Integration walkthrough: adding AI Act Radar to a B2B AI product in one afternoon
Illustrative scenario. The team and product below are composite: every step, file, and command works against the live AI Act Radar API. Information only, not legal advice.
You ship an AI product that your enterprise customers deploy in the EU. Their procurement teams have started forwarding you AI-Act questionnaires, their privacy counsel asks which articles your system touches, and your sales engineering Slack now has a dedicated #ai-act-questions channel. The information your customers want already exists across twelve official EU and national sources: you just don't want to hand-curate it. This walkthrough shows how a two-person team integrated AI Act Radar in a single afternoon and turned regulatory updates into a product surface their customers actually use.
The starting point
Picture a Series-A AI document-processing company headquartered in Munich, fifty employees, one in-house counsel, no compliance team. Their customers (banks, insurers, medical-device makers) sit squarely in Annex III high-risk territory. The team had already done the regulatory mapping once, in February 2026, in a Notion page. Six weeks later that page was already stale: the AI Office had revised the GPAI Code-of-Practice template, the BfDI had published a new guidance note on automated decision-making, and CEN-CENELEC had moved an Annex-III-relevant standard into formal vote. Their customers asked which of those affected them. The team didn't know.
Step 1: Subscribe to the relevant feeds (10 minutes)
The team signed up for the Team plan, generated an API key, and pointed the webhook at a Cloudflare Worker they already had for internal Slack notifications. They filtered to the topics that matter to their product:
POST https://api.aiactradar.com/v1/subscriptions
Authorization: Bearer aia_live_…
{
"url": "https://hooks.acme.dev/aia",
"topics": ["high-risk", "data-governance", "human-oversight", "transparency", "national-de"],
"min_severity": "medium"
} The first delivery arrived four minutes later: a backlog of the seventeen events from the past 30 days that matched the filter, signed and replayable.
Step 2: Wire the webhook into the existing Slack handler (20 minutes)
The handler already verified HMAC signatures from a different vendor; adding AI Act Radar's signature scheme was twenty lines:
// src/aia.ts
import { verify } from './hmac';
export async function handleAIA(req: Request, env: Env) {
const raw = await req.text();
const ts = req.headers.get('x-aia-timestamp') ?? '';
const sig = req.headers.get('x-aia-signature') ?? '';
if (!await verify(raw, ts, sig, env.AIA_SECRET, 300)) {
return new Response('invalid signature', { status: 401 });
}
const event = JSON.parse(raw);
// De-dupe on event id
const seen = await env.EVENTS.get(event.id);
if (seen) return new Response('ok', { status: 200 });
await env.EVENTS.put(event.id, '1', { expirationTtl: 7 * 86400 });
// Route by topic
if (event.topics.includes('national-de')) {
await postSlack(env, '#counsel', formatGerman(event));
} else if (event.severity === 'critical' || event.severity === 'high') {
await postSlack(env, '#ai-act-radar', format(event));
} else {
await postSlack(env, '#ai-act-radar-firehose', formatBrief(event));
}
return new Response('ok', { status: 200 });
} By 14:30 every Slack channel was getting the right firehose. Engineering saw critical events; counsel saw German-only events; the firehose channel collected everything else for weekly review.
Step 3: Add an MCP server for the customer-facing assistant (45 minutes)
The product already had a chat assistant: a thin wrapper around an internal LLM that answers customer support questions about deployed models. The team registered the AI Act Radar MCP server alongside their existing tools so the assistant could answer questions like "is there a recent BSI publication that affects how we document the human-oversight handoff?" using live data, not stale training context.
// mcp.config.json
{
"servers": {
"aiactradar": {
"command": "npx",
"args": ["@aiactradar/mcp-server"],
"env": { "AIA_API_KEY": "${env.AIA_API_KEY}" }
}
}
} The assistant immediately gained four tools: search_events, get_event, list_topics, and get_source_status. No prompt engineering, no scraping, no curated knowledge base to maintain.
Step 4: Surface relevant events inside the customer dashboard (90 minutes)
The big lift. The team added a new dashboard widget that shows the customer the three most relevant AI-Act events for their specific deployment. The mapping logic is simple: they already tagged each customer deployment with its Annex III category and primary processing member state, so they just queried the public API filtered to those two dimensions:
// app/api/customer-radar/route.ts
const events = await fetch(
'https://api.aiactradar.com/v1/events?' + new URLSearchParams({
topics: `high-risk,national-${customer.country.toLowerCase()}`,
annex_iii_category: customer.deploymentCategory,
min_severity: 'medium',
since: thirtyDaysAgo,
limit: '3'
}),
{ headers: { Authorization: `Bearer ${process.env.AIA_API_KEY}` } }
).then(r => r.json()); The widget shipped to staging at 17:00 and to production the next morning. It surfaces under "Compliance context" on every customer's main dashboard. Clicking through opens the official source: never a re-write, never legal interpretation.
What changed in the next two weeks
- Three customer questionnaires that previously took two days to research were closed within an hour, because counsel could point to specific recent events instead of researching from scratch.
- Sales engineering closed two enterprise deals where the customer explicitly cited the in-product AI-Act surface as a procurement signal.
- The internal Notion page got deleted. There was no longer a reason to maintain a hand-curated copy of regulatory state.
- The team's monthly bill for AI Act Radar was €299 (Team plan). Their fully-loaded cost of a single hour of in-house counsel time was higher.
What this team did not do
They did not build a legal-advice surface. They did not write summaries of articles. They did not interpret severity for the customer. Every event the customer sees links to the official source and carries the disclaimer "information only: not legal advice". The team's product gained a regulatory data surface, not a legal opinion engine. That distinction is what kept them, and us, on the right side of the unauthorised-legal-services line.
The honest part
Two parts of this took longer than the headline number suggests. The annex_iii_category tag on each customer deployment did not exist before the integration: the team had to add it to their internal data model, which took a half-day of upstream work. And the customer-dashboard widget went through three design rounds before they removed every word that read like advice and kept only the title, date, source name, and a link. The integration is fast; the discipline of staying on the data side of the line is the part that took thought.
Try the same thing
The AI Act Radar quickstart walks through the same four steps with the boilerplate code you can copy. Pricing for the Team plan is documented at aiactradar.com/pricing. The MCP server is open source at github.com/ligea-gmbh/aiactradar-mcp: clone it and run it locally to see the tool surface before signing up.
AI Act Radar is operated by ligea GmbH, Karlsruhe. Not legal advice. Information only.