A chat transcript is a written record
Agents type quickly, informally, and under time pressure, which makes it easy to forget that a chat
is a durable written statement from your company. It can be screenshotted, forwarded, quoted back
months later, and in a dispute it is evidence.
That is not a reason to make agents stiff and defensive. It is a reason to be clear about the small
number of sentences that create obligations, so that everything else can stay relaxed.
Advice you are not qualified to give
The clearest line is professional advice. If your product touches health, money, law, tax,
insurance or safety, visitors will ask agents questions that sound like product questions and are
actually requests for professional judgement.
“Should I take this instead of my usual one?” “Is this deductible?”
“Am I allowed to fire someone for this?” A well-meaning agent who answers has given
medical, tax or legal advice on your company's behalf, and the fact that they were being kind does not
help afterwards.
The move is to answer the product half and decline the professional half in the same message. You
can say precisely what the product does, what the options are, and that the choice between them is one
for their doctor, accountant or solicitor. Agents need explicit permission to do this, because without
it the instinct to be helpful wins.
Guarantees nobody can keep
The second class is the accidental promise. These slip out as reassurance rather than commitment,
which is exactly why they are dangerous.
“That will definitely be fixed by Friday.” “You will not be charged
again.” “It is completely secure, nobody can access it.” “We will never
change that.” Each is a commitment on behalf of engineering, billing, security or the company's
future, made by someone with no authority over any of them.
The honest version is nearly as comforting and costs nothing: what is true now, what is expected,
and who owns the outcome. “The fix is scheduled for this week's release and I will message you
when it ships” is a promise you can keep. “Definitely Friday” is one you cannot.
Absolute security claims deserve their own mention because they are both unkeepable and easy to
disprove. Describe what you actually do — conversations run over an
encrypted connection, access is limited to the people who need
it — rather than reaching for the word “unhackable”.
Speculating about a cause
During an incident, agents get asked what happened, and the temptation to theorise is strong. An
agent guessing at a cause is your company's public position on an event you have not investigated.
“I think we were attacked.” “It was probably the vendor.” “Someone
must have deleted it.” If any of those turn out to be wrong, you have published a false
statement about a security event or blamed a partner in writing. Agents should describe impact and
status, not diagnosis, until there is an agreed line.
The same applies to blaming a named colleague or another department in a transcript. Whatever
internal frustration exists, it should not be written down in a customer-facing record.
Confirming things you have not verified
The fourth class is quieter: confirming information about an account to whoever is in the window.
Chat feels like a conversation with a known person, and it is not. You know the page they are on and
whatever your site passed along, and none of that establishes identity.
So reading back an address, listing recent orders, or confirming that an account exists at a given
email are all disclosures to an unverified party. The rule that holds up is that agents may discuss
what the visitor has already proven they know, and anything further moves to a channel where identity
is established. Never take card numbers or passwords in chat at all — not to be careful with
them, but because they should not enter the transcript in the first place.
Write the list, then make it easy to follow
A policy nobody can recall under pressure is decoration. Keep the list to a single page, phrased as
“instead of X, say Y” rather than as prohibitions, because agents need a sentence to type,
not a rule to remember.
Then put the replacements where the typing happens. Saved replies for the three or four hardest
declines — the advice request, the deadline question, the identity wall — mean the careful
version is one shortcut away, at the moment when improvising is most likely.
Review the list after real conversations rather than in the abstract. The best source is your own
transcripts: search for the phrases that overpromise and you will find where the pressure actually
falls, which is rarely where the policy assumed.
What to measure
Sample transcripts for the promise words — definitely, guaranteed, never, always, immediately.
A rising count is an early warning that the team is under enough pressure to reassure past what they
know, and it usually tracks staffing rather than training.
Count how often agents correctly decline an out-of-bounds request rather than how often they get
one. A team that never declines anything is not being asked easier questions than everyone else.
Track the conversations that had to move channel for verification. If that number is very low,
either your product touches nothing sensitive, or the wall is not being enforced — and the
transcripts will tell you which.