Telling people they are talking to a bot, and why it helps

The instinct is to hide it. A bot that seems human feels like better service, so businesses give it a name, drop in some warmth, and let customers assume. It works until somebody realises, and the moment of realisation is what does the damage.

The realisation always comes. Somebody asks something unexpected, notices a reply arrived at 3am, or simply gets the feeling. What they remember is not that a bot answered them. It is that they were allowed to believe otherwise.

What we'll cover

• Why concealment fails
• Disclosing without sounding cold
• What customers actually want to know
• Impersonation, and what a number proves
• Building it with CodeWords
• Being wrong well
• The handover, which is where trust is made
• What you are holding about them
• The behaviours that cost trust quietly
• Making it survive
• Limits worth knowing about
• What to change first
• Frequently asked questions

Why concealment fails

Three reasons, and none is about ethics.

It is discoverable. Anyone can ask something off-script, and most people eventually do.

It sets the wrong expectations. Somebody who thinks they are talking to a person assumes the business now knows something. When it turns out nobody read it, that is a service failure produced entirely by the pretence.

It makes you look like the thing customers are afraid of. People are increasingly wary of who is on the other end of a message, for good reason. A business being evasive about that is behaving like the messages they have been told to distrust.

Disclosure costs almost nothing. A business that says plainly what is automated gets the benefit of the doubt on everything else, which is worth considerably more than the impression of a larger team.

Disclosing without sounding cold

The objection is that disclosure kills the warmth. It does not, if it is done once and briefly.

Say it in the first message, in a subordinate clause rather than a disclaimer. Something to the effect that this is the automated line, it can help with the common things, and a person is available.

Do not name it like a person. A human name on a bot is the concealment done subtly, and it is what people feel deceived by afterwards.

Keep the tone. Disclosure is one sentence. Everything after it can be as warm as you like, and warmth from something honest about what it is reads better than warmth from something pretending.

Do not over-apologise for it. A bot that is constantly reminding you it is only a bot is tiresome and signals low confidence.

Said once, clearly, at the start, it stops being a topic.

What customers actually want to know

Less than businesses assume, and specific.

Whether a person will see this. The most common underlying worry. Somebody who has typed something important wants to know it has not vanished into a system.

How to reach a person, easily and without arguing with a menu.

When they will hear back, if not now. A real timeframe rather than "shortly".

Whether their message is kept, especially if they mentioned something sensitive.

Answer those four and most of the anxiety in an automated conversation disappears. None of them requires the bot to seem human.

Impersonation, and what a number proves

Worth being direct about, because it is the trust problem that actually costs people money.

Businesses are impersonated on messaging channels constantly. Somebody sets up a number, copies the branding, and messages customers asking for payment or details. It works because customers have no easy way to tell.

Three things help.

Use a verified business profile with your real registered name, and tell customers what it will look like.

Tell them what you will never ask for. Card details, passwords, or a payment to different bank details. Say it in the confirmation messages where people will actually see it, not in a policy nobody reads.

Never change payment details by message. If yours genuinely change, that is a letter and a phone call, because redirection fraud works by looking like an ordinary update from a supplier somebody trusts.

The one thing that does not establish identity is the number itself. Numbers are recycled, devices are shared, and a display name is whatever somebody typed.

Building it with CodeWords

CodeWords connects to more than 3,000 integrations, and the connection is made once and reused.

1. Open CodeWords and start a new automation.
2. Describe what should happen in plain language to Cody, the automation builder, including how it introduces itself and what it says when it does not know.
3. Authorize the connections to your systems and your WhatsApp Business Platform account.
4. Describe the exceptions: anything it should not answer, anything a person must see, and what it must never ask a customer for.
5. Ground every answer in your own information rather than leaving the model to answer from general knowledge, which is what prevents the confident wrong answer.

You describe the outcome; Cody builds it, connects it, and deploys it. The free plan covers light use, with Pro at $39 per month and Business at $100 per month as usage grows; details are on the pricing page.

Being wrong well

Every automated conversation will be wrong sometimes. What decides the damage is the manner.

Say what it does not know rather than producing something plausible. A confident wrong answer costs far more than an admission, because the customer acts on it.

Do not guess at prices, dates or availability. These are the three things customers act on immediately, and a wrong one produces somebody arriving expecting something.

Make the failure message useful. "I can't help with that, but Jo can, and she's here until five" is a good answer. An apology followed by nothing is the one people screenshot.

