What happens when the AI gets it wrong?

Nobody selling one of these writes this article, which is a good reason to read it. A reply written by software and sent under your name will, at some point, say something you would not have said. The useful question is not whether that happens but what kind of wrong it can be, and what stands between each kind and your customer.

The four kinds of wrong

Wrong facts. A price that is not your price, stock you do not have, a delivery window you cannot meet. This is the expensive one, because the customer acts on it and you are the one who has to unwind it.

Wrong tone. Accurate, and still not you. Too formal for a business whose regulars get first names, too breezy for one that does not. Nobody cancels an order over this, but it erodes the thing that made them your customer.

Wrong moment. Technically fine, sent into a situation that needed a person. A complaint answered smoothly. A bereavement met with an upsell. This is the one that does real damage, and it is not a language problem.

Wrong at all. The reply that misunderstood the question and answered a different one. Usually obvious, usually harmless, occasionally not.

What actually prevents each one

Being precise here, because this is where marketing usually goes vague.

Wrong facts are the one that can genuinely be engineered away, and the mechanism is unglamorous: the reply may only state things you gave it. Prices, stock and policies you have uploaded and approved are facts it can assert. Facts you never gave it are not available to be asserted, so the reply hedges or hands the message to you rather than filling the gap with something plausible. If a tool cannot tell you where a number in a reply came from, it is improvising.

Wrong tone is reduced, not eliminated. You set the register yourself, and on the paid plans it learns from messages you actually sent, which gets it closer than any setting. It will still occasionally land a sentence you would have phrased differently.

Wrong moment is the one worth choosing a product on. What matters is whether the tool knows when to stop. Anything that looks like it needs a person should not be answered automatically at all: the reply waits for you instead, and it tells you why it stopped. No tool catches every one of these, and any vendor claiming otherwise is describing a product that does not exist.

Wrong at all is mostly a function of how much context it had. A tool with no knowledge of your business misreads more questions, because it has less to anchor on.

The control that matters more than any of it

One switch turns automatic sending off. With it off, the product works identically except for the last step: every reply waits in the app for you to read, edit and send.

This is not a downgrade and it is not a lesser plan. It is the same engine with the final decision handed back to you, and for some businesses it is the right setting permanently. A clinic, a lawyer, anyone whose wrong sentence is a regulatory event rather than an embarrassment should probably run this way, and we would rather tell you that than sell you the automatic version.

When one gets through, do this

Correct it fast, in the same thread, in your own words. Do not explain the software. The customer does not care and the explanation reads as an excuse.
Honour it if it cost them something. If a wrong price went out and they ordered on it, the cheapest resolution is usually to eat the difference once.
Fix the source, not the sentence. A wrong price in a reply is a wrong price in what you uploaded. Correcting the item stops it happening again; apologising does not.
Read a day’s worth of replies afterwards. One error is usually a symptom of stale details rather than a one-off.

Who is responsible

Plainly: you are. A reply goes out from your account, under your name and your number, and it counts as a message you sent. That is what our terms say, and it would be true regardless of what they said, because that is how the customer will see it.

What we are responsible for is narrower and worth stating. The product should not invent facts you did not give it. It should stop when it should not answer. It should be switchable off in one action. And it should let you read what it sent. If it fails at those, that is on us.

What this means before you install anything

The reasonable version of this decision is not “will it ever be wrong” but “what does wrong cost me here”. For a shop answering stock questions, an occasional imperfect sentence is cheaper than a day of unanswered messages. For a business where a wrong answer is a legal problem, run it with sending off and use it as a drafting tool.

Both configurations are the same product and the same price. What each plan gives the model to work with is on the features page and the plans themselves are on the pricing page.

Reading a week of replies, properly

The single habit that separates people who get value from these tools from people who get burned is reading what went out. Not all of it, and not forever.

Week one, read everything. It is fewer messages than you think, and it is the week your uploaded details are least accurate. Nearly every error you will ever see is visible here.

After that, read the ones that got held back. A reply the tool declined to send is a signal about the edges of what you taught it, and it is a much higher yield of information per minute than reading the successful ones.

Once a month, read a normal day. Tone drifts as your business changes. Somebody asks about a service you stopped offering in March and gets a cheerful answer about it.

What you are looking for is not perfect sentences. It is the specific ones you would have written differently, because each of those is a correction to make once at the source rather than a sentence to regret repeatedly.

The comparison worth making

The honest benchmark for automated replies is not perfection. It is what actually happens today, which for most small businesses is that a proportion of messages get no reply at all, and a further proportion get one so late that the customer has already gone elsewhere.

An unanswered message is also a wrong answer. It is just one nobody counts, because it produces no artefact to be embarrassed by. Any decision about automation should weigh the errors it introduces against the silence it removes, and for a lot of inboxes that arithmetic is not close.

That is an argument for automating, not for automating carelessly, which is why this article exists at all.

The two questions that usually come before this one are whether an auto-reply gets your number banned and what these apps can read.