Chatbot Development Cost in 2026: Build Price and Per-Conversation Math

Chatbot cost by type: the short answer

"Chatbot" covers a decision tree that collects leads and an assistant that reads a customer's order history, explains a delay and issues a refund. Those are different products. Price follows three questions: where the answers come from, what the bot is allowed to do, and how many channels it lives in.

These ranges come from the estimates we prepare and the competing quotes clients show us in first calls. They cover a production bot: the conversation logic, an admin view for transcripts, analytics, a human handoff, QA and deployment. They don't include your own team's time to write or clean help content, which is often the hidden half of the project.

This guide is about conversational assistants: support, sales, internal help desks. If the AI is one feature inside a bigger app, read AI app development cost. If the system plans and executes multi-step work across tools on its own, AI agent development cost is the closer match. Machine learning projects without a chat interface are covered in AI development cost.

Build cost by bot type and what drives it

The jump from a rules-based bot to an AI assistant is mostly content and testing work. Someone has to collect the help articles, remove the three outdated versions of the returns policy, decide what the bot must never answer, and build a set of real customer questions with the correct answers. The jump from an assistant to a transactional bot is integration work. Each system the bot touches (Shopify, Salesforce, Zendesk, a booking engine, your own API) adds authentication, permission checks, error handling and tests. One integration of medium complexity typically adds $5k–$20k; our API integration cost guide breaks that down.

The other drivers, roughly in order of how much they move the price:

• Channels. A web widget is the cheapest. WhatsApp, SMS, Slack, Microsoft Teams and in-app SDKs each have their own message formats, limits and approval rules. Two to three channels add 15–30%.
• Voice. A phone agent needs speech-to-text, text-to-speech, telephony, interruption handling and latency work. It also needs far more testing, because people speak less precisely than they type.
• Languages. The model may handle Spanish fine, but your test set, content review and escalation rules have to exist in Spanish too. Plan 10–20% for two or three languages.
• Compliance. HIPAA requires a business associate agreement with every vendor that touches protected health information, including the model provider, and that narrows your choice of models and plans. SOC 2 reviews add logging, retention rules and access controls.
• The cost of a wrong answer. A bot that recommends blog posts can be wrong now and then. A bot that quotes refund terms creates obligations: in Moffatt v. Air Canada (February 2024), a British Columbia tribunal held the airline liable for a refund policy its website chatbot made up and rejected the argument that the chatbot was responsible for its own words. Guardrails and evaluation are what you pay to avoid that.

The running cost: what one AI conversation costs

A rules-based bot costs about the same to host at 100 conversations or 100,000. An AI chatbot doesn't. On every turn the app sends the model its instructions, the help articles it retrieved, the conversation so far and the new message, and the provider bills each token in and out. Because history is resent on every turn, a long conversation costs more than the sum of its messages.

The math behind the figure. Anthropic lists Claude Haiku 4.5 at $1 per million input tokens and $5 per million output tokens, Claude Sonnet 5.5 at $2 and $10, and Claude Opus 5.5 at $4 and $20 at the time of writing. OpenAI's GPT-5 nano sits at $0.05 and $0.40. Multiply by 36,000 input and 1,800 output tokens and you get the four rows. OpenAI and Google price their mid-size models in the same broad band, so the order of magnitude holds whichever provider you pick.

Three things change these numbers in practice:

• Prompt caching. The system prompt is identical on every turn. Providers bill cached input at a fraction of the normal rate (on Anthropic's current models, cache reads cost a tenth of the input price or less), which removes most of the system prompt cost.
• Retrieved context. Sending three good help articles instead of eight mediocre ones cuts the largest input line by half and usually improves answers too.
• Conversation length. Because history is resent each turn, a 12-turn conversation costs about three times a 6-turn one, not twice. Summarizing older turns or ending idle sessions keeps it in check.

Voice changes the scale. A phone agent pays for telephony (Twilio lists US inbound local calls at $0.0085 per minute), speech recognition and speech synthesis on top of the model. In the estimates we prepare, an all-in voice stack lands at roughly $0.05–$0.15 per minute before the model, so a four-minute call costs tens of cents rather than a few cents.

That first pair of numbers is why companies build chatbots. Gartner's 2019 customer service poll put live channels (phone, chat, email) at $8.01 per contact and self-service at $0.10. The same poll found only 9% of customers fully resolved their issue through self-service. A bot that costs pennies per conversation but resolves nothing simply adds a step before the $8 contact.

Chatbot cost calculator: build, run and break-even

The calculator estimates what the bot costs to build, what it costs per month once it handles real volume, and how that compares with paying a SaaS agent per resolution. Hours behind it are the ranges we use for first estimates. If you request a quote, your inputs go with it.

Play with the volume first. At 1,000 conversations a month a custom AI assistant costs more to maintain than a SaaS agent charges in fees, and it never pays back on cost alone. At 100,000 conversations the build pays back within months, and the model tier becomes the line worth optimizing. In between, the answer depends on the resolution rate you can honestly expect, and on whether the SaaS tool can do what your bot has to do.

Build or buy: where the lines cross

Off-the-shelf AI agents have gotten good at answering from a help center. Intercom lists Fin at $0.99 per resolution, with a 50-resolution monthly minimum (about $50) when it runs on another helpdesk; other helpdesk vendors also sell AI resolutions on usage-based pricing. For many companies that is the right first step, and we say so in first calls. The calculation changes with volume.

Fee math is only half of it. In practice, these are the reasons we see companies build a custom chatbot even when a SaaS agent looks cheaper on paper:

• The bot has to act in a system the SaaS tool can't reach, such as a homegrown order system, a legacy booking engine or an ERP.
• The bot is part of the product (inside a mobile app, a SaaS dashboard, a device), not a support widget.
• Data must stay in a specific region or never leave a specific cloud account.
• The company wants to choose and switch models, or run open models on its own infrastructure.
• Per-resolution pricing punishes success: the better the bot gets, the bigger the bill.

