Arbitr
Use cases▾
Generic environmental claimsAI interaction disclosureAccessibility statement and conformity
Pay range in job postings
Prescribed risk warning in promotions
Auto-renewal and cancellation disclosure
Content localisation
AI discoverability
PlatformPricing
Resources▾
Blog
About▾
CompanyCareers
Book a demo
Home > Use cases > AI interaction disclosure (EU)

Your chatbot has to say it's AI.

On every chat surface you run, in the language of each market - a notice your customer cannot read has not told them anything. And that's only half the story. Are the claims your AI assistant is making accurate in every market and language you serve?

Trusted by
Oracle
Entrata
Trane Technologies
Ingersoll Rand
IBM
Spirax Sarco
SEAT
Varian
Mitutoyo
Panasonic
Oracle
Entrata
Trane Technologies
Ingersoll Rand
IBM
Spirax Sarco
SEAT
Varian
Mitutoyo
Panasonic
EU AI Act · Regulation (EU) 2024/1689, Article 50

Since 2 August 2026, people must be told when they are interacting with AI.

In force since
2 Aug 2026
Maximum exposure
€15m or 3%
Who it binds
Any provider or deployer whose AI output reaches people in the EU, wherever based
What it takes
A notice on the surface
That's the obligation. But is it the smallest part of your exposure?
Agent liability

Your assistant speaks for you.

An answer from your chat assistant is a statement your company made. Not a suggestion and not a draft - a commitment a customer can act on, and hold you to.
Returns

"You can send it back any time in the first 60 days."

The policy says30 days from delivery.
The company honours 60, or argues with a customer holding a transcript.
Capability

"Yes, that integrates with your existing system."

The product doesNot support it.
A deal closed on a promise nobody can deliver, and a refund conversation.
Price

"That discount is still available."

The offerExpired last quarter.
Honoured at margin, or withdrawn and complained about.
None of these needed a bug. The assistant was working exactly as built. What it's missing is a place to find and use your approved rules, claims and company information in real time and in any language.
Answer drift

It answers in every language you sell in. You have read one.

Your policy exists in English and was signed off once. The assistant does not translate it - it writes a fresh answer in whatever language the customer typed in. Nobody on your team has read most of them.
Source · approved
Returns accepted within 30 days of delivery
Policy v4 · Jan 2026
→
ENYou have 30 days from when it arrives.Pass
DESie können innerhalb von 60 Tagen zurückgeben.Hold
NLU kunt binnen 14 dagen retourneren.Hold
FRLe délai est de 30 jours après réception.Pass
The German answer is wrong. So is the Dutch. Both went out unnoticed, because the only version anyone has read is the English one.
1
Approved policy
11
Languages you answer in
1
Anyone has read
That is ten languages of answers going out today that nobody has reviewed.
See your AI answers checked.
Book a demo
The platform

How it works

You already have the answers. They live in your published policies, your approved claims register, your terms, and in what the product actually does. What none of those can do is check what your chatbot told a customer at four o'clock on Tuesday. Arbitr takes those sources and runs your assistant against them - continuously, in every language you answer in.
The old way

Slow, manual review

Send the file to human reviewers. Wait. Pay by the word. Find out what changed after it published.

With Arbitr

Content governance as a platform

Your rules run against every change before it goes live, in every market, at the speed of AI.

How a check runs

Whether it is a web page, a campaign or an answer from your assistant, everything runs through the same three stages.

STAGE 01

Checks

Content is read against your approved policies, claims and terms for that market.

Output · a decision per claim
→
STAGE 02

Flags

Anything that breaks a rule is held and the rule that held it is named, so the person who has to fix it knows what to change and who can approve it.

Output · a hold, with a reason
→
STAGE 03

Publishes

What passes goes live in the market it was written for, and the decision stays on the record.

Output · live content, and an audit trail

A hold is not a block. It is the product working, and it comes with a named owner and a next step.

See the platform in full ›The trust layer between AI and your live systems.
Remediation and governance

Fix today. Then stay that way.

Two jobs, in order. First we find every chat surface you have, check the notice is present and readable in that market, and correct the answers that do not match your approved sources.

Then we stay on. Every model update, prompt change and knowledge base edit moves the answers, and none of them go through a publication review.

Getting today's answers right is finite work. Every conversation after it is not.

Proof

How we help

37,000
Expert reviewers in the network, across 120+ languages
25+
Years working with regulated multinationals on multilingual content
IBM Gold Partner
Technology partner
FAQ

Questions we get asked

Have a surface or market this page has not covered? Tell us where your assistant answers, and in which languages.

Ask a pricing question ↗
Does the notice have to be in the local language?
The rule is that the person has to be informed. A notice in English on a surface serving German customers has not informed anybody, whatever it says. We check the notice is present and in the language of the market that surface serves.
What about our voice agents and email assistants?
Our assistant came from a vendor. Is this their obligation or ours?
How do you check what an assistant says without reading every conversation?
How much does it cost?
See your AI answers checked.
Book a demo