Say it is an agent. Every team that hid it lost CSAT

Three of the teams we work with wanted to leave the disclosure out. The argument is always the same and it is not a stupid one: say it is a machine and people will disengage before they have read the answer.

All three put it back within a month.

What actually happened

Satisfaction did not fall when customers were told. It fell a week or two later, when somebody worked it out on their own — usually from a reply that arrived at 03:14 on a Sunday, or from the fourth message in a row that answered the question they had not asked.

In two of the three cases we can date it fairly precisely, because both teams ran a public forum. In each one the decline begins within days of the first post where a customer says, in their own words, that they think they are talking to a bot. Nobody in those threads is angry about the automation itself. They are annoyed at having been allowed to assume otherwise.

The size of the fall was consistent enough to be worth stating plainly: between four and seven points of CSAT in each of the three teams, sustained for the rest of the measurement window, against no detectable movement in the weeks when disclosure was added.

Being told costs you nothing. Being caught costs you the benefit of the doubt on every reply after it, including the ones a person wrote.

The wording matters less than you would think

We have watched teams spend a fortnight on the sentence. In the data it barely registers. “You’re chatting with our support assistant”, “This is an automated reply”, and a plain “I’m an assistant, not a person” all land within noise of each other.

Two things do register. The first is where it sits: at the top of the first reply, above the answer, not in a footer under the signature. The second is what follows it — one obvious way to reach a person.

Teams that disclose and then bury the handover do worse than teams that never disclosed at all. That is the only result in this whole exercise that genuinely surprised us, and it is the one we now lead with. A disclosure without an exit reads as a warning that nobody is coming.

The teams that never told anyone

The obvious question is what happens if you simply never disclose and never get caught. We have two inboxes in the sample that ran that way for most of a year, and the honest answer is that the second half of that sentence does not hold for long.

Both were found out — once through a customer pasting two replies side by side and noting they shared a sentence, once through a journalist. In both cases the recovery took longer than the eight weeks we were measuring, and in one it required a public note from a founder. Set against a disclosure that costs nothing measurable, it is a strange trade to keep making.

There is also a quieter cost while the secret holds. Agents write to a standard of warmth that a machine can imitate but not sustain across a difficult thread, and teams that are hiding the automation tend to over-correct: longer replies, more apologies, more padding. Customers read that as evasiveness long before they read it as software.

Do not give it a name

A named assistant with a friendly first name is the fastest route to the failure above. It reads as a person, it gets thanked as a person, and when it turns out not to be one, the disclosure you wrote at the top is remembered as fine print.

We ask teams to use a role rather than a name. “Support assistant” is fine. Anything that would fit on a lanyard is not. The same goes for avatars: a photograph of a person attached to a machine is a claim, not a decoration.

What the handover has to do

Asking for a person should be one sentence, and it should work without the customer repeating themselves. If the human who picks it up opens with “can you tell me a bit more about the issue”, the disclosure has bought you nothing — you have simply moved the frustration one step down the queue.

In practice that means the thread, the order, the article the agent was working from and the draft it had written all arrive together. The person reads for ten seconds and answers. Teams that get this right see handovers resolved faster than conversations that started with a person, because the groundwork is already done.

What we ask teams to do

Four rules, and they take an afternoon:

Disclosure is not only a compliance question, though it may also be one where you operate. It is the thing that keeps the rest of the queue believable. A customer who knows which half they are talking to reads both halves more generously — and the half written by your team finally gets credit for being written by your team.