Correct visibly. Where the business gets something wrong, saying so plainly rebuilds more trust than never erring would have preserved, which is counterintuitive and consistently true.

The handover, which is where trust is made

The moment somebody asks for a person is the highest-stakes moment in the conversation, and the commonest place to lose them.

Do not make them ask twice. A bot that tries once more to help after being asked for a human is the single most irritating pattern in the category.

Pass the conversation over. Somebody who has explained their problem should not explain it again. Repeating themselves is the specific thing that makes people feel the automation wasted their time.

Be honest about the wait. If it is tomorrow morning, say tomorrow morning. A vague promise that turns into a day of silence is worse than an honest delay.

Make sure the person actually exists. A handover to an unattended inbox is worse than no handover, because the customer is now waiting.

What you are holding about them

Automated conversations accumulate more than people expect, and customers increasingly ask.

Be able to answer plainly what is stored, for how long, and who can see it. Keep clinical, financial and personal detail out of message content where a prompt to get in touch would serve, since notification previews are visible to whoever is near the phone. Apply a retention period you chose rather than letting history accumulate by default. And honour a request to stop immediately and across every automation, not just the one that sent the message.

Doing this well is invisible. Doing it badly is the thing that ends up being complained about publicly.

The behaviours that cost trust quietly

Messaging people who never opted in.

Ignoring a stop request on one automation while honouring it on another.

Sending at antisocial hours, which reads as a machine regardless of disclosure.

Over-personalising. Referencing something the customer did not tell you in this conversation is unsettling rather than impressive.

Asking the same questions repeatedly across conversations, which tells them nothing is being retained and nobody is paying attention.

Making it survive

Disclose in the first message, every time, briefly.

Keep a list of what you will never ask for, and repeat it where customers see it.

Route handovers to somebody real with a stated time.

Read a sample of conversations weekly, looking for confident wrong answers.

Honour opt-outs across every automation, immediately.

Limits worth knowing about

Disclosure does not excuse a bad bot. Being honest about what it is does not make a useless one acceptable.

A phone number proves nothing about identity, in either direction.

You cannot prevent impersonation, only make it easier for customers to spot.

Warmth is not trust. A friendly bot that gets things wrong is less trusted than a plain one that does not.

Trust is asymmetric. It is built slowly through ordinary competence and lost in a single conversation.

What to change first

The first message. One sentence saying this line is automated, what it can help with, and that a person is available. It takes five minutes, it removes the moment of realisation entirely, and it is the single highest-return change available here.

Two habits make the difference. Never let the bot guess at a price, a date or availability, because those are the three things customers act on immediately and a wrong one produces somebody standing in front of you expecting something. And when a person is asked for, hand over on the first request with the conversation attached, since being made to ask twice and then repeat yourself is what people actually remember.

Frequently asked questions

Should I tell customers they are talking to a bot?

Yes, in the first message, briefly. Concealment is discoverable and the moment of realisation is what damages trust, far more than the automation itself ever would.

Does disclosure make the experience feel cold?

Not if it is one sentence at the start rather than a repeated disclaimer. Warmth from something honest about what it is reads better than warmth from something pretending.

Should a bot have a human name?

No. A human name is concealment done subtly, and it is the specific thing customers feel deceived by when they work it out.

What do customers most want to know?

Whether a person will see their message, how to reach one, when they will hear back, and whether what they said is kept.

How do I protect customers from impersonation?

Use a verified business profile with your registered name, tell customers plainly what you will never ask for, and never change payment details by message. A number and a display name prove nothing.

What should a bot do when it does not know?

Say so, and point at somebody who does with a real timeframe. A confident wrong answer costs far more than an admission, because the customer acts on it.

What should a bot never guess at?

Prices, dates and availability. Those are the three things customers act on immediately, and a wrong one produces somebody arriving expecting something you did not promise.

What is the worst handover mistake?

Making somebody ask twice, then making them explain again. Hand over on the first request with the conversation attached and a stated time.

What quietly erodes trust?

Messaging people who never opted in, honouring a stop request on one automation but not another, sending at antisocial hours, over-personalising, and re-asking questions already answered.

Can trust be rebuilt after a mistake?

Yes, and correcting something visibly often leaves more trust than never erring would have preserved. What is hard to recover from is having allowed a false impression deliberately.

Related reading

• What you are now holding
• Grounding, refusal and tone
• Keeping conversations separate