The answer is already written when you open the inbox. You read it, edit it, press send.
ai backends
the reply is already written.
An AI email assistant for business, that drafts and never sends.
Email drafts, a chat agent, a dashboard. Each reads your own material and shows its sources. A person still presses send.
it works on what happens from now.
An AI backend is the lighter cousin of the business brain. The difference is the archive: a backend works on what arrives from now, the brain on what already happened.
"Answer this enquiry" is a backend. "What did we quote them in 2019, and who agreed it" is not. the business brain →
where the line falls
A backend sees what turns up after it is switched on: the enquiry that arrives this morning, the booking made after closing, the chat window on your site, this month's numbers.
The business brain is built on what already happened. The one live deployment reads twelve years of mail alongside the business's document archive, so it can answer questions about earlier prices, promises and deadlines. No backend can reach that history.
Most businesses start here. The archive comes later, once it is obvious which questions keep going unanswered.
three quiet robots, none of which send.
Built into the site, not embedded from somewhere else.
Answers out of your own pages, links the one it used, and says when it does not know.
Enquiries, bookings, what people read before they wrote to you.
email that answers itself.
Most of a small business's inbox is the same questions arriving again. Opening hours. Do you deliver here. What does this cost. Those are worth handing over.
the rule the whole thing is built on
it drafts. you send.
The assistant writes the reply out of your own material — your price list, your pages, the answers you have already given — and leaves it in your drafts, unsent. Nothing reaches a customer that a person has not read. There is no sending in the build at all.
your inbox
drafting
draftsnever sends
your keyyour billing page
citedfrom your own files
draft ready · sitting in drafts, unsent
"Hello — yes, we deliver to Utrecht. For three units the delivery charge is…"
- It writes from your material, not from the internet. Your price list, your pages, your documents, your past replies.
- Prices come from a table you approve. If the enquiry asks about something not on the table, the draft says so.
what it refuses to draft, and how it shows its work
A complaint, a legal question, something it has no material for — it does not draft those, it flags them. The point is a shorter inbox, not a fuller one.
Every draft shows its sources: which page, which document, which past reply. If one looks wrong you can see in a click where it got that.
It is not asked to be knowledgeable about your trade. It is asked to find what you already wrote down. Corrections can be added to the source material or a rule file you control, so the next draft starts from the corrected record rather than an invisible learning claim.
a chat agent that cites the page.
Two things most chat widgets skip: it links what it read, and it admits when your site has no answer.
- It reads your site, not the whole web. Your pages, your PDFs, your price list — nothing else is in scope, which is why the answers stay yours.
- Every answer carries a link to its source. The page it came from, in the answer, clickable.
- "I don't know" is a real answer. When your material does not cover the question it says so and offers the contact form. That refusal is the feature.
handover, and why there is no cookie banner
A conversation it cannot finish becomes an enquiry in your inbox with the transcript attached — and, if you have the email side too, a draft reply waiting next to it.
It is built into the page, not a third-party widget loaded from someone else's server with someone else's cookies attached. That is what keeps the site free of tracking cookies and consent banners.
your own numbers, in your own account.
First-party data, collected by your own site and kept in your own account — not rented back to you by an analytics product.
- The numbers your business runs on. How many enquiries this month, what they were about, how many turned into work. Not a wall of charts nobody opens.
- Yours to keep, export, or switch off. None of it depends on us still being here.
how it counts, and where it leads
Aggregate, cookieless, one-click opt-out. Which pages get read, where people arrived from, what got clicked, which screen widths. Never a profile of a person, and nothing handed to an ad network. None of it lives in an account you would lose access to.
It is also the on-ramp. A business that owns its own record has the first dataset an agentic system can read. That is where the business brain starts. The rest of what a build includes is on custom-coded websites.
it runs on an account you control.
A site-attached assistant uses a metered API account in your name. A consumer chat subscription is separate and is never assumed to cover this work.
Claudeavailable through your provider account
OpenAIavailable through your provider account
Geminiavailable through your provider account
Your files, your account
The price list, pages, documents and rules sit with you. Provider credentials and metered usage stay visible in the account you own.
Vendor marks are shown to say what the system connects to. They are not a partnership and not an endorsement.
on your account, where you can see the meter.
Not a seat you rent from us. The assistant calls the AI on your own account, so every call and its cost is a line you can look at.
- Your account, in your name. If you want to stop it for a week, you stop it — you do not ask us to.
- It meters, and the bill moves with use. There is no honest generic monthly number. The written scope includes an estimate based on the model, context and volume your build needs.
- We bill the build, not a seat. Website development starts at €500, excluding VAT, and a backend is quoted in the same written scope.
how the running cost actually works
Send ten times as much through the assistant and it costs more. That is arithmetic, and it happens on your billing page where you can watch it. We would rather you saw a small bill than were told there wasn't one. There is no per-user monthly fee to us for using something you own, and any software engineer can pick the build up afterwards.
An eligible static site can still cost €0 a month to host, but AI provider usage is separate and metered. The site's other exclusions remain your own domain, transactional email past free tiers, and unbranded multi-user booking — explained on hosting & care.
A backend attaches to a site we built, and to one you already have if the material it needs to read is reachable. That is the first thing we check.
what it sits next to.
A backend attaches to a site, and it is the smaller half of the AI work.
Hand-built, no templates. From €500, €0 a month to keep online — outside that, your own domain (~€10–15 a year, registered to you), transactional email past the free tiers, and unbranded multi-user booking.
custom-coded websites →Years of mail and documents made askable, with a source on every answer. Quoted after we understand the archive.
the business brain →The €0 a month side, with its three exclusions priced.
hosting & care →questions we get asked.
The five that come up on nearly every first call.
Will it email my customers on its own?
No. It writes a draft and leaves it where you will see it, and a person reads it and presses send. Sending is not a switch that is turned off — the build has no route to send mail by itself, which is why nobody can turn it on by accident.
What happens when it does not know the answer?
It says so. The assistant answers from your own pages and documents and shows which one an answer came from. When nothing in there covers the question, it says it does not know and hands the conversation over rather than inventing something that sounds right.
What does it cost to run?
It runs through a metered API account you control, so each call and its cost is visible there. The number moves with the model, context and volume. The written scope includes an estimate. KUUP bills the build, not a recurring seat.
Which AI does it run on?
Claude, OpenAI or Gemini, through a provider API account in your name. A consumer chat plan is separate and is not used as third-party automation credentials. Showing the marks says which providers a build can connect to. It is not a partnership or endorsement.
How is this different from the business brain?
The archive. A backend works on what arrives from now — the enquiry that came in this morning, the chat window on your site, this month's numbers. The business brain is built on what already happened: the one live deployment reads twelve years of mail and the documents kept beside it. If the question is about 2019, no backend can reach it.
tell us what the inbox keeps asking.
Tell us which questions you answer again every week. If a backend would not save you anything, we will say so.
start a project →