Cost

What an AI assistant costs on top of a bot

Retrieval adds roughly six to twelve delivery days over the same bot with scripted answers. Most of that is not model work: it is preparing your content into retrievable chunks, building an evaluation set, and designing what the assistant does when it does not know. Token spend is a running cost, not a build cost.

What an AI assistant costs on top of a bot: effort, cost and what moves it

Fixed price: $6,720 USD
Delivery: Awaiting sign-off

Effort: 16 delivery days
Biggest cost driver: Corpus preparation — unstructured source documents, not the model
Commercial basis: Fixed price agreed before work starts; scope changes quoted separately
Year one, beyond the build: Support retainer, Managed hosting, plus model api usage, content upkeep

As of 2025-10-01, Telegram Bot API 13.4

Why this costs what it costs

The pricing conversation about AI assistants is distorted by how easy a demo is. Connecting a language model to a Telegram bot and getting plausible answers takes an afternoon, and the result is genuinely impressive in a meeting. The gap between that and something a business can put in front of customers is where the entire cost lives, and it is almost never about the model.

The first real cost is your content. A retrieval assistant is bounded by what it can retrieve, and most businesses do not have a corpus — they have a website, some PDFs, a policy document somebody wrote in 2019, and a great deal of knowledge that exists only in a manager's head. Turning that into chunks that retrieve well is preparation work with a real day count, and it is work no model quality improvement removes. A well-built assistant over bad content underperforms a scripted FAQ bot over good answers.

The second is evaluation. Without a fixed set of questions with known-correct answers, there is no way to tell whether a prompt change made the assistant better or worse, and every adjustment becomes a guess defended by whoever last read a transcript. Building that set is unglamorous and is the difference between an assistant you can improve and one you can only re-roll. It is also what makes a hand-over meaningful, because it gives your team a way to check the thing after we leave.

The third is refusal. An assistant that confidently invents a refund policy creates a liability that a scripted bot cannot create, and in a consumer-facing UAE business that is a real exposure rather than a theoretical one. Designing the boundaries — what it will answer, what it escalates to a human, what it declines — and testing them adversarially is a named part of the build here. It is also the part clients most often ask to cut, and the part we most often decline to cut.

The breakdown

Work	Days	Cost	What it covers
Content audit	1	$420	What material exists, what state it is in, and what proportion of your real questions it can actually answer. Frequently ends with a recommendation to write three documents before building anything.
Corpus preparation	3	$1,260	Extraction, cleaning and chunking with boundaries that respect meaning rather than character counts. The single biggest determinant of answer quality and the least visible line in this table.
Retrieval pipeline	2	$840	Embedding, vector storage, query rewriting and reranking, tuned against your actual question distribution rather than against a benchmark.
Answer synthesis and citation	2	$840	Prompting, grounding, and returning the source alongside the answer so a customer can check it and a manager can audit it.
Refusal and escalation design	2	$840	What the assistant will not answer, how it says so, and how it hands over to a human without losing the conversation. Includes adversarial testing of the boundaries.
Evaluation set	2	$840	Fifty to a hundred real questions with known-correct answers, plus the harness to run them. The only way to know whether a change helped.
Telegram integration	2	$840	Streaming behaviour within Telegram's edit limits, conversation memory, feedback buttons on each answer, and the handover into a human topic.
Hardening and launch	2	$840	Rate limiting per user, cost caps, prompt injection defences, logging that supports review, and a kill switch that reverts to scripted answers.
Total	16	$6,720	Fixed on signature. Scope changes are quoted separately, never absorbed silently.

What moves the number

Answer quality is bounded by corpus quality, not by model choice.

If your documentation is thin, the cheapest improvement is writing it rather than upgrading the model. We will say so before quoting, because building over a bad corpus wastes your money and our reputation.

Token spend scales with conversation length far more than with conversation count.

Ten thousand short lookups can cost less than five hundred long open-ended conversations. Model your usage on message depth, and set per-user caps before launch rather than after the first surprising invoice.

Every additional language multiplies both corpus preparation and evaluation.

