What an AI receptionist cannot do, including ours

A receptionist, in the ordinary sense, answers a ringing telephone. Nothing on this page does that, ours included: mesej cannot answer a phone, cannot start a conversation, and has no channel connected to it today, so it has not answered anybody’s customer. If the job you are hiring for is picking up when a customer calls, go and look at a voice product instead, and what a telephone answering service does that a messaging tool cannot is the comparison worth reading before you buy either.

The name describes a job the software is not allowed to do

On a messaging channel the limit is not ambition or engineering effort. It is written into the platforms, by the platforms. Telegram states it on its introduction for developers, in the section comparing bots to ordinary accounts:

Bots can’t start conversations with users. A user must either add them to a group or send them a message first.

That is categorical, not rate-limited and not priced: Telegram’s own FAQ notes that “bots are able to message their users at no cost”, and its paid tier buys throughput to people who already wrote to you, not permission to open a chat.

WhatsApp is the near miss rather than the parallel. Meta defines a window instead of a prohibition, and what opens that window and what resets it is the whole of the mechanism:

When a WhatsApp user messages you or calls you, a 24-hour timer called a customer service window starts.

Inside that window a business writes freely. Outside it, the WhatsApp Business Messaging Policy is blunt:

Outside the 24-hour customer service window, you may only send messages via approved Message Templates.

Meta’s own summary of what that leaves is the plainest line published on it: “non-template messages can only be used to respond - not reach out - to users”.

A template is the exception that proves it. Meta calls one “the only type of message that can be sent to WhatsApp users outside of a customer service window”, approves it in advance, and warns that “Review can take up to 24 hours”. A pre-worded, pre-cleared, separately billed message reviewed the day before is marketing machinery, not answering the door.

So an assistant on these two channels can do one thing: answer somebody who wrote to you, in the period the platform allows. That is a responder, not a receptionist, and every honest description of this category starts there.

What mesej specifically cannot do

It cannot answer a phone. Voice is a different product with different failure modes, and a wrong answer spoken aloud cannot be reviewed before it lands. On WhatsApp a customer’s call opens the same window a message does, in Meta’s own wording, and we still cannot take the call.

It never messages anyone first. No send in the repository accepts a raw destination: the two senders take a branded target, whose one constructor is named for the single caller allowed to mint it, and nothing but a test calls either sender today. So no follow-ups, no promotions, and nothing sent to people who never wrote to you: not an appointment reminder, which needs a template approved before the first send, and not a text back to a number that rang and got no answer. That is an unforgeable type plus a convention you can read, rather than a setting: a weaker claim than impossibility, and the accurate one.

It cannot reopen a conversation that has gone quiet. Once the window shuts you cannot reply through us either, and the conversation waits for the customer. The same clock is what leaves an enquiry that arrived at 11pm still answerable in your own words at 8am.

It cannot invent a price, a time or an address. It is specified to answer from what you have written down and to hand you anything it cannot find. A signed-in owner cannot supply those notes yet: there is no screen and no API route for them, so that half is a design and not a feature.

It cannot read a calendar kept anywhere else. If your day lives in another booking system, or in a paper book on the counter, it is specified to take the details and pass them to you.

It will say something wrong and you will carry it

You will, and Meta writes that into the contract. Hiring an AI product to work on your WhatsApp account makes it what the WhatsApp Business Solution Terms call a Third Party Service Provider, and the allocation of risk is one sentence: “You are solely and fully liable for all acts and omissions by your Third Party Service Providers”. No vendor can sign that away for you.

The objection is conceded. The only honest mitigation is an approval step, a doubtful reply reaching you rather than the customer, and here that check is a function called guard_reply, frozen behind a contract test, which returns either a send-ready envelope or a list of problems, and whose documented rule is that a reply with any problem does not auto-send. A reply claiming to be a person is one of those problems.

