Giacomo Mantovani

Article · November 19, 2026 · By Giacomo Mantovani

The AI Act if you already run AI in production: what to do now

High-risk deadlines moved to 2027, chatbot transparency did not. What a team with AI already in production should prepare, from inventory to audit trails.

  • AI Act
  • compliance
  • AI in production
  • regulation
Part of it applies now. The rest has a date. Inventory, Classification, Transparency, Logging, Human oversight.

Every AI system I have taken to production sits in a different place under the AI Act. One is an LLM agent that routes more than a million internal help desk tickets a year at a major Italian bank. Another checks inheritance documents against the bank's records, and a third answers the phone for a clinic with 40+ doctors. The regulation asks different things of each, and in July 2026 the EU changed several of its dates. Below is the checklist I would give an engineering team that already has AI running and needs to know what applies now and what can wait. I am an engineer, so treat it as the technical half of the work; your legal team has the last word on classification.

What changed in the AI Act timeline in 2026?

The AI Act, Regulation (EU) 2024/1689, entered into force in August 2024 with obligations phased in over several years (Official Journal, July 2024). In 2026 the EU amended it with the Digital Omnibus on AI, Regulation (EU) 2026/1744, published in the Official Journal on 24 July 2026 (European Parliament Legislative Train, September 2026). Law-firm summaries put its entry into force on 27 July 2026 (Cooley, August 2026).

The main change is a delay for high-risk systems. Stand-alone high-risk systems listed in Annex III, such as those used for hiring or credit scoring, now have until 2 December 2027 instead of 2 August 2026. High-risk AI built into regulated products under Annex I moves from August 2027 to 2 August 2028 (European Parliament Legislative Train, September 2026). According to Cooley, the transparency obligations in Article 50 kept their date of 2 August 2026 (Cooley, August 2026).

AI Act dates for AI already in production (status in October 2026)
DateObligationWho it hitsWhat to prepare
2 Feb 2025Prohibited practices (Art. 5) and AI literacy (Art. 4)Every provider and deployerConfirm no use case is banned; record staff training
2 Aug 2025General-purpose AI model rules (Art. 51 to 56); penalties regimeModel providers such as OpenAI, Anthropic, Google, MistralCollect your model vendors' documentation; check contracts
2 Aug 2026Transparency (Art. 50): people are told they talk to an AI; AI content disclosedProviders and deployers of chatbots, voice agents, generative toolsDisclosure in the first message or greeting; labels on generated content
2 Dec 2026Machine-readable marking (Art. 50(2)) for generative systems on the market before 2 Aug 2026Providers of generative systemsWatermarks or metadata on generated output
2 Dec 2027High-risk obligations for Annex III systemsHiring, credit, insurance, education and other listed areasRisk management, logging, human oversight, documentation, registration
2 Aug 2028High-risk obligations for Annex I systemsMakers of regulated products such as medical devices and toysConformity assessment inside the product's existing regime

The table is my reading of the amended text and of law-firm summaries, so check a date against the consolidated text on EUR-Lex before you plan around it. Some details, such as what an Article 50 label must look like, will depend on codes of practice that the amended Article 50(7) encourages (AI Act Explorer, July 2026).

Which AI Act obligations already apply to a company that uses AI?

The bans on prohibited practices, such as social scoring or emotion recognition in the workplace, have applied since 2 February 2025. The AI literacy duty in Article 4 started on the same day: staff who operate or rely on AI systems need enough understanding of them for their role. The Omnibus softened the wording, and the amended article says it does not require a specific level of AI literacy for any individual (AI Act Explorer, July 2026). The duty to take measures remains, so a short training with a record of who attended is still worth doing.

The rules for general-purpose AI models have applied since 2 August 2025, and they fall on the companies that train and release those models. If you call a model through an API and build a product on it, you are normally the provider of your own AI system and a user of theirs. For most companies that makes Article 50, in force since 2 August 2026, the obligation most likely to touch a system that is already live.

Is my AI system high-risk under the AI Act?

For back-office and customer service tools the answer is usually no, and the reasoning is worth writing down. Annex III lists the areas that make a system high-risk. For a typical company the relevant ones are hiring and worker management (CV screening, promotions, terminations), creditworthiness assessment and credit scoring of individuals, risk assessment and pricing in life and health insurance, and access to essential services. Article 6(3) adds an exception: a system in one of those areas is not high-risk if it only performs a narrow procedural task or a preparatory task for an assessment that a person makes, unless it profiles people.

How the systems I have built would likely be classified
SystemAnnex III area?My reading
Internal help desk routing, 1M+ tickets a year, 400+ categoriesNoNot high-risk: it routes employee requests and decides nothing about people
Inheritance document checks at a bankNoNot high-risk: a failed check goes to an operator, who decides
Financial statement extraction, 2,500+ documents a dayDepends on downstream useExtraction is not listed; check what consumes the data if it feeds credit decisions about individuals
Clinic voice agent for bookingsNoNot high-risk, but Article 50 applies because callers talk to an AI

Write the same table for your own systems and date it. If an auditor asks, a document that explains why a system is outside Annex III is much easier to defend than a verbal answer. If you rely on the Article 6(3) exception for a system inside an Annex III area, the Omnibus kept the duty to register it in the EU database, with a simplified procedure (Cooley, August 2026).

