The phrase "AI chatbot" now covers everything from a $50-a-month widget you paste onto a page to a system that answers questions from your private data, takes actions, and represents your company to every customer who lands on the site. Those are not the same product, and pricing one as if it were the other is how projects go sideways.
So before anyone quotes you, it's worth understanding what actually drives AI chatbot development cost. It's rarely the conversation part — models are good at chatting out of the box. It's everything around the conversation: what the bot knows, what it's allowed to do, and how you keep it from embarrassing you. Duskel builds custom chatbots from $2k, with the more involved ones running as retainers from $3k a month. Here's how to tell which one you need.
Off-the-shelf versus custom
If your needs are genuinely generic — a FAQ bot, basic lead capture, canned answers — don't pay for custom development. A hosted tool will do it for a monthly subscription and you'll have it running this week. Anyone who quotes you thousands to reskin a SaaS widget is selling you the wrong thing.
Custom development earns its cost the moment the bot needs to know something the public internet doesn't: your product catalog, your policies, your account data, your specific way of doing things. That's a retrieval system underneath the chat, and building it so the answers are accurate and cited is real work. The line between "buy the widget" and "build it custom" is basically: does it need to answer from your data, and does a wrong answer cost you anything? Two yeses means custom.
The cost tiers, honestly
Simple: a chatbot grounded in one clean knowledge source that answers questions and hands off to a human when it's out of its depth. Support deflection for a well-documented product, a smart assistant on your docs. Accurate retrieval, citations, a sensible "let me connect you to someone" fallback. This is the from-$2k end and it covers a lot of real business needs.
Mid: a bot that pulls from multiple sources, remembers context across a conversation, handles authentication so it can speak to a logged-in customer about their own account, and maybe does a few safe actions like checking an order status. Now you're integrating with real systems and handling permissions, so it's a monthly engagement rather than a one-off build.
Complex: a chatbot that takes meaningful actions, operates in a regulated or high-stakes context, handles high volume, or is effectively the front door to your business. It needs guardrails, escalation logic, monitoring, and continuous tuning against real conversations. This is ongoing retainer work, because a bot talking to thousands of customers is something you maintain, not something you launch and forget.
The cost drivers that sneak up on you
Integrations are the quiet budget-killer. A bot that only talks is cheap. A bot that reads a customer's order, checks inventory, or updates a ticket has to connect to systems that each have their own quirks, auth, and failure modes. Every integration is a small project inside the project.
Tone and safety take more time than people expect too. A chatbot speaks for your brand to everyone who uses it, so it needs boundaries: what it won't discuss, how it handles someone trying to bait it, when it refuses and hands off. Getting that right is testing work, and it's the difference between a bot that helps and a screenshot that ends up on social media for the wrong reasons.
Then there's the running cost. Every conversation is model tokens, and a busy bot generates a lot of them. It's usually modest per chat but it scales with traffic, so it's a real line item to plan for rather than a rounding error.
Where teams overspend
The biggest waste is building a custom bot for a generic job. If a subscription widget covers it, use the widget and spend the money where you're actually different. The second is launching without measuring deflection or satisfaction, so you can't tell whether the thing is working. A chatbot you can't measure is a chatbot you can't improve, and you end up paying to maintain something nobody can prove is worth it. Start narrow, ground it well, measure honestly, and expand into the expensive tiers only when the cheap one has earned it.