Chatbot Development Cost: Every Line Item, Priced

Chatbot development cost splits into two numbers, and the smaller one is the run rate: a conversation costs about $0.004 on a raw model API. Anthropic publishes the arithmetic on its own pricing page: an average support conversation runs roughly 3,700 tokens, or about $37 per 10,000 tickets on Claude Haiku 4.5. The same conversation, once resolved, costs $0.99 on Intercom's Fin. Two published numbers, and a gap of about 270 times between them.

Most of what people call chatbot development cost lives inside that gap. It is not quite a like-for-like comparison, because Fin charges only when it resolves the conversation while the token figure covers every conversation, resolved or not. Assume a 50% resolution rate and the gap narrows to roughly 130 times, which is still enormous. What the extra buys is the part nobody prices in a spreadsheet: the retrieval, the handoff, the content pipeline, the evaluation, and an engineering team you do not have to hire. Building your own means paying for those yourself, up front, in engineer-weeks.

This guide prices both routes with real published figures. The chatbot is the worked example because it is the AI project most companies buy first, and the one where the numbers are actually knowable. Everything below applies to the wider question of AI development cost too, but a chatbot is where the abstractions turn into invoices.

The short version

• Model tokens are the cheapest line. About $0.004 per conversation, or $37 per 10,000, on a small production model.
• Managed platforms invert the economics: near-zero build, then $0.99 per resolution on Intercom or $2 per conversation on Salesforce Agentforce.
• The build, not the model, is what you are really buying when you go custom. Twelve engineer-weeks is around $32,000 in base salary at US median rates.
• The costs that break budgets are content preparation, integration, evaluation, and the retuning that never stops.
• Cost scales with tokens, not users. Context length and turn count drive the bill far more than headcount does.
• Gartner expects generative AI cost per resolution to exceed $3 by 2030, higher than many offshore human agents, so the cheap-forever assumption is worth stress-testing now.

What a chatbot costs to build and run

Three routes exist, and they price completely differently. A managed platform charges almost nothing to start and bills per outcome forever. A framework build on your own stack costs engineer-weeks up front and cents per conversation after. A custom or fine-tuned model adds a data and training program on top of the framework build, and almost nobody needs it for a chatbot.

The published anchors behind that table: Intercom lists Fin at "From $0.99 per Fin outcome" with seats from $29 a month, and notes that on an existing helpdesk you pay the per-outcome fee with no seat cost at all. Salesforce prices $2 per Agentforce conversation, or $500 per 100,000 Flex Credits if you meter by action instead. Microsoft sells Copilot Studio at a $200 monthly capacity license or pay-as-you-go credits. Amazon prices Lex at $0.00075 per text request, which is the cheapest managed option on the list and also the least capable out of the box.

For the framework route the honest anchor is labour. Stack Overflow's 2025 Developer Survey puts the median US full-stack salary at $138,000 a year, which is about $2,650 a week per engineer in base salary alone, before employer taxes, benefits, equipment and management. Twelve engineer-weeks, a realistic figure for two people over six weeks building a retrieval chatbot wired into one system, is therefore roughly $32,000 of salary. An agency quote will be a multiple of that, because it carries overhead and margin. The useful discipline is to convert any quote back into engineer-weeks and ask whether the weeks are credible for the scope described.

Chatbot development cost by route: platform, framework, or custom

Start on a platform unless you have a specific reason not to. That is the honest default, and the reason is arithmetic: at 1,000 resolved conversations a month, Fin costs about $990, and a year of that is under $12,000, which is a third of what twelve engineer-weeks of salary costs before you have shipped anything. Below a few thousand conversations a month, buying wins on cost alone, and it wins on time to value by weeks.

The platform stops winning at three thresholds. Volume is the obvious one: per-resolution pricing scales linearly, so ten thousand resolutions a month is roughly $9,900 and a hundred thousand is roughly $99,000, at which point the build looks cheap. Data is the second: if the chatbot has to reason over documents, records or business rules that live inside your systems, you end up building integrations anyway, and the platform becomes an expensive shell around your own work. Interface is the third: if the chatbot is part of your product rather than a widget bolted onto support, the vendor's UI constraints will fight you.

