Answers customers in seconds from content you have approved, then hands off to a person when unsure. Lives on your website chat and in your helpdesk.
The console beside this is one conversation, drawn. The question lands, the bot reads two of your documents, answers with both cited, and hands the hard one to a person with everything attached.
An example support conversation. A customer asks whether a sale item can be returned after 20 days; the bot reads the returns policy and the shipping FAQ, answers with both cited, and the answer is sent. A second customer reports a double charge on a gift card; the bot scores it under the confidence line and hands it to a person with the ticket, the sources it checked and the thread attached.
One example conversation, on a loop. Yours reads your documents, speaks in your tone and hands off by your rules.
Your documents, your policies, your best past answers, your product pages. That is the whole library, and every answer names the page it came from.
When a document changes, the register updates and the old answer goes with it. When a question has no source in the register, the bot says so and hands the conversation to a person instead of guessing.
no made-up answers, every one cited
Email, website chat, WhatsApp: the same cited reply, in the shape that channel expects. Inside your helpdesk it drafts the reply into the ticket for a person to send.
Re: return after 20 days
support@yourcompany · 10:42
customer wrote: Can I still return a sale item after 20 days?
Yes. Sale items can be returned within 30 days for store credit [1]; the 14-day window is for card refunds [2].
sent · 2 sources cited
Chat · yourcompany.com
online · replies in seconds
Can I still return a sale item after 20 days?
Yes. Sale items can be returned within 30 days for store credit [1]; the 14-day window is for card refunds [2].
delivered · 2 sources cited
WhatsApp Business
yourcompany · 10:42
Can I still return a sale item after 20 days?
Yes. Sale items can be returned within 30 days for store credit [1]; the 14-day window is for card refunds [2].
delivered · 2 sources cited
Also your helpdesk: the reply is drafted inside the ticket, cited, for a person to send. Hand-offs post to Slack or email with the thread attached.
The hand-off
Every answer is scored before it is sent. Above your line it goes out, cited. Below it, the conversation lands with your team: the ticket, the thread so far and the sources it checked, all in one place.
Some topics never take the test. Payments, complaints and anything legal go to a person on sight, whatever the score. That is the human-in-the-loop promise, and it is a rule you set, not a setting we hide.
Can I return a sale item after 20 days?
92%sent, cited
Do you ship to the islands?
81%sent, cited
Charged twice, paid with a gift card
41%to a person
Payments · complaints · anything legal
no scorealways a person
attached
Illustration of a real gate. The scores are examples.
Setup covers the reading, the rules and two weeks of tuning. The monthly fee keeps it connected, monitored and corrected. Every number is in writing before we start.
We load your help centre, policies, past tickets and product pages into the register. You approve the set.
How it speaks, what it never says, and the exact words it uses when it does not know.
The topics that always go to a person, where your confidence line sits, and who picks up.
After it goes live we read every answer with you, correct what needs correcting, and re-index.
You watch it answer your real questions before any customer does.
Model usage is passed through at cost, never marked up. Month to month, cancel with 30 days notice.
A support bot is a reader, not a writer of policy. It holds an index of the documents you approved, finds the passages that answer a question, and writes a reply from those passages in your tone. Every reply carries a citation to the passage it came from, so your team can check any answer in seconds and a customer can see it is not a guess.
It lives where your customers already write to you: the website chat, the inbox, WhatsApp, and inside your helpdesk, where it drafts the reply into the ticket for a person to send.
It is not a general chatbot with your logo on it. It does not answer from the open internet, from other customers, or from what the model remembers from training. If the answer is not in your documents it says so and hands the conversation to a person, which is slower than a guess and worth it every time.
It does not replace your support team. It takes the repeats off their plate so the conversations that need a person get one sooner.
Each document is split into passages when it is indexed. When a question comes in, the closest passages are retrieved and scored; the reply is written only from those, and each sentence is tied back to the passage it came from. The small marks in an answer, [1] and [2], are those ties.
A reply that cannot be tied to a passage is not sent. That single rule is what makes the bot safe to put in front of customers.
When someone on your team edits a reply, the correction is saved against the source passage, and the next answer to that question uses it. It never learns from customers, only from you.
When a document changes, the old passages are dropped and the old answers go with them. The register on this page is that mechanism: what is in it can be cited, what is not cannot be said.
Low confidence, a sensitive topic (payments, complaints, anything legal) or a plain request for a human all route the conversation to your team, with the question, the thread so far and the sources it checked attached. The customer sees one honest line: a person is picking this up.
The line is yours to set, per topic, and we tune it with you for two weeks after launch by reading the answers together.
We publish a number only once we have measured it on a live build, and the support bot's are yours to see, not ours to claim. What we can show is the report every client gets, and the fleet the same hand-off design already runs for.
Measured on every live bot from the first week. Yours, not a claim of ours.
live AI reception fleet across Greece, now expanding into English-speaking markets
counted from production deployments, 2021 to today
The same answers we give on the first call.
That is what we design against. The bot is grounded: it answers only from documents you have approved, not general knowledge, so it cannot invent policy. When a question falls outside what it can answer confidently, it hands off to a human with full context instead of guessing.
It is agreed before any work starts. Chatbots start from €2,000 to set up; the total depends on the integrations yours needs, like your helpdesk, CRM, or customer database. On top comes a monthly running fee from €200 a month, scoped the same way, with no per-conversation surprises. You get the exact numbers on a 15-minute call, and larger custom builds are quoted after a scoping call.
No. We build it, load your content, and test it against your real questions before it touches a customer. After launch the monthly running fee covers monitoring and tuning it as your documents change.
Yes. Code2b is built to SOC 2 controls and GDPR aligned, DPA on request, with encryption at rest and in transit, role-based access, and audit logging. When data cannot leave your servers, it runs inside your own infrastructure, or as a model of your own built by yourtrained, on your hardware or a single-tenant EU GPU pod.
It answers in natural language matched to your tone, not a rigid decision tree. And it knows its limits: it escalates to a person the moment the question is sensitive, the confidence is low, or someone asks to talk to a human.
No. It removes the repetitive work so your team spends time on the conversations that actually need a person. Every hard or sensitive case still routes to a human, with the context already gathered.
Most builds go live in 2 days to 2 weeks, a simpler single-channel bot in about one. You see it answering your real questions before it ever speaks to a customer, with a fixed scope and fee agreed up front. The first two weeks after launch are tuning: we watch every answer with you and correct what needs correcting.
Yes, in a controlled way. When someone on your team edits a reply, the correction is saved against the source it came from, and the next answer to that question uses it. It never learns from customers, only from you. When a document changes, the old passages are dropped and the old answers go with them.
Book a 15-minute call. We map the questions your customers ask, the documents that answer them, and where the bot lives, then give you a fixed scope, fee, and timeline before you commit.
Every support bot is led personally by Code2b's founders, Alex and Dusan.