Arabic and English is not one assistant with a translation layer — it is two retrieval paths and two evaluation sets, because your Arabic content is rarely a faithful mirror of your English content.

Telegram limits how frequently a message may be edited, which constrains streaming.

Token-by-token streaming is not available the way it is on the web. The assistant streams in chunks, and the perceived latency design is a real decision rather than a default.

A hallucinated policy statement is a liability a scripted bot cannot create.

Refusal design and citation are not optional in a consumer-facing UAE business. This is the line item we decline to remove, and the reason is legal exposure rather than craft.

Below roughly thirty distinct questions, retrieval loses to scripted answers.

A small, stable question set is answered better, faster and more cheaply by authored text. Count your questions before buying an assistant; the comparison page has the full argument.

What year one actually costs

Running cost	Year one	Detail
Support retainer	$3,480/yr	Assistants need more attention than scripted bots: the corpus goes stale, question patterns shift, and the evaluation set needs rerunning after any provider model update.
Managed hosting	$660/yr	Runtime, vector storage and the conversation log. Modest, and the vector index for a business corpus is smaller than most people assume.
Model API usage	billed by the provider	Billed per token by the provider on your own account. Driven by conversation depth and by how much context each answer carries, both of which are tuneable and worth tuning.
Content upkeep	billed by the provider	Yours, and the most important recurring investment. An assistant over a corpus nobody updates degrades steadily while appearing to work perfectly.

Buying the same thing elsewhere

Where you buy it	What the money buys	What you carry
An off-the-shelf AI chatbot subscription	A monthly fee, a document upload box, and an assistant answering within a day. For a straightforward FAQ over clean content this is a reasonable purchase and it will work.	No control over chunking, retrieval or refusal behaviour, and no evaluation set, so quality is whatever it is and cannot be systematically improved. Escalation into your own workflow is usually shallow or absent.
A general AI consultancy	Deep model expertise, strategy work, and comfort with the parts of this that are genuinely research. Valuable when the problem is unusual.	Telegram-specific delivery — edit limits, conversation state, escalation into a forum topic, cost control per user — is usually the thinnest part of the engagement, and it is the part your customers actually experience.
Your own team with an API key	Complete control, knowledge that stays internal, and a fast first prototype. The right call when the assistant is core to your product.	The demo is the easy 20%. Evaluation, refusal design and cost control are where in-house builds stall, and they are the parts with no visible progress to show a stakeholder.

Questions about the money

Do we need AI at all, or will a scripted bot do?

Count your distinct customer questions. Below roughly thirty, authored answers win on cost, latency and trust, and they cannot invent anything. Above it, maintaining scripts becomes the bottleneck and retrieval starts paying for itself. The full threshold argument is on the comparison page.

Which model do you use?

Whichever fits the task and the budget, and the architecture keeps that swappable rather than baked in. Model choice is genuinely the least consequential decision in this build, which is the opposite of how it is usually sold.

Can it answer in Arabic?

Yes, and it is priced as a second assistant rather than as a translation setting. It needs its own corpus, its own retrieval tuning and its own evaluation set, because your Arabic material is almost never a faithful mirror of your English material.

How do we stop it from making things up?

Grounding every answer in retrieved passages, returning the source alongside the answer, and designing an explicit refusal path for anything outside the corpus. Then testing those boundaries adversarially before launch. It is not a prompt; it is a section of the build.

What does it cost to run per conversation?

It depends almost entirely on conversation depth, which is why we cap tokens per user and log spend from day one rather than discovering it. A short factual lookup and a long open-ended exchange differ by more than an order of magnitude on the same model.

Can it take actions, or only answer questions?

It can take actions — booking, lookups, ticket creation — but every action needs its own confirmation step and its own audit trail, which is priced as ordinary integration work rather than as AI work. An assistant that acts without confirmation is a support incident waiting to happen.

What happens when it does not know?

It says so plainly and offers a human, and the conversation moves into a topic your team already watches with the full history attached. A confident non-answer is worse than an honest handover, and customers forgive the second one.