The framework route means an LLM API, a retrieval layer, your own application code, and an evaluation harness. It is the route most companies land on once the chatbot stops being a support add-on and starts being a feature. Cost is dominated by engineering time, and after launch it drops sharply, because tokens are genuinely cheap. The mechanics of wiring that call into an existing request path, including key management, retries and provider fallback, are covered in our guide to AI API integration in production.

Custom or fine-tuned models are the rare route for a chatbot. Fine-tuning helps with tone, format and narrow classification, and it does not help with knowing your facts, which is what retrieval is for. Unless the model itself is the product you sell, the data preparation and training cycles buy you very little that a good system prompt and a decent retrieval layer would not.

The most expensive chatbot decision is building one before you have proved anyone will use it. Renting the capability first costs a fraction, and it tells you whether the custom build is worth having.

What a chatbot costs to run per conversation

Per-conversation cost is the number that decides whether the project survives its second year, and it is knowable before you build. Anthropic's own worked example gives the shape: about 3,700 tokens for an average support conversation, priced on Claude Haiku 4.5 at $1 per million input tokens and $5 per million output, coming to roughly $37 per 10,000 tickets. That is $0.0037 a conversation, or put another way, a million conversations a year costs about $3,700 in tokens.

Model choice moves that number by two orders of magnitude. Here is the current published spread for the small and large ends of the three major providers, per million tokens:

Sources: OpenAI, Anthropic and Google publish these openly, and they change often enough that any figure in a vendor proposal should be checked against the source rather than taken on trust. Note the Sonnet 5 row in particular: Anthropic lists $2 and $10 as introductory pricing through 31 August 2026, after which the standard $3 and $15 apply. A quote built on introductory pricing is a quote with a scheduled 50% increase in it.

The practical read is that the same chatbot costs 100 times more on a flagship model than on a nano-class one, and for a support bot answering questions from retrieved documents, the nano-class model is often good enough. The way to find out is to measure, which is what the evaluation harness is for, and which is why skipping it costs money rather than saving it.

Why token volume, not user count, drives the running bill

Chatbot cost scales with tokens, and tokens scale with how much context each turn carries rather than how many people are talking. This catches teams out constantly, because the mental model everyone brings is per-seat software.

Every turn resends the system prompt, the retrieved documents and the conversation history so far. A ten-turn conversation therefore pays for the system prompt ten times and for an ever-growing history on each turn. Double the length of a conversation and you more than double its cost. Retrieve six documents instead of two and every subsequent turn carries them. A chatbot with a 4,000-token system prompt and aggressive retrieval can cost twenty times what a lean one costs while answering exactly the same questions.

Caching is the direct counter, and both major providers price it explicitly. Anthropic charges cache reads at 0.1 times the base input rate, with a 5-minute cache write at 1.25 times, so caching pays for itself after a single read. OpenAI does the same in effect: gpt-5.4-mini is $0.75 per million input tokens and $0.075 cached, a tenfold reduction on the static part of every prompt. For a chatbot, where the system prompt and product documentation are identical across thousands of conversations, this is the cheapest large saving on the list, and it is a few lines of code.

Retrieval adds its own line. Embedding the content is close to free, at $0.02 per million tokens for text-embedding-3-small. Storing and querying the vectors is where the floor sits: Pinecone's Standard plan carries a $50 monthly minimum, with reads at $16 to $18 per million read units and storage at $0.33 per GB per month. For a small chatbot, that $50 minimum is larger than the entire token bill.

The chatbot costs nobody puts in the quote

Five line items go missing from most proposals, and together they usually outweigh everything that was quoted.

