The AI Act: what it actually changes for a company deploying an agent
The AI Act has stopped more projects in Polish companies than any technical barrier — usually not because something is prohibited, but because nobody established which category the planned system falls into. The difference matters: between the high-risk and the limited-risk regime sit a conformity assessment, registration in the EU database and technical documentation. Below we set out where a typical agent handling correspondence sits in that structure and what actually has to be done before launch. This is not legal advice — it is the list we go through with clients during the audit and leave with them in writing.
- An agent that reads and classifies email is a limited-risk system — no conformity assessment, no registration in the EU database.
- The high-risk obligations from Annex III have been pushed to 2 December 2027; that is the single most common reason projects get put on hold.
- Four real obligations remain: disclosing that it is AI, AI literacy in the team, GDPR documentation and deciding whether a DPIA is needed.
Where an email-handling agent sits in the AI Act
A system that reads enquiries, classifies them, pulls data from the ERP and drafts replies is a limited-risk system. Which means: no conformity assessment, no registration in the EU database, no Annex IV technical documentation. What remains are transparency obligations and general duties of care — a considerably lighter load than most legal departments expect.
The classification depends on the intended use, though, not on the technology. The same model pointed at recruitment, creditworthiness assessment or any other use listed in Annex III falls into the high-risk regime with all its consequences. That is why the system's intended purpose goes into the contract, and why changing it should require a fresh analysis — one sentence in the paperwork saves a quarter's work later.
The misunderstanding that stalls the most projects
The most common scenario looks like this: the legal department reads about the obligations for high-risk systems, sees the date August 2026 and puts the project on hold pending a full analysis. The problem is that those obligations apply to a category the planned agent usually doesn't belong to — and the deadline for Annex III systems has itself been moved by Regulation 2026/1744 from August 2026 to 2 December 2027.
The result is companies postponing implementations because of rules that don't cover them — and if the rules did cover them, they would have over a year to adapt. This is not an argument for ignoring regulation. It is an argument for establishing the category first and estimating the effort second: those two hours of work change a project schedule more than any technical decision.
The four obligations that remain
Limited risk does not mean "nothing to do." Four things have to be done, and all four can be closed before the pilot starts.
Disclose that it is AI. Art. 50 requires the recipient to know they are dealing with an AI system. In practice that is one element of the reply text — zero cost, matter closed. The obligation applies from 2 August 2026.
Take care of AI literacy in the team. Art. 4 imposes a duty to ensure AI literacy among the people operating the system; it applies from 2 February 2025. A short training session for the team that will work with the agent is enough, documented with an attendance list.
Complete the GDPR documentation. On 6 August 2026 UODO published sets of preliminary questions, including a version for companies using off-the-shelf AI tools. It is completed separately for each process and kept as evidence of accountability.
Check whether a DPIA is needed. Processing customer correspondence usually calls for a data protection impact assessment under Art. 35 GDPR. That determination should be made at the audit stage — after launch it costs considerably more.
The real scope of obligations for an agent handling correspondence, provided its intended purpose has been correctly classified as limited risk. For Annex III use cases the picture is entirely different — which is why classification is the first step, not the last.
The Polish context: the national act and the supervisory authority
On top of the EU regulation there is the national act on AI systems (Journal of Laws 2026 item 1003), in force since 11 August 2026, which designates the supervisory authority — KRiBSI. Inspections, proceedings and penalties start on 28 October 2026. For a company that carries one practical implication: the documentation described above should exist in a form you can produce, not one you have to reconstruct from the team's memory.
Our role here is limited and we describe it as such: we prepare the technical and process documentation on the system side and we go through the checklists with you. The legal classification of the use case inside your company belongs to your legal department or adviser — we give them the full picture of what the agent does, which data it reaches for and what it records.
Frequently asked questions
Do we have to register the agent in some database?
No, provided the system is a limited-risk system — registration in the EU database applies to high-risk systems. The classification is established before implementation and recorded together with the system's intended purpose, so that it can be demonstrated later.
Does the customer have to know an AI is answering them?
Yes — Art. 50 requires the recipient to know they are dealing with an AI system. In practice we handle that with one sentence in the reply. We have not seen it affect reception negatively; leaving it out and having the customer work it out for themselves costs far more.
What if we later want to use the agent for recruitment?
That is a change of intended purpose which moves the system into a different risk category — an Annex III one. It requires a separate analysis and a different set of obligations, so we treat it as a new project rather than an extension of the existing one. Which is why the intended purpose goes into the contract.