Techbanq

AI · conversational

A chatbot is judged by the questions it refuses to guess at

A chatbot worth deploying answers from your own product pages, policies and help content, not from a model's general memory, and it recognises the question it should not attempt — a complaint, an edge case, a request that needs a person — and hands it over with the conversation intact. Getting it to talk is a weekend; getting it to know its limits is the engagement.

Serving Saudi Arabia from our Riyadh office and our engineering hubs in Lahore and Karachi

Shoppers moving through a modern retail concourse
  1. 01

    Market

    Saudi Arabia

  2. 02

    Written for

    Head of CX, Head of Digital

  3. 03

    Decision

    Deploy a website or app chatbot grounded in your own content

  4. 04

    Languages

    EN · AR

Information security for software delivery

  • ISO

    ISO/IEC 27001

    ISO/IEC 27001 certified

Certifications & registrations

The path of a question

What happens between a question and an answer that can be trusted

A deployable chatbot does not answer straight from the model. It reads the question, finds the passage in your content that addresses it, answers from that passage, and escalates when it cannot. The escalation branch is the part that makes the rest safe.

A question flows through four steps — read the intent, find the passage in your content, answer from the source, know when to stop — with a branch handing the conversation to a person with the full thread.

  1. 01

    Read what was asked

    The chatbot works out the intent behind the wording — including a question typed half in Arabic and half in English.

  2. 02

    Find it in your content

    It retrieves the passage from your pages and policies that actually answers the question, rather than composing from general knowledge.

  3. 03

    Answer from the source

    It replies in the customer's language, grounded in that passage, so the answer can be traced to where it came from.

  4. 04

    Know when to stop

    When no passage answers the question, or the request needs judgement, it says so plainly instead of inventing a reply.

Always available

Hand it to a person

When the chatbot reaches its limit, it passes the conversation to an agent with the full thread attached, so the customer never re-explains and the agent starts informed. A chatbot without this branch does its worst work exactly when the question mattered most.

The demo that misleads

The confident wrong answer is worse than no answer

A chatbot wired straight to a language model demos beautifully, because it answers everything fluently. That fluency is the trap. A model asked a question outside your content will not fall silent; it will produce a confident, plausible reply that may be wrong — a refund window that does not exist, a policy you never wrote. On a support channel that answer is not a curiosity; it is a commitment a customer will hold you to, and cleaning it up costs more than the chatbot saved.

The fix is to ground the chatbot in your own content and to make refusal a first-class behaviour, not a failure state. It answers when a passage supports the answer and says it does not know when none does — and that honesty is what earns it the right to handle the easy majority so your team keeps its attention for the questions that need a person. A chatbot that filters and routes is worth deploying; one that guesses to seem helpful is a liability with a friendly avatar.

The deployable difference

Four things that separate a real chatbot from a demo

Any of these missing is why a chatbot that impressed in a meeting becomes a complaint in production.

01

It cites where the answer came from

An answer traceable to a passage in your content can be checked and corrected; one generated from general memory cannot be verified until a customer disputes it.

02

It refuses rather than invents

When the content does not cover the question, saying so is the correct behaviour. A chatbot tuned to always answer is tuned to sometimes lie.

03

It hands over without a reset

The escalation carries the whole conversation to the agent, so the customer does not start again from "how can I help you" after already explaining.

04

It works in both languages, mid-thread

A Gulf customer may open in English and switch to Arabic in the next message. The chatbot follows the switch instead of losing the thread.

Our role

What we build and what your team keeps deciding

  1. 01

    Grounding, not just wiring

    We connect the chatbot to your content so its answers come from your pages and policies, and we tune it to refuse rather than guess when the content is silent.

  2. 02

    The handover, designed first

    We build the escalation path into your existing support tools before launch, because a chatbot with no clean way to reach a person is one you cannot trust with a real customer.

  3. 03

    A record you can improve from

    We instrument the conversations the chatbot could not answer, so your team can see the gaps and close them by adding content rather than by guessing.

Which questions the chatbot should never attempt, and where an escalation lands, are your service decisions; we build a chatbot that keeps to the first and delivers the second with the context a person needs to help.

Common questions

What buyers ask

How is a website chatbot different from a WhatsApp or voice agent?+

The reasoning is shared; the channel changes everything around it. A website or in-app chatbot has a screen, so it can show a link, a form or a table, and the customer reads at their own pace. A voice agent has no screen and must handle interruption, silence and being misheard, and a WhatsApp agent lives in an asynchronous thread where a reply may come hours later. We build the grounding and the handover logic once and shape the surface to the channel, rather than pretending a voice call and a chat window are the same product. This page is about the on-screen chatbot; the voice and messaging channels are their own work.

What stops the chatbot from inventing an answer when it does not know?+

Two things working together. First, the chatbot answers from retrieved passages of your own content rather than from the model's open-ended memory, so an answer has to be supported by something you actually published. Second, it is built to treat "the content does not cover this" as a valid, expected outcome that triggers a plain admission and a handover, not as a failure it should paper over. We cannot promise it is never wrong — that is one of the two claims this page will not make — but grounding plus a designed refusal makes a confident fabrication the exception rather than the default a raw model produces.

Does it need to speak Arabic and English in the same conversation?+

For a Gulf audience, yes, and treating it as an afterthought is a common way these projects disappoint. Customers here routinely open in one language, switch to the other mid-conversation, and mix both in a single sentence. A chatbot that resets or degrades when the language changes reads as foreign to the people it is meant to serve. We build and test the switch as a normal case from the start — the chatbot follows the customer's language rather than forcing a choice at the beginning — because a bilingual audience is not two monolingual audiences taking turns.

Next step

Scope it with us

A build plan that names the content the chatbot answers from, the questions it must escalate, and how the handover reaches a person.