Content preparation is the hidden half of chatbot development cost

A chatbot answers from what you give it, and most companies' documentation is not in a state to be given to anything. Gartner's October 2025 survey of 321 customer service leaders found 58% planning to upskill agents into knowledge management specialists, on the explicit grounds that AI systems need accurate, continually updated content. That is an organization-wide admission that the content work is the work.

Integration and the human handoff

A chatbot that only reads a help center is cheap. One that checks an order status, looks up an account, or opens a ticket has to authenticate, respect your permissions model, and fail safely when a system is down. The handoff to a human is its own build: transcript, context, routing, and a queue that does not drop people. Getting this wrong is expensive in a way that does not show up on the invoice.

Evaluation, which is what lets you cut cost later

You cannot ship what you cannot measure, and chatbots fail in ways ordinary software does not, being confidently wrong rather than visibly broken. A test set of real questions with known good answers, run on every prompt or model change, is what lets you swap to a cheaper model later without guessing whether you broke something. Skipping it means never being able to cut your own costs safely.

Channel fees on SMS and WhatsApp

The web widget is free. Everything else is metered. Twilio charges $0.0083 per SMS in the US plus carrier fees that run from $0.0025 to $0.007 per message depending on the network. WhatsApp is more forgiving for inbound support: Meta charges per message by category, and service messages are free inside the 24-hour customer service window that opens whenever a user writes to you and resets with each of their messages. Marketing messages are always charged. A support chatbot on WhatsApp can therefore run its conversations at no channel cost, while the same infrastructure used for outbound campaigns bills every send.

The tuning that never stops

Prices move, policies change, models get deprecated, and answers drift. Gartner's January 2026 prediction that cost per resolution for generative AI will exceed $3 by 2030 attributes the rise to data center costs, vendors moving from subsidized growth to profitability, and use cases that consume more tokens as they get more capable. The same release notes that by 2028, regulation guaranteeing access to a human is expected to push assisted service volume up by 30%, which means the human capacity you were planning to remove has to stay.

How chatbot development cost changes with scope

Scope moves chatbot development cost more than any technical choice, and it moves it in steps rather than smoothly. Each rung on the ladder below adds a category of work rather than a percentage.

The step from retrieval to account-aware is where most budgets break, because it is the first rung where the chatbot touches systems that were built for humans with sessions rather than software with tokens. The step from account-aware to transactional changes the risk profile entirely, and at that point you are no longer costing a chatbot.

Gartner's July 2026 customer survey adds a useful pressure to this ladder: among customers who use generative AI, 58% have used it to complete a task on their behalf, rising to 74% in B2B. Expectations have moved past answering. Scoping a question-answering chatbot when your customers expect to get something done is a cheap build that fails at its actual job.

Chatbot or AI agent: two cost models, not one

A chatbot is priced by conversation and an agent by action, and that single difference is why the two budgets behave nothing alike. A chatbot produces text inside a conversation: its cost tracks turn count and context size, so the levers are model choice, caching and prompt discipline. An agent plans its own steps, calls tools, reads results and adapts, so its cost tracks how many steps it takes, which varies per request and is much harder to forecast.

The vendors price exactly this distinction. Salesforce charges $2 for a conversation but $0.10 per Agentforce action, which means an agent that takes twenty steps to resolve one request costs the same as a whole conversation, and a badly scoped one costs more. Intercom's per-outcome pricing hides the same variance behind a flat fee, which is a real argument for buying rather than building at the agent end of the spectrum.

If you are pricing autonomous systems that take actions in your business rather than conversational interfaces that answer questions, the economics, the guardrails and the evaluation requirements are all different. Our guide to AI agents for business covers that side, including build versus buy and what a deployment actually takes, and the AI agent development page covers how we build them. Everything in this article stays on the chatbot side of that line.

Where chatbot cost sits inside AI development cost

Chatbots are the cheapest well-understood entry point into AI development cost, which is why so many companies start there. The same five drivers govern every AI build: the approach you pick, how ready your data is, how many systems you integrate with, the evaluation and guardrails you put around the output, and the inference you pay for on every request afterwards. What makes a chatbot tractable is that all five are small and measurable at the start.

