Sona

Most chatbots will guess. This one tells you it doesn't know.

A website assistant for businesses where a wrong answer costs real money — a quoted duty figure, a dosage, a delivery date. It answers only from knowledge you have loaded and classified, and refuses any figure it cannot point back to a source for.

Create an account See pricing

Three tiers. Every guardrail is in all of them — the safety layer is the product, not the upgrade.

real output from the engine — not a mock-up

Roughly what duty will I pay on a 2019 Prado?
I can't give duty, tax, or landed-cost figures — KRA publishes the rates and assesses the actual amount when the vehicle lands at Mombasa. I can walk you through the process, and a person can help you get a proper number.
Refused — operator-locked rule G-DUTY. The model was never called.

Where this is up to, before you read any further

Stated here rather than in a footnote, because a product whose pitch is “it does not assert what it cannot support” cannot have a marketing page that does.

Independent audit
not signed
Seventeen of seventeen criteria pass. An author does not sign their own audit, so until someone who did not build it does, no chat surface goes on a customer's website.
Language model
not connected
Today it restates what it found rather than discussing it. The seam is built and the checks that police a model's output are written and tested — nothing is plugged in.
Semantic retrieval
off
Matching is on keywords. A question phrased unlike the page that answers it will miss, and you will see that as “I don't have a confirmed answer” rather than as a wrong one.
Guardrail engine
built and tested
The refusals, the citation rule, the provenance model and the audit trail all work today, and are what the rest of this page describes.

Three stages, and the middle one is the product

The common design in this category is two stages: point it at your website, and it answers. What is missing between them is anybody deciding what the thing it read is allowed to be said as.

stage 1

Load

Your website is the starting point. Add PDFs, price lists, policies and your own written answers on top of it. A crawl reports what it did not keep, so silence is never mistaken for coverage.

stage 2

Classify

Every passage is typed before it can be served: your own service, your operating experience, a carrier's document, published guidance. Anything unclassified is not served — there is no route that skips this, including the API.

stage 3

Serve

The type decides how it may be phrased. A figure with no source behind it is refused. A question nothing covers goes to a queue for a person, and their answer is what the next visitor gets.

Four things it does that a prompt cannot

A rule written into a prompt is a request. A long conversation, a determined visitor or an ordinary edit can talk it past one. These are enforced in the database and in code the tenant does not control.

Rules you can read, and cannot weaken

A locked rule is enforced in the database. The business owner can see exactly what their assistant refuses and why — and cannot switch it off, by any route in the product, including the API.

Every figure traced to a source

A number the assistant cannot cite back to a specific piece of your knowledge is refused. Not matched against a list of suspicious words — refused because nothing supports it.

Facts carry how they may be spoken

A carrier's terms are attributed to that carrier. Your own experience is framed as yours. Published guidance is quoted as published. The difference is legal, not stylistic.

A record of what it nearly said

Every change to a rule keeps the previous wording, with who changed it and when, and can be put back. Answers that were generated and then refused are kept too — that is the interesting half.

What it sounds like when it has nothing

The second of these is the one worth reading. Most assistants have no behaviour for “I don't know” and produce something fluent instead.

a question the knowledge base does not cover

Do you ship motorbikes to Kisumu by rail?
I don't have a confirmed answer for that one, and I'd rather not guess. I've flagged it for the team to answer properly.
Answered — nothing covered it, so the question went to the unanswered queue.

and what the owner sees afterwards

The question appears in the console with the visitor's wording kept. The owner writes the answer once, it is classified like anything else, and the next visitor who asks gets it. Nothing is invented in the meantime, and nobody has to notice a wrong answer to find out it happened.

Why “it learns your website” is the problem, not the feature

Point a tool at a site and it will read whatever is there — your prices and a regulator's rates, your delivery promise and a page you forgot you published, all filed as one undifferentiated pile of “knowledge”. From that moment the tool will speak all of it in the same confident voice, because nothing ever recorded which was which.

There is no moment in that flow at which anybody could have said “this sentence is a claim about the law, and we may not assert it”. That is not a gap in the settings. It is a gap in the model of what knowledge is, and it cannot be closed by a better prompt.

Sona's answer is the middle stage above, and it is the reason this product is slower to set up than the alternative. That is a real cost and it is the right one.

Who is behind this

Sona is built by the people who run the businesses it is used on. Imani Freight — a diaspora shipping company moving vehicles and cargo to East Africa — is tenant zero, and the reason the product exists: a customer asking what duty they will pay is asking a question where a confident wrong number costs them thousands and costs us the relationship.

Every site it runs on today is one of ours. We are our own first customer, which is a smaller claim than a testimonial and one you can check — and it is why the refusals on this page are real output rather than a mock-up. SmartBot AI is the agency that sells and delivers it; Sona is the platform underneath.

There is no venture funding, no sales team and no support queue yet. When you write to us, one of the people who built it answers.

What it deliberately does not do

Listed because the absences are choices, and because a product that claims everything is one nobody can check.

The obvious objections

If no model is connected, what am I buying?

The part that is hard. Generating fluent text is the commodity — every provider sells it by the token. Deciding what a business is allowed to say, enforcing it where a conversation cannot reach, and keeping a record of what was refused is the part that takes a year. When a model is connected, the same checks sit in front of its output, and it is allowed to find knowledge rather than invent it.

Can I put it on my website today?

No, and that is enforced rather than promised — the embed script returns a 404 on the server. Until the independent audit is signed, the assistant is reachable from your own console and nowhere else. You can sign up, load your knowledge and see exactly how it behaves before that changes.

What happens when my prices change?

You correct the passage and the old wording is kept. Pages you have asked it to watch are re-read on a schedule, and a page whose content has changed goes back into the review queue rather than being served quietly. A source nobody has confirmed for long enough stops being served at all, and your checklist tells you it happened.

Who can see my data?

Each tenant's data is separated by their own database role, not by a filter in application code that somebody could forget to apply. The operator view used to run the platform shows counts and never content.

Is it cheaper than the alternatives?

No. It is priced in the same range and it does less, on purpose. If what you want is a bot that answers everything, the cheaper products are better at that and you should buy one.

Start with your own site

Create an account, point it at your website, and see what it keeps, what it drops and what it refuses. That is the whole evaluation, and it takes an afternoon.

Create an account See pricing

Sona is the platform. SmartBot AI is the agency that sells and delivers it. No language model is connected, the independent audit is unsigned, and neither of those sentences will be removed from this page before it stops being true.