What does Article 50 mean for chatbots and voice agents?

From 2 August 2026, a provider must design an AI system that interacts with people so that they are told they are dealing with an AI, unless it is obvious from the context. Deployers who publish deepfakes, or AI-generated text meant to inform the public on matters of public interest, must disclose that too (Official Journal, July 2024).

For a text chatbot this usually means a line in the first message and a label in the interface. A voice agent has to say it in the spoken greeting, because a caller sees no label. Response speed makes this more pressing. IGEA, the voice agent I co-founded for a clinic with 40+ doctors, replies in under 1 second, and at that speed the conversation gives the caller no hint that software is on the line. Any voice agent in that position needs the disclosure in its greeting, which costs a few seconds per call.

What should an engineering team prepare for the AI Act now?

The delay to December 2027 buys time, and most of the work it buys time for is engineering: logs, review screens, documentation. This is what I would put on a backlog today.

AI Act preparation checklist for AI already in production
ItemWhat it looks like in practiceDone when
1. InventoryEvery AI system and model call in production, with owner, model, vendor and data processedOne list with an owner who keeps it current
2. ClassificationPer system: Annex III area or not, Article 6(3) reasoning, Article 50 or notA dated document per system
3. TransparencyDisclosure in chat, in voice greetings and on generated contentVisible in the product and covered by tests
4. Logging and audit trailEach decision traceable to input, model version and output, with no sensitive data in logsYou can reconstruct any decision from last month
5. Human oversightExceptions routed to a person with the reason; the person can overrideOverrides are logged and reviewed
6. Data location and vendorsWhere models run, which contracts cover them, what the vendor documentsContracts and vendor documentation on file
7. AI literacyShort training for people who operate or rely on each systemMaterial and attendance recorded

What do logging and human oversight look like in a real system?

In the banking projects I worked on, the design choices that got a system past compliance are the same ones the AI Act now describes. In the inheritance document checks at a major Italian bank, standard files clear automatically, and a file that fails a check goes to an operator with the specific reason. The operator reviews one discrepancy instead of the whole file, and handling time fell by 80%. Human oversight with a clear trigger is also where the saving came from.

The financial statement pipeline processes more than 2,500 documents a day and runs about 30 validation checks on each. Only documents that pass every check clear without a person. Models ran on Azure OpenAI inside the client's EU cloud, every production document had an audit trail, and sensitive data never went into logs. Keeping sensitive data out of logs settles a conflict between traceability and GDPR: you need to reconstruct decisions, and you must not keep personal data you do not need. Log identifiers, model versions, check results and reasons, and leave the documents where they already live.

What are the most common AI Act mistakes when AI is already live?

The first is reading the Omnibus as a general postponement. Only the high-risk deadlines moved; transparency, AI literacy and the bans apply now. The second is deciding a system is not high-risk without writing down why. The third is logging everything, personal data included, in the name of traceability, which solves an AI Act problem by creating a GDPR one. The fourth is oversight that exists on paper only: a reviewer who approves every case by default, or a review screen so slow that people stop opening it. The last is handing the whole topic to the legal team. Inventory, logging, disclosure and review screens are code, and they need sprint time.

Next step

Need to map your AI systems against the AI Act?

I can review what you run in production, classify it together with your legal team and plan the logging and oversight work.

Let's talk

For banks and financial services, see how I approach AI consulting for banks and finance. For document pipelines with audit trails and EU hosting, see the document AI service for regulated industries.

Frequently asked questions

Does the AI Act apply if we only use ChatGPT or Claude through an API?

Yes, in part. The model provider carries the general-purpose AI obligations. If you build a chatbot or an assistant on the API and offer it to customers or staff, you are the provider of that system, so Article 50 transparency applies to it and the AI literacy duty applies to the staff who use it.

What are the fines under the AI Act?

Article 99 sets fines of up to €35 million or 7% of worldwide annual turnover for prohibited practices, and up to €15 million or 3% for most other obligations, transparency included. For SMEs the lower of the two amounts applies, and the Digital Omnibus extended that rule to small mid-cap companies.

Do we have to register our AI systems in an EU database?

Only high-risk systems under Annex III, plus systems in Annex III areas that you have classified as not high-risk under Article 6(3). The Omnibus kept that second duty with a simplified procedure. Back-office and customer service tools outside Annex III do not need registering.

Does the AI Act delay change anything under GDPR?

No. GDPR applies to any personal data your AI systems process, whatever the AI Act dates. Systems that process personal data at scale or in sensitive areas such as health may already need a data protection impact assessment.

Author

Giacomo Mantovani

AI Engineer & Consultant for Companies and Agencies

LinkedInGitHub

Next step

Tell me the workflow. We scope it on a free call, then a working version is live within 10 days.

Book a scoping call

Related articles