joins steps scattered across systems (ERP, DMS, email): gathers data, prepares a document, waits for approval and closes the case.
Engineering proof
Production AI discipline — proven on our own product.
100%
AI extraction accuracy on packaging data (validation set, n=200)
$0.0006
cost per AI scan in production
302,516
drug-interaction records in the production knowledge base
Our strongest public proof, mojApteczka, is a production GenAI/RAG system on AWS — not an autonomous agent. The same foundations that keep it in production — AWS architecture, automated quality evaluations, inference cost control, logs and governance — are what we carry into the AI agents we build for clients.
See the full case studyQuestions
Before you ask
A chatbot and RAG answer questions. An agent performs tasks — it plans steps and acts in your systems through granted tools. RAG is often one of the agent's tools: first find the right information from a source, then act on it.
No. We build agents that act within set boundaries, with per-role permissions and logs. For irreversible or high-stakes actions the agent stops and waits for human approval. Autonomy without boundaries is a risk, not a feature.
The agent runs in your AWS account (EU region) and only through the tools and permissions you grant it. Data does not reach public models or their training, and every action is logged.
We build AI agents for clients, while our public production proof is mojApteczka — a GenAI/RAG system on AWS. It shows the engineering we carry into agents: quality evaluations, cost control, logs and governance. The scope of an agent for your company is set in an audit.
First an audit (2–4 weeks) and an optional proof of concept on a real slice of your process. The production timeline depends on the number of integrations and the stakes of the actions — we set it after the audit, on facts, not on a promise of "in X weeks".
Have a process AI could close — not just describe?
Let's start with an audit: we'll check which steps an agent can safely take over, where human oversight is needed, and how fast we can ship a working system.