Techbanq

Cost guide · AI

The model is the cheapest part of a chatbot that works

A working assistant is not a model with a prompt; it is the knowledge it can accurately draw on, the systems it is allowed to act in, the guardrails that stop it inventing answers, and the evaluation that keeps it right as your business and its language change. The per-message price everyone asks about is real but small. What actually sets the cost, and the risk, is everything the model connects to, which is what this page is about.

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

Market
Saudi Arabia
Written for
Head of CX, COO
Decision
Understand what drives the cost of an AI chatbot
Languages
EN · AR

Enterprise delivery credentials

  • ISO

    ISO/IEC 27001

    ISO/IEC 27001 certified

Certifications & registrations

Outside the model

What sets a chatbot's cost once the model is chosen

Each driver is listed with the direction it pushes the figure. The model call is a small, predictable line; these are where a chatbot's real cost and risk live.

VariableEffect on priceWhy
The knowledge it answers fromThe largest driver of a useful botA chatbot that answers accurately does so from your content, not the model's general memory, which means gathering, cleaning and structuring the material it retrieves from and keeping it current. A tidy, well-maintained knowledge source is modest; scattered, contradictory or out-of-date content is where most of the effort goes, because a bot grounded in bad material is confidently wrong.
Whether it answers or also actsActing raises cost and risk sharplyA bot that answers questions is one thing; a bot that books an appointment, checks an order or changes an account is doing work in your systems, and each such action is an integration with its own permissions, testing and failure handling. The step from answering to acting is where most of the cost appears, and it is the step a per-message price ignores entirely.
Guardrails and escalationNot optional for anything customer-facingA chatbot in front of customers needs limits on what it will attempt, a way to recognise when it is out of its depth, and a clean handover to a person. Building the boundary and the escalation path is real work, and skipping it does not save money so much as move the cost to the day the bot gives a confident wrong answer to a customer.
Arabic, dialect and toneA build requirement in this marketServing customers here means handling Arabic well, including the way people actually write it, and holding a tone that fits your brand. Getting this right is testing and tuning against real messages, not a language setting, and a bot evaluated only in English will read as foreign to the customers it is meant to serve.
Ongoing evaluationA recurring cost, not a one-offA chatbot is accurate on the questions it was tested on and drifts as your products, policies and customers change. Keeping it right means measuring its answers against reality on a schedule, catching where it has started to fail, and correcting it. This loop is the difference between an assistant that stays useful and one that quietly degrades, and it is a running line a launch price leaves out.

Ask any supplier quoting this

  • Where does the bot get its answers, and who keeps that knowledge current?
  • Does it only answer, or does it act in our systems — and which actions are in scope?
  • What stops it inventing an answer, and how does it hand over to a person?
  • How is Arabic accuracy tested, and against what real messages?
  • What is the ongoing cost of evaluating and correcting it after launch?
  • Does the figure separate the model usage from the build and the running cost?

The per-message illusion

A price per message answers the one question that barely matters

Because the model is billed per token, it is easy to reduce a chatbot to a cost per message and compare suppliers on that. It is the wrong number to fixate on. The model call is cheap, predictable and roughly the same wherever you buy it; what varies enormously, and what decides whether the bot is worth having, is everything around it — the quality of the knowledge it draws on, whether it can act safely in your systems, and the work of keeping it accurate. A supplier who competes on the per-message rate is competing on the part of the cost that was never the problem.

The honest way to budget for an assistant is to size the parts a per-message price hides: the effort to get the knowledge into shape and keep it there, the integrations if the bot is to act rather than only answer, the guardrails and escalation, and the ongoing evaluation. Some of these are one-off and some recur, and together they dwarf the model bill. A chatbot scoped this way costs more on paper than one sold as a rate per message, and it is the one that still gives correct answers a year after launch.

What earns trust

The conditions a chatbot has to meet before customers can rely on it

Each is a requirement that carries cost, and each is a place a cheap build cuts. A blank in any row is where a confident wrong answer will eventually come from.

Answers grounded in your content

A maintained knowledge source the bot retrieves from, rather than reliance on the model's memory.

What a gap here produces: Plausible answers that are wrong, because the model filled a gap with invention.

A known boundary on what it does

Explicit limits on the actions it may take and the topics it will attempt.

What a gap here produces: A bot that acts beyond its remit or answers where it should refer.

A clean handover to a person

A path that recognises when the bot is out of its depth and passes the customer on with context.

What a gap here produces: A customer stuck in a loop with a bot that cannot help and will not let go.

Verified Arabic handling

Testing and tuning against real Arabic messages, including how people actually write.

What a gap here produces: A bot that reads as foreign to the customers it exists to serve.

A running evaluation loop

Scheduled measurement of answers against reality, with correction when it drifts.

What a gap here produces: Quiet decay, as the bot keeps answering old questions about a business that has moved on.

Our role

What we size before we put a number on an assistant

  1. 01

    The knowledge, assessed first

    We look at the content the bot would answer from and how maintained it is before quoting, because grounding the assistant in your material is the largest part of the work and the part that decides whether it is accurate.

  2. 02

    Answering and acting, priced apart

    We separate a bot that answers from one that acts in your systems and price the integrations and guardrails explicitly, so the step that carries the risk is a decision you make rather than a surprise you inherit.

  3. 03

    Evaluation as a standing line

    We quote the ongoing evaluation as its own recurring cost rather than folding it into the build, because an assistant is kept accurate, not finished, and a price that hides that has priced a bot that will drift.

How much you want the assistant to do, and how accurate it must stay, are yours to decide; we size the knowledge, the integrations and the evaluation so the figure covers what keeps the bot right rather than only the messages it sends.

Common questions

What buyers ask

Suppliers quote us a price per conversation — why won't you?+

Because a price per conversation covers the cheapest and most predictable part of a chatbot and hides the parts that decide whether it works. The model usage behind a conversation is small and roughly the same wherever you buy it. What varies, and what we would be quoting on if we quoted honestly, is getting your knowledge into shape, building the integrations if the bot is to act, the guardrails and escalation, and the ongoing evaluation. A per-conversation rate lets a supplier look cheap while quietly assuming a thin bot that answers from the model's memory, which is exactly the bot that gives confident wrong answers. We price the parts that matter and show the model usage as the small line it is.

Why does keeping a chatbot accurate cost money after it launches?+

Because your business does not hold still and the bot was only ever accurate on the world it was tested against. Products change, policies change, prices change, and customers ask new things in new ways. Without a loop that measures its answers against current reality and corrects them, a chatbot keeps confidently answering last quarter's questions, and the failure is invisible until a customer is misled. The evaluation is not a defect being fixed; it is the ordinary cost of keeping a system that talks to customers honest as the thing it talks about changes. A bot sold as finished at launch has simply moved that cost out of the quote and onto your customers.

Is a chatbot that just answers questions much cheaper than one that takes actions?+

Yes, and it is the single biggest fork in the cost. A bot that answers questions from your knowledge is largely about getting that knowledge right and putting guardrails around it. A bot that also books an appointment, checks an order status or changes an account is doing work in your systems, and every one of those actions is an integration with its own permissions, testing, error handling and audit. It is a genuine step up in both cost and risk. The useful move is to be deliberate about which actions are worth that step: many assistants deliver most of their value by answering well and handing off cleanly, and adding actions one at a time, where each earns its integration, keeps the cost tied to the value.

Scope an assistant that works

A scoping session that sizes the knowledge base, the integrations and the evaluation loop, so a chatbot is budgeted for what keeps it accurate rather than for the model calls alone.

Start the conversation