The limits on that are ours to state. Nothing has been sent through that check to a real customer, and there is no owner-facing approval queue in the console, where a signed-in owner can move or cancel an appointment and change a password. A gate that has never held back a live reply is a design, not a track record.

Google closed the route that looked most like a switchboard

Nor can the enquiries arriving through your Google listing be piped into an assistant. Google says so on its own help page: “As of July 31, 2024, the chat and call history features are no longer available in your Business Profile.” What exists now is a link rather than an inbox, since “Businesses with a claimed and verified Business Profile can add a WhatsApp or text messaging option to their contact info”. A profile is where you answer reviews, not where you hold conversations; the chat feature Google withdrew has the rest.

WhatsApp’s own rule says a person has to be reachable

The Business Messaging Policy permits automation and then immediately requires an exit from it:

You may use automation when responding during the 24-hour window, but must also have available prompt, clear, and direct escalation paths.

The paths it names run from a human agent transfer through to a support form. A product sold as a replacement for the person answering enquiries is contradicted by the policy of the channel it runs on.

On disclosure, Meta’s published requirement sits on the Instagram side of its documentation. Its Instagram Messaging overview repeats the escalation rule and adds one the WhatsApp policy does not:

When required by applicable law, automated chat experiences must disclose that a person is interacting with an automated service: at the beginning of any conversation or message thread, after a significant lapse of time, or when a chat moves from human interaction to automated experience.

We do not support Instagram, and quote it as the clearest thing Meta publishes on the subject. The WhatsApp policy we read carries the escalation requirement and no equivalent disclosure sentence, so on WhatsApp telling people is your local law’s question rather than Meta’s.

Whether it gets your number banned

Answerable from Meta’s pages rather than from optimism. Quality rating scores recipient behaviour, and Meta names the inputs:

WhatsApp determines message quality from how WhatsApp users have received your messages over the past seven days, weighted by recency. It bases this score on user feedback signals such as blocks, reports, mutes, archives, and reasons users provide when they block you.

Enforcement is a separate system with its own published ladder. Meta’s policy and spam enforcement page says that “Initially, WhatsApp Business Accounts will get a warning with information on the policy they violated”, that “Accounts can become restricted or disabled depending on the number and severity of issues”, and, connecting the two, that “We might also limit or offboard your business from WhatsApp if your account receives excessive negative feedback from users”.

Two qualifications. Neither Meta page cited here states a threshold: no block rate, no report count, no score at which anything happens. And that a reply to a question somebody just asked is unlikely to be blocked is our reading of Meta’s inputs, not a claim of Meta’s.

It will paste the same reply everywhere

The strongest objection in the category is published. Google’s v4 Business Profile reviews reference names a rejection reason for exactly this:

Content that violates our policy on repetitive content isn’t allowed. Repetitive content includes identical or similar content added multiple times by one or more accounts.

The Maps policy behind it prohibits “Posting the same content multiple times either from the same account or multiple accounts”. Repetition counts across accounts, not only within one, so a reply library shipped to a whole customer base is the pattern that sentence describes, and Google moderates a reply before anyone sees it, which the published rejection reasons covers. The only reliably non-repetitive reply is one the specific message made necessary.

The number we are not going to publish

There is a monthly search figure attached to this phrase. We measured it three ways, got three materially different answers, and one was visibly distorted by an artefact in the series, so we are publishing none: a figure we cannot stand behind is worth less than the admission that we cannot.

What is actually left, and whether it is enough

What survives every limit above is narrow and real: answering the question a customer already asked, in seconds, from what you have already written down, at three in the morning, and flagging anything it is unsure about. Replying to what people post about you is a separate discipline, covered in how to answer an angry one and what the tools in that market actually publish.

One more limit, on us rather than on the platforms: no channel is connected to mesej today, so its reply quality has never been measured against live traffic, and we will not assert a number for it from our own test fixtures. When there is one, it will carry the date we measured it.

If your enquiries arrive by phone, none of this helps. If they arrive as messages and you want them answered rather than chased, the limits on this page are the product rather than the small print.