Somebody wants to buy from you at 11pm. They have one question — does this ship to their state, is the blue one in stock, can they return it if it doesn’t fit — and they cannot buy until they know the answer.
Nobody is awake. The question goes into an inbox, and by morning they have bought it somewhere else.
That is the problem a chatbot solves. Notice how narrow it is.
Why people hate these things
Say “AI chatbot” to most people and they picture the worst customer service experience of their year. A widget that will not let them reach a person, that answers a question they did not ask, that loops them back to an article they already read.
That reaction is earned, and it is not really about the technology.
It is about what the bot was put in front of. When a company has people available and installs a bot between you and them, that is a wall. Everyone can feel the difference between a company helping them and a company avoiding them.
At 11pm, there is no wall, because there was nobody behind it. The bot is not standing between the customer and a human — it is standing where nothing was.
Same technology. Opposite experience. The entire difference is what it replaced.
The question that sorts good from bad
Before building one, answer this: if the bot does not exist, what happens to this customer at this moment?
If the answer is “they reach a person,” you are building a wall. Every dollar you save is coming out of someone’s experience, and they will notice.
If the answer is “nothing happens, they leave,” you are adding coverage where there was none. Nobody is being deprived of anything, because the alternative was silence.
Most bad chatbot experiences come from businesses that answered that question honestly, did not like the answer, and built the thing anyway.
Two follow-ups sharpen it further.
Can the customer reach a human, and how fast? Every bot needs an exit. Not a phone number buried in a menu — a visible path out, taken as soon as someone asks for it or gets frustrated. When I designed an intake assistant for a nonprofit, we built an explicit escape path for families whose situation did not fit the options, because the ones who need a person most are exactly the ones a decision tree fails.
Does it complete anything, or does it only deflect? Telling someone where to find the shipping policy is deflection. Looking up whether you ship to their zip code, checking whether the item is in stock, and taking the order is service. The first one makes people angry. The second one makes them customers.
The measurement problem nobody talks about
Here is the part that decides whether a bot turns soulless, and it happens before anyone writes a line of it.
The moment you measure a chatbot on containment — tickets deflected, percentage of conversations that never reached a human — you have guaranteed the version everyone hates. You have told the system that not reaching a person is the win, and everything downstream gets designed to make escape harder.
Measure it on questions actually answered, and on orders completed outside business hours, and the same technology gets built differently. The exit gets easier to find, because a customer who needed a human and got one is a success rather than a failure.
Design does not make a bot corporate. The number you judge it by does.
What a real one has to be able to do
A bot that answers from a static FAQ is a search box with a personality. It will disappoint everyone, and it is most of what gets sold.
A useful one reaches live data. For an e-commerce business, that means the actual catalogue, real stock levels, real shipping rules for the customer’s actual location. I built one for a Mexican e-commerce operation that handles product questions on WhatsApp overnight and checks coverage by postal code before promising anything, because a promise the business cannot keep is worse than no answer.
It also has to know what it does not know. A bot that invents a shipping date is generating a complaint with a delay built in. Refusing to answer and handing off is the correct behavior, and it has to be designed in deliberately.
That is the difference between the version that works and the version people complain about online. Not the model. The plumbing behind it.
What it costs
Setup runs $3,500 to $6,000 for a deployment on an existing system.
That covers scoping what it must handle, training and configuring it against your real products and policies, channel setup and verification, testing against actual customer conversations rather than a demo script, training your team, and the first thirty days of tuning while real people find the gaps.
Running it is $400 to $700 a month, covering management, monitoring, and message volume up to an agreed ceiling, with overage beyond that.
Two things move you inside those ranges. The channel matters: a widget on your own site is straightforward, and messaging platforms like WhatsApp bring business verification and message template approval, which is more waiting than work but it is real. And whether the bot only answers or also transacts — booking, quoting, checking coverage — which is the difference between reading and doing.
The tuning line is the one people skip. A bot is not finished at launch, it is finished after a month of real customers asking things you did not anticipate. Budgeting zero for that is how you end up with the bot everyone complains about.
When not to build one
If you have people available during the hours your customers ask questions, and they answer promptly, do not build this. You would be putting a wall where there is currently a person, and your customers will like you less for it.
If your question volume is low, do not build this either. Ten questions a week is a person’s job for twenty minutes a day, and no arrangement of the numbers above makes that worth thousands of dollars.
If the honest reason you want one is that responding to customers is annoying, that is worth sitting with. A bot built from that motive tends to produce exactly the experience the motive implies, and customers are better at detecting it than most owners expect.
And if your product needs a real conversation to sell — high value, high consideration, lots of variables — a bot is not going to have it. It can qualify and hand off. It should not be closing.
What to check before you build anything
Pull your customer messages from the last month and sort them by the hour they arrived. Count how many landed outside the hours anyone was working.
Then read them and mark which ones a bot could have actually resolved with access to your real data, not just pointed at a page.
That second number is what you would be buying. If it is small, the answer is no, and you have saved yourself several thousand dollars and a chatbot your customers resent.
Ready to see how a chatbot can help your business? Book a free Workflow Review and let’s map out where we could help.