And the reasons to buy: low volume, a standard helpdesk, answers that come from a help center, and no engineering team to own the bot after launch. A middle path is common too. Keep the SaaS agent for general questions and build a small custom service that it calls for order lookups or account actions. If you want to compare vendors who do that integration work, we keep a list of chatbot integration companies.

Built by Gilzor

Results we’ve shipped

70+products launched

98%delivered on time

85%clients come back

Aviation33%faster baggage weigh-inWeb and Mobile application for Airlines→E-commerce · AITaking an AI-built e-commerce SEO platform to production→Healthcare+30%lab test ordersMedical Lab Web and Native Mobile Applications→

Talk to the people who build it. Tell us about your project and get a free estimate of scope, timeline and cost.

See how we’d approach yours →

Quiz: build, buy or fix the content first?

Six questions. The result tells you which route fits your chatbot today, which matters more for the budget than any single feature. We ask the same questions in first calls.

What vendors in each region charge

Region moves the build price more than any single feature. The rates below are for senior engineers hired through a vendor in 2026; engineers with production LLM experience usually bill 15–30% above a regular backend developer in the same region.

For an in-house comparison: the US Bureau of Labor Statistics put the median software developer wage at $135,980 in May 2025, before benefits, payroll taxes and recruiting. Two engineers who have shipped LLM features cost more than most first chatbot builds, every year. Gilzor sits in the CEE row, with teams in Poland and Cyprus. That gives a few shared hours with New York when schedules shift and very little with California, which works for a chatbot build because most of the work is asynchronous: content review, test sets, integration tickets. Our nearshore rates guide compares Latin America and Europe country by country.

Where chatbot budgets blow up

Gartner predicted in March 2025 that by 2029 agentic AI will resolve 80% of common customer service issues without human help. Plenty of current projects are nowhere near that, and from what we see in estimates and first calls, the reasons are rarely the model:

• The knowledge base isn't ready. "Our help center has everything" usually means half the answers are outdated and the other half live in agents' heads. Content work is the line most often missing from cheap quotes, and the one that decides the resolution rate.
• No test set. Without 100–300 real customer questions and the correct answers, nobody can say whether a prompt change or a cheaper model made the bot better or worse. Teams end up arguing from screenshots.
• Integrations discovered late. "It should also check the order" turns a knowledge assistant into a transactional one, usually in month three. That can add 30–60% to the budget. Decide the actions before the estimate.
• Too many channels at launch. Every channel multiplies testing. Teams that launch on web, WhatsApp and Teams at once spend the first months fixing formatting, not answers.
• No handoff design. A bot that traps people is worse than no bot. The escalation path, what context passes to the agent, and what the bot says when it doesn't know all need to be designed and tested.
• Model churn. Providers retire model versions on a published schedule. Each switch means re-running your test set and sometimes rewriting prompts. Budget a few days per switch, a few times a year.

Hidden costs to put in the budget

Our QA team treats every reported bad answer as a bug with a test case attached. Only 5% of tasks sent to QA come back to our developers, an internal metric we track; on chatbots, the test set is what keeps that number meaningful.

How to reduce the cost without breaking the bot

1. Fix the top 50 answers firstPull the most common questions from your helpdesk and make sure each has one current, written answer. This is the cheapest work in the project and it moves the resolution rate more than any model choice.
2. Launch on one channelUsually the website widget or in-app chat. Add WhatsApp, SMS or voice after answers are reliable, using the same backend.
3. Read before you writeStart with lookups (order status, account details) and add actions like refunds or rebooking in a second phase, with confirmation steps.
4. Route by difficultySend simple questions to a small model and only the hard ones to a larger model. With caching and trimmed context, this commonly cuts model fees by half or more.
5. Build the test set on day oneA few days of work. It lets you switch to a cheaper model with evidence and catches regressions before customers do.
6. Use a SaaS agent where it fitsLet an off-the-shelf agent handle generic questions and build only the custom part: the integration or the action your business needs.
7. Design the handoff earlyA clean escalation with full context saves agent time on every conversation the bot can't finish, which protects the business case.

What we cut first from a first release: extra channels, extra languages, voice, and actions that change data. What we don't cut: the test set, logging, the human handoff and usage limits. Cutting those makes launch cheaper and everything after it more expensive. If the chatbot is part of a new product, our MVP development cost guide shows how to size the first release around it.

How to get comparable chatbot quotes

Send each vendor the same one-page brief: channels, systems the bot touches, languages, compliance needs, monthly conversation volume and who owns the content. Then ask for answers in the same shape:

• Build cost split into conversation logic, content preparation, integrations, channels, QA and project management.
• Which model they plan to use, why, and the estimated model cost per 1,000 conversations at your volume.
• How answer quality is measured, on which test set, and what counts as done.
• The monthly running cost at today's volume and at three times that.
• The plan for model retirements, provider price changes and outages.

A vendor who can answer these in a first call has run chatbots in production. Our AI/ML team prices chatbots in exactly this format, and the business analysis phase is where we settle the questions that change the price most: which actions, which channels, which content.

FAQ

Where Gilzor fits

We build AI assistants and the systems around them: integrations, web and mobile apps, QA, and the analysis that decides what a bot should and shouldn't do. 85% of our clients come back for the next project, and we have launched 70+ products for startups and SMBs over more than seven years. Send us your channels, your systems and your volume, and we'll price both numbers that matter: the build and the cost per conversation. If a SaaS agent is the better answer for you, we'll say that too.