About
CasaReply is an AI employee that knows its limits.
It answers your customers on WhatsApp using your own business information, and when it does not have the answer, it says so and fetches a person. That second half is the product.
The problem
Every small business now runs a customer service operation it never chose to run.
A small business today runs a customer service operation it never chose to run.
Customers message on WhatsApp because it is where they already are. They ask the same things: the price, the hours, whether you do this, whether you can help. They ask at nine in the evening and on a Sunday, and they compare how quickly you answer to how quickly the last business answered.
The person replying is usually the person who should be doing the actual work. The answers live in four places: a price list in a PDF, a policy on the website, opening hours that changed last month, and everything else in somebody’s head. Two people answer the same question slightly differently and the customer remembers the cheaper version.
Meanwhile the conversation that mattered, the one from someone genuinely about to buy, is somewhere in the same unread list as a wrong number.
Why CasaReply exists
The gap between hiring and a generic chatbot.
Two things already exist for a business tired of answering the same WhatsApp message for the fifth time today, and CasaReply is neither of them.
One is hiring: putting a person on messages full time, which works, and means paying a salary to answer a handful of questions asked slightly differently, all day, every day.
The other is a generic chatbot: cheap, always on, and willing to answer anything, including questions it has no real information about. It will quote a price it invented and sound completely sure while doing it.
CasaReply sits in the gap between those two: an assistant that only ever says what a business actually told it, so it is safe enough to leave running, and specific enough to that business to actually be useful.
What it does
Five things, in order, every time a customer writes in.
- 1
Connects to WhatsApp
The number your customers already message you on, through the official Meta Business API.
- 2
Learns your business
Prices, hours, services, policies, whatever you give it, stored as the only facts it is allowed to state.
- 3
Answers, or says so
Replies from what it knows. Asked something it does not, it says that plainly instead of guessing.
- 4
Keeps a record
Every contact, every conversation, and who became a real, verified client rather than a claim.
AI approach
A component with a known failure mode, not an oracle.
CasaReply treats a language model as a component with a known failure mode, not as an oracle.
The failure mode is specific and well documented: given a question it cannot answer from what it has been shown, a model will produce something that reads exactly like an answer. It will not hesitate, hedge or sound uncertain, because sounding certain is what it was trained to do.
Every reasonable use of one in a customer-facing product follows from taking that seriously. So the model here does not talk to your customers directly. It returns a structured decision (what it thinks the customer is asking, what it proposes to say, and which of your information it drew on), and that decision is checked before anything is sent.
An answer about your business with nothing behind it does not get softened or qualified. It is discarded, and the conversation goes to a person.
What that means in the code
- The model never sees another business’s information. Retrieval is scoped in the database, not filtered afterwards.
- Customer messages are never placed where instructions go, so a customer cannot rewrite the assistant’s rules by typing them.
- Text extracted from uploaded documents and web pages is fenced and labelled as data: quoted from, never obeyed.
- A reply claiming an action, such as booked, checked, sent or paid, is rejected, because none of those happened.
- The assistant can promote someone to lead. Only a person on your team can confirm a client.
Human + AI
Not instead of your team. Alongside it.
The assistant handles what is genuinely repetitive. The moment a conversation needs judgement, it goes to a person, with the full history attached.
The AI handles the conversation
A situation requires a person
Your team takes over
Core principles
Five decisions, and the reasoning behind each.
These are positions, not values. Each one costs something, and each one is visible in how the product behaves.
The AI answers from your business, or not at all
A general-purpose model has read a great deal about businesses like yours. None of it is about yours. It will produce a plausible price, and plausible is the problem.
A customer who arrives quoting a price you never set is worse off than one who waited an hour for a real answer.
“I don’t know” is a feature, not a failure
Most of the risk in automating customer conversations is not a rude reply or a clumsy one. It is a confident, specific, wrong one, delivered in your name, at speed, to someone who believes it.
An assistant that declines ten questions a day and gets nothing wrong is worth more than one that answers all ten and gets two wrong.
A person is always one message away
The assistant steps aside when someone asks for a human, when a complaint arrives, when it lacks the information, and when the same misunderstanding repeats.
Automation that makes a customer fight to reach a person is worse than no automation. Everyone has met that phone menu.
The business stays in control of what is said
You choose what it knows, how it sounds, what it does outside your hours, and whether it is on at all. There is no prompt to write and no model to configure.
Those controls look like power and function as a trapdoor: a free-text prompt is how an assistant gets talked into saying something it should not.
A record you can check
Every decision is logged with its reason, including what information was missing when it could not answer. Settings changes and client verifications name who made them.
An assistant you cannot audit is one you have to take on faith, and taking software on faith is not a reasonable thing to ask of someone’s livelihood.
Trust and reliability
What you can rely on, not just what the model does.
Nothing is stated without a source
A reply about your business is checked against your own information before it is sent. Unsupported, it is discarded rather than softened.
Claims are not facts
A customer saying they booked or paid is recorded as a claim. Only a person on your team can confirm it, and the record shows who did.
Every decision leaves a trail
What was asked, what was found, what was said, and why it was or was not, logged as it happens, not reconstructed afterwards.
Your data leaves when you do
Your business information, contacts and conversation history are yours. No feature is held back to make leaving harder.
The legal pages describe what the software actually does today, plainly, rather than reading like a template. See Privacy.
Try it on the questions you actually get asked.
Add a few of your services and prices, and see what your assistant will and will not say.
No card required. CasaReply is not charging yet.