Cost guide
What an AI chatbot really costs to build
A chatbot can be a widget over your help centre or a system with write access to your order database, and the gap between those two is most of the answer.
Get a scoped estimateWhy two quotes for the same brief look nothing alike
Ask three studios to price a chatbot and you will get three different products, because nobody agreed what the word covers. One quote is a retrieval widget over your public content. Another includes read access to your order system, a handoff that lands in your ticket queue, and a test set somebody has to sit down and write.
The question that moves the number most is whether the bot needs to do anything, or only to say things. Answering is a content problem, and content problems stay cheap when your documentation is already in order. Acting is an integration problem, and integrations are where the months go.
Cost drivers
What actually moves the number
The model licence is rarely the largest line on a chatbot build. These are the things that change the shape of the project, and most of them are decisions you control.
Answering versus acting
Answering "where is my order" needs one read from your order system. Cancelling that order needs write access, a refund policy, and a person who signs off.
The state of your documentation
Retrieval is only as good as what it retrieves. A help centre holding three versions of the returns policy becomes a cleanup project before it becomes a bot.
How many systems it touches
Each integration brings its own auth, rate limits, sandbox and failure modes. Zendesk plus HubSpot is two integrations, not one integration done twice.
Channels
A website widget is one build. WhatsApp adds Meta template approval and a business verification queue whose timing you do not control.
Where the model runs
A hosted API is the default. Self-hosting an open-weight model for data residency means GPU capacity, an on-call rota and your own evaluation harness.
Languages
One language is a prompt instruction. Five means retrieval tuned per language, plus a reviewer per language who can tell a wrong answer from a clumsy one.
Evaluation and red-teaming
Someone has to try to make the bot promise a refund you do not offer. That test set is the work that makes a launch date defensible.
Compliance and data handling
PII redaction before text leaves your network, a data-processing agreement with the model vendor, and a retention policy your legal team has actually read.
Quality of the handoff
A clean escalation writes the transcript into the ticket your agents already work in. A handoff that starts a fresh conversation saves nobody any time.
Scope tiers
Four scope tiers, smallest first
Start at the smallest tier that answers your question. Most teams learn more from shipping the first tier and reading a month of transcripts than from a larger build that lands half a year later.
- TIER ONE — ANSWER ONLY
- Answers from content you already have. Ships in 3-5 weeks.
- ₹1.5L – ₹2.5L$1.7k – $2.8k
- Includes the site widget, retrieval over your public help centre, escalation to a human by email or live chat, and a conversation log. Suits teams testing whether deflection is real. Does not include CRM access, order lookups or any action on your systems.
- TIER TWO — GROUNDED SUPPORT BOT
- Grounded answers and a real handoff. Ships in 6-9 weeks.
- ₹5L – ₹7L$5.7k – $8k
- Adds retrieval over private docs and past tickets, read-only lookups into one system of record, handoff into your ticket queue with the transcript attached, and an evaluation set. Suits support teams with real volume. It does not write to your systems or handle refunds.
- TIER THREE — ACTING AGENT
- The bot acts in your systems. Ships in 3-4 months.
- ₹14L – ₹18L$16k – $20k
- Write access behind an allowlist, spend and action caps, approval checkpoints above your threshold, two or three channels, an admin dashboard and full run tracing. Suits e-commerce and SaaS support. Does not include self-hosted models or data-residency guarantees.
- TIER FOUR — REGULATED PLATFORM
- Self-hosted models and audit trails. Ships in 5-8 months.
- From ₹35LFrom $40k
- Open-weight models running in your own environment, role-based access, per-conversation audit trails, PII redaction, multiple languages and A/B infrastructure. Suits healthcare, fintech and anyone whose data cannot leave the network. Does not include the regulatory sign-off itself; bring your own counsel.
Indicative ranges for scoping conversations, not quotes. They reflect what work of this shape has cost us to deliver — your figure comes out of a paid discovery, against a written scope, and is fixed before any code is written. Gas, licences, cloud and third-party audit fees sit outside these numbers.
How it gets scoped
From a ticket queue to a live bot
Timelines below are typical for a first chatbot. The pilot exists to discover early that the scope is wrong, which is a much cheaper place to find out than launch week.
Read the real conversations
1 weekWe take a sample of your actual tickets and chats, sort them by how often they recur and how hard they are to answer, then pick one job.
Pilot on the hard cases
2-3 weeksA working bot on your own documents, run against the queries it is most likely to get wrong. Scope usually shrinks here, and occasionally a project stops.
Build and harden
4-10 weeksIntegrations, guardrails, the escalation path, and the evaluation set we score against before the bot is allowed to speak to a customer.
Watch and tune
OngoingReading transcripts weekly at first, fixing the answers that were confidently wrong, and cutting token spend once you can see where the bot wastes it.
Is a chatbot worth budgeting for yet?
Worth reading before you take a number upstairs. For a good number of teams the honest answer this quarter is "not yet", and that is a defensible thing to report.
Worth budgeting for now if…
- The same handful of questions fills your inbox every single week
- Your answers already exist in writing and someone owns keeping them current
- You can name the person who decides an unclear case
- You have real transcripts to test against, not just a hunch
Wait, if…
- Your help centre is out of date and nobody owns it
- One person absorbs today's support volume without breaking a sweat
- The real problem is a broken checkout the bot would only explain
- Nobody on your side has time to read conversations after launch
Frequently Asked
Questions
Common questions about scoping a chatbot, cutting it back, and what it takes to run afterwards.
Integrations, almost always. Answering questions from your own content is the cheap half of the work. The moment the bot has to look up an order, update a CRM record or open a ticket, you are paying for authentication, rate limits, sandbox environments and every failure case that system can produce. Two integrations is not one integration done twice. The second common surprise is documentation: if your help centre contradicts itself, somebody has to fix that before retrieval is worth switching on.
Channels, languages and actions, in that order. A bot that lives only on your website and answers in one language still deflects the questions that repeat. Multi-agent orchestration is usually the first thing to cut and the last thing missed. What you should not cut is the evaluation set and the human handoff. A bot with no clean escalation does not save support time. It moves the same conversation to a worse starting point and irritates the customer on the way.
Data preparation and red-teaming. Retrieval quality is decided by the documents, so cleaning, deduplicating and tagging your knowledge base is real work that arrives before anything looks impressive. Red-teaming is the other one: somebody spends days trying to make the bot promise a refund you do not offer, or repeat another customer's details back to the wrong person. Teams also forget the reviewer. Someone on your side needs a standing slot to read transcripts, and that time is part of the cost.
Three buckets, and only one of them is a vendor invoice. Model usage scales with conversation volume and with how much context you push into each turn, which is why retrieval tuning repays the effort quickly. Infrastructure stays modest unless you self-host a model, where GPU capacity dominates everything else. The third bucket is human: reviewing conversations, adding answers for questions the bot has started receiving, and re-testing when a model provider changes behaviour without announcing it.
Phase it. The first version teaches you which questions actually arrive, and that list is never the one the team predicted at kickoff. Ship a narrow bot on one channel, read a month of transcripts, then spend the next phase on whatever those transcripts prove. Building the full agent up front commits you to integrations you may not need and delays the only feedback worth having. The exception is a fixed compliance deadline, where the audit trail cannot wait.
Have a project in mind?
Fixed price after a paid discovery — no hourly billing. A real engineer reads every enquiry, and we reply within 24 hours.