The wider lesson transfers. AI development cost has no list price because the approach swings it by orders of magnitude, and because unlike most software, the meter keeps running after launch. A chatbot makes that visible early and cheaply: you can see your token bill within a week of shipping, adjust the model, and know your unit economics before committing to anything larger. That is a much better way to learn what AI costs than commissioning a platform and finding out in month nine.

The same honest-by-drivers logic applies to software generally, which we lay out in the custom software development cost guide. For where chatbots and AI sit within a support operation specifically, including deflection rates and where the returns actually come from, see our piece on AI customer support. And for a worked vertical example of what a chatbot needs before it talks to real buyers, our write-up on real estate chatbots covers the live-data and handoff requirements in one industry.

How to spot an underestimated chatbot quote

Cheap quotes are usually not cheaper. They are quotes with things left out, and the missing items arrive later at a worse time. Six questions separate a real number from an optimistic one.

1. Who prepares the content, and how long do they have? If the answer is "we will use your existing documentation", the quote assumes your documentation is already correct and structured. It almost certainly is not.
2. Which systems does it read from, and who builds those integrations? A quote that names no systems is quoting a document-answering widget, whatever the sales deck showed.
3. What is the evaluation approach? If there is no test set and no measurement plan, nobody can tell you whether the chatbot is working, including you, and you will not be able to change models later without risk.
4. What happens when it cannot answer? The handoff, with transcript and context, is a build. A quote without one is quoting a dead end that customers will hit within the first hour.
5. What is the assumed token cost, and at what model? Ask for the per-conversation figure and the model it assumes. Then check the model's published price yourself, and check whether it is introductory.
6. Who owns it after launch? Content updates, prompt changes, model deprecations and retuning are continuous. A quote that ends at launch is quoting half the project.

A quote that answers all six with specifics will be higher than one that answers none. It will also be the one that resembles what you end up paying.

How to cut chatbot development cost without wrecking it

The largest saving available is not building a chatbot nobody uses, and the evidence that this is a real risk is uncomfortable. Gartner's survey of 3,566 customers in early 2026 found people about three times more likely to reach for a third-party generative AI tool than a company-provided chatbot, with company chatbot use statistically unchanged since 2022. In a parallel survey of 1,303 senior leaders, service and support teams had put a median of 12% of their 2025 budget into AI, more than any other function, and only 24% reported positive financial returns. Spending is not the constraint. Adoption is.

With that established, the levers that actually work:

• Buy first, build second. Prove people use it on a platform, then build once the volume or the integration requirements justify it. The build is much easier to scope when you have real conversation logs.
• Route by difficulty. Send straightforward questions to a nano-class model and escalate only the hard ones. Given a 100-fold price spread between the cheapest and most expensive models, getting most traffic onto the cheap one is the biggest recurring saving available.
• Cache the static context. The system prompt and the product documentation are identical across every conversation, and cache reads cost a tenth of fresh input tokens. This is a small change with a permanent effect.
• Trim the context. Retrieve fewer, better chunks. Truncate old conversation history. Every token you do not send is one you never pay for again, on every turn, forever.
• Set hard budget caps. A runaway loop should fail loudly rather than quietly draining an account overnight. Per-conversation and per-day ceilings cost nothing to add and have saved several teams a very bad month.
• Instrument before scaling. Know your cost per conversation and your resolution rate before you widen the traffic. Scaling something you have not measured is how a $37 monthly bill turns into a surprise.

What not to cut: the content work, the evaluation harness, and the human handoff. Those three are cheap to do at the start and expensive to retrofit, and a quote that is low because it skips them is a loan rather than a saving. If you are weighing whether to build in-house or bring in a team that has shipped this before, our guide to what it costs to hire AI engineers has the US market rates and the roles a chatbot build actually needs, and the AI integration services page describes how we connect models to systems you already run.