How to Word an AI Disclosure in a Support Email Without Killing CSAT

Two disclosure lines. Same legal obligation. Very different outcomes.

The first: “Hi Anna, I'm Robin, the AI assistant for [Company]. I've pulled up your order and here's where things stand.” The second: “This message was automatically generated by an artificial intelligence system. Do not reply to this email.”

The second line is the reason so many support leaders believe disclosure costs them satisfaction scores. It isn't the disclosure doing the damage. It's the sentence that comes after it.

The 79.7% figure everyone quotes, and why it doesn't describe your inbox

Every time a support team debates AI disclosure, someone brings up the Luo et al. field experiment published in Marketing Science in 2019. Across more than 6,200 customers, revealing chatbot identity before the conversation started cut purchase rates by over 79.7%. It's a genuinely startling number and it has done more to discourage honest disclosure than any other single piece of research.

It also describes a situation almost nothing like yours.

That study measured outbound sales calls to people who hadn't asked to be contacted, and the outcome was whether they bought something. A support email is the opposite setup on every axis. The customer started the conversation, they already want an answer, and the outcome that matters is whether their problem got fixed. Later service-context research points the other way: in an experimental study in the German fashion retail sector, disclosing chatbot identity moved neither trust nor customer retention, while service outcome moved both.

Which lines up with what we've seen across deployments. When teams tell us disclosure tanked their scores, the AI was usually also failing to resolve anything, and the disclosure just gave customers a name for their frustration. Fix the resolution rate and the disclosure debate mostly evaporates.

What the law actually makes you say

Since 2 August 2026, Article 50 of the EU AI Act has applied to any AI system that interacts directly with a person, whatever its risk tier. A support agent answering customer email is squarely inside that. Non-compliance carries administrative fines reaching EUR 15 million or 3% of worldwide annual turnover, whichever is higher.

Article 50(5) sets the standard for how you say it. The information has to be clear and distinguishable, delivered at the latest at the point of first interaction, and in an accessible format. The Commission's guidelines, adopted on 20 July 2026, are unusually specific about what falls short:

• A generic reference to an “assistant” doesn't satisfy the obligation. Naming your agent Robin is charming, but on its own it tells the reader nothing about what Robin is.
• Machine-readable marking alone doesn't count as informing a person.
• Anything buried in terms and conditions fails, as does a faint line in a page footer.
• Where an AI agent acts autonomously, it should also disclose on whose behalf it's acting, and reaffirm that at key moments such as authorisation and reporting back.

There's an exception where the AI involvement would be obvious to a reasonably well-informed, observant person. Don't build your compliance posture on it. An email that opens “Hi Anna, thanks for getting in touch” and reads like a human wrote it is the definition of not obvious.

What the regulation conspicuously doesn't do is dictate wording. That's the part worth paying attention to, because it means the CSAT outcome is entirely in your hands.

Four jobs the line has to do

A disclosure that survives contact with real customers does four things at once:

1. Identifies: says plainly that an AI is answering, in words a reader processes in under a second.
2. Attributes: names the company the agent is acting for, which is both the Article 50 expectation for autonomous agents and the thing that stops the message feeling like it came from nowhere.
3. Offers an exit: tells the reader how to reach a person, in the same breath.
4. Doesn't apologise: no hedging, no pre-emptive excuses. A line that sounds embarrassed to exist teaches the reader to distrust everything under it.

Miss the third one and you get the “do not reply” disaster. Miss the fourth and you get something worse: an answer the customer discounts before reading.

15 lines you can ship today

These are grouped by what the agent is actually doing, because the right wording changes depending on whether it's answering, acting, or stepping aside. Swap the bracketed values for your own.

The agent resolves it end to end

• “Hi [Name], I'm [Agent], the AI support assistant at [Company]. I've looked into your order and here's where it stands.”
• “You're being helped by [Company]'s AI support agent. Prefer a person? Reply with HUMAN and I'll hand this over straight away.”
• “[Company] AI Support · answered automatically · reply to this email any time to reach our team.”

AI drafted it, a human pressed send

This is the mixed-mode case, and it's the one teams get wrong most often by claiming more human involvement than there was.

• “Drafted by our AI assistant, reviewed and sent by [Agent Name], [Company] Support.”
• “This reply was prepared with AI assistance and checked by someone on our team before it went out.”
• “[Agent Name] here. I used our AI assistant to pull your account history together, and I've checked everything below myself.”

The agent is doing something, not just saying something

• “I'm [Company]'s AI support agent, acting on behalf of [Company]. I've issued your refund of EUR 41.00 to the card ending 4417.”
• “AI assistant for [Company]. Before I cancel this booking, can you confirm the reference is BK-88231?”
• “Done, your delivery address is updated. This change was made by [Company]'s AI agent and is live on your account now.”

Notice that all three name the company alongside the AI. That's the “on whose behalf” requirement, and it costs you four words.

Handing the thread to a human

• “I'm the AI assistant and this one's beyond what I should settle on my own. The full thread is now with [Team], who'll reply by [time].”
• “AI agent here. I've stopped short deliberately, because this needs a judgement call. A specialist is picking it up now.”
• “Handing over to [Name]. Everything you've already told me has gone across with the thread, so there's nothing to repeat.”

