AI Act · retrofit
Your chatbot or voicebot is already running, and since 2 August 2026 it has to tell people they are talking to a machine. In five working days I change the widget, the logging and the route to a human, and hand you the AI system register plus the paperwork an inspector asks for.
5 days
kickoff to live
2 Dec 2026
AI content marking deadline
12
checkpoints
The report covers the technical state and the paperwork; the reading of the law is signed off by your lawyer.
Law firms wrote the requirement down. Nobody touched the widget or the code afterwards.
Article 50 of the AI Act applies from 2 August 2026 and requires that a person is told at the first interaction they are dealing with a machine, clearly and in an accessible way. A line in your terms of service or in the footer does not cover it. The notice belongs where the conversation starts.
Law firms describe the duty and state outright that technical implementation is out of scope. So nobody rewrites the greeting, nobody configures consent logging, nobody documents the escalation path. The document goes into a folder and the system stays exactly as it was.
In Poland the KRiBSI commission gets inspection powers on 28 October 2026 and will ask which AI systems run in the company and who owns each one. Without a register and without logs the answer is that you do not know. The regulation's upper penalty bands are up to €35M or 7% of turnover for prohibited practices and up to €15M or 3% for other breaches.
Article 4 on AI literacy has applied since 2 February 2025 to every company deploying AI, including a firm that only uses ChatGPT in customer service. A one-off message in your work chat is not a programme. You need role-adapted material and proof that the session happened.
Each one is visible either in the product itself or in a document you can hand over.
The AI notice goes into the opening of the conversation: in the chat widget above the input field, in the voicebot as the first sentence before any question. It stays visible after a page reload and when someone returns to the thread. You get the wording in four languages.
I build an explicit way out of the bot: a button in the widget, trigger phrases that force escalation, and a handover that carries the transcript into your inbox or agent chat. Outside working hours the bot collects a contact and states a real response time.
I set what gets stored: the consent text, the timestamp, the version of the disclosure shown, the session id. Retention period and storage location are written down, and the export for an inspection is set up in advance rather than assembled by hand.
I list every AI system actually running: vendor, model, purpose, input data, the owner on your side, the legal basis. You get the register as a sheet your own team can keep current without me.
Product descriptions, blog posts and generated images get a visible label and metadata in the file. This is the specific duty that carries a 2 December 2026 deadline for systems placed on the market before 2 August 2026.
Briefing material adapted per role: support, marketing, sales. A 60-90 minute online session, an attendance list, and a one-page internal policy on using AI. This is the piece of evidence that is missing most often.
I need access to the bot panel and the site. Everything after that is on me.
I list what is genuinely running: the widget on the site, the voicebot on the phone line, the description generator, the plugins in the shop. I collect access and pin down who owns each system on your side.
I work through the checklist: disclosure, placement, escalation, consent, retention, content marking, register, staff briefing. You get a report with a verdict on every checkpoint and a list of the gaps.
I change the widget, the system prompt, the logging configuration and the escalation branch. Everything lands on staging first and moves to production once you sign it off.
AI system register, retention policy, briefing material and a one-page summary for your lawyer. The files stay with you in an editable format.
I run the session for your team and leave the attendance list behind. Before 2 December 2026 I do one free re-check to see whether anything drifted after your own edits to the bot.
Packages
One chatbot or voicebot, one language, one channel. Audit and implementation fit into the same week.
A Shopify shop running a Tidio widget, a clinic with a voicebot on the phone line.
The full set for a company with one bot and a couple of other AI tools inside. It closes the product and the paperwork in one pass.
A company with a site chat, a description generator and ChatGPT in support.
Up to four bots, brands or language versions, with a separate register per legal entity.
A salon chain running a bot on the website, WhatsApp and Messenger.
Prices are net and payable once you accept the report. This is engineering and documentation, not legal advice: your lawyer signs off the reading of the law, I own what actually changes in the system.
Four situations where this engagement gives you nothing.
If the chatbot is still only a plan, the right place to start is the page on Services
Describe the task in two sentences. You get a fixed price, a timeline and a plan. If it isn't a fit for us, we say so straight away and point you elsewhere.
Yes. Since 2 August 2026, Article 50 of the AI Act requires that a person interacting with an AI system is told so at the first interaction, clearly and in an accessible way. The duty sits with you as the deployer, not with the vendor of the widget. Placement matters as much as wording: the notice has to be where the conversation begins.
The disclosure duty already applies, from 2 August 2026. Systems placed on the market before that date have until 2 December 2026 to bring the marking of AI-generated content into line. In Poland there is a third date: 28 October 2026, when the KRiBSI commission gains the power to inspect, run proceedings and impose fines.
Up to €35M or 7% of worldwide turnover for prohibited practices, and up to €15M or 3% for other breaches, transparency duties included. The higher of the two figures applies. For a small company the practical risk is rarely the ceiling, it is the proceeding itself and the time it eats.
Komisja Rozwoju i Bezpieczeństwa Sztucznej Inteligencji, KRiBSI for short, created by the Polish act on artificial intelligence systems. The substantive part of that act entered into force on 11 August 2026, and the commission's powers to inspect, run proceedings and impose fines start on 28 October 2026. The chair is appointed in October and the commission is operational from November 2026.
No, the transparency duty falls on you as the deployer. In practice the Tidio, Intercom or Voiceflow panel gives enough control: I change the greeting, the widget copy, the escalation and the retention without touching vendor code. The whole job takes five working days and does not require rewriting the bot.
Yes. Article 4 of the AI Act on AI literacy has applied since 2 February 2025 to every organisation using AI, regardless of size or risk class. A single email or a generic handout is not enough: you need a role-adapted programme with material and evidence that the session took place. A breach falls into the band of up to €15M or 3% of turnover.
No. It is engineering and documentation: I change the product, configure the logs, compile the register and prepare the material for your team. The reading of the law and the final risk assessment are signed off by a lawyer, yours or one I can point you to. The one-page summary is written so a lawyer can review it in fifteen minutes.
What clients say
A few words from founders and teams we've shipped for.
Launched our store in 11 days — multilingual, fast, and already on page one for three of our main keywords.
The Telegram bot handles the first conversation with every lead now. My sales rep only steps in when the deal is hot.
Leads land in our CRM with every field already filled. No more copy-pasting from chats — the pipeline moves itself.
Shipped to the App Store and Google Play in three weeks. The handover docs let my team take it from there without me.
Set up our Google and Meta ads from zero. Cost per lead dropped by about a third in the first month.
He rebuilt the site and six weeks later ChatGPT was recommending us by name. Didn't know that was even possible.
Free 30-minute call. No sales pitch - just a real conversation about what you need and whether we're the right studio to build it.
No sales pitch — just a clear read on your scope.