The third line doesn't re-disclose, and it doesn't need to. The obligation attaches to the first interaction. Once the customer knows, a handoff message can just be useful.

High-stakes, regulated, or already-angry threads

• “You're reading a reply from [Company]'s AI support agent. Given what you've described, I've flagged this to our complaints team and a person will contact you within one working day.”
• “AI assistant for [Company]. I can confirm your policy details below. Anything involving a claim decision goes to a licensed adviser, who will be in touch.”
• “This is [Company]'s AI agent. I've answered what I can from your account record. I haven't made any decision about your dispute, and I can't; that sits with our payments team.”

Pick three of these, not fifteen. Consistency across a thread matters far more than finding the perfect sentence, and a customer who sees three different disclosure styles in one conversation trusts the whole thing less. The same discipline applies to the rest of the reply, which is why auto-replies that don't sound robotic depend as much on consistent voice as on good copy.

Wording that reliably backfires

• “As an AI language model...” Model-speak. It reads as raw, unedited output and immediately lowers the perceived care behind the response.
• “Do not reply to this email.” The single most expensive sentence in support automation. You've disclosed and slammed the door in the same breath.
• “I'm just a bot, so I might get this wrong.” Pre-apologising invites distrust before the customer has read a word of the answer.
• A cute name with no company attribution. Fails the Article 50 guidance, and reads as evasive to boot.
• Eight-point grey text in the footer. Not clear, not distinguishable, and a regulator will say so.
• “Powered by AI” as a badge. That's branding. Disclosure is a statement to a person, not a logo.

Placement beats phrasing

Once the wording is decent, position is where the remaining CSAT movement lives. Teams tend to reach for a banner above the greeting because it feels safely prominent. It isn't. A disclosure banner sitting above “Hi Anna” is the email equivalent of a terms-and-conditions checkbox: technically present, cognitively absent. Readers have been trained for two decades to skip anything that looks like system chrome.

Signature-only placement has the opposite problem. On a four-paragraph reply about a billing discrepancy, a line at the bottom arrives long after the reader has formed a view about who wrote it, which is hard to square with disclosure “at the latest at the time of first interaction”.

What works is the boring middle option.

Put the disclosure adjacent to the greeting on the first AI-handled message in a thread, where it reads as part of the introduction rather than a legal notice. Then carry a short persistent marker in the signature of every subsequent AI message, so the customer never loses track of who they're dealing with three replies in. That pairing satisfies the timing requirement and the “maintained throughout the experience” expectation without ever feeling like a warning label.

Testing it without guessing

Most teams change the disclosure line, watch CSAT for a week, and declare a result. That's not enough signal to act on, and CSAT is the least sensitive of the four numbers worth tracking here.

• Human-request rate: the share of AI-handled threads where the customer explicitly asks for a person. In our experience this moves earlier and further than CSAT, and it's the cleanest read on whether the disclosure reassured or alarmed.
• Reopen rate: whether “resolved” threads come back. A disclosure change shouldn't touch this. If it does, something else changed too.
• Time to first human touch on threads that do escalate.
• CSAT, segmented by intent. Aggregate CSAT hides the thing you need to see, which is that disclosure lands very differently on a WISMO query than on a complaint.

Run it as a holdout on 10–20% of eligible volume for a fortnight rather than a hard cutover, and keep the segments large enough to mean something. One more thing, easy to forget in the rush: log which variant each customer received, and log it in a way you could produce eight months later. If a regulator asks what a specific customer was told on a specific date, “we were running an A/B test at the time” is not an answer. The evidence side of this is a discipline in itself, and we've written separately about the audit evidence Article 50 actually calls for.

Where disclosure sits alongside the broader question of customer confidence, the transparency playbook for AI email responses covers the surrounding practices: confidence framing, error recovery, and how much of your reasoning to show.

Where this sits in the stack

The operational risk isn't writing a bad line. It's that a good line quietly drifts. Disclosure text that lives in an editable email template will get shortened by someone tidying up the signature block, and nobody will notice for months.

Robylon's email agent treats disclosure as a property of the agent rather than a string in a template, with per-locale variants across 40+ languages and the acting entity bound to the reply itself, so multi-brand deployments can't cross-attribute. Escalation keywords route to a human on the first request rather than looping the customer through another AI turn, which is the behaviour that makes an exit line credible. The agent resolves 60–80% of email autonomously and takes action across 60+ write-access integrations, and every one of those actions carries the acting-agent identity with it. Deciding which threads it should hand over in the first place is a separate design problem, covered in our guide to when to resolve versus route to a human.

Write the line once, well, and stop treating it as a tax. Customers have never objected to knowing who they're talking to. They object to being stuck with them.

Ready to disclose confidently and still resolve at scale? Robylon AI resolves 60–80% of customer emails autonomously with AI agents that take action across Zendesk, Shopify, Stripe, Salesforce and 60+ other integrations. Start free at robylon.ai