Chatbot Development Cost: Rule-Based vs. LLM Budgets

If you are trying to pin down chatbot development cost before talking to vendors, the short version is this: a narrow rule-based bot typically runs $10k-$40k to build, an AI-assisted support or sales chatbot lands between $40k and $120k, and an LLM chatbot wired into your business systems, compliance workflows, and human review can reach $120k-$300k or more. Those are planning ranges, not market averages. Your actual number depends on scope, integrations, data work, and how much conversation design you need before the first message goes out.

What catches most buyers off guard is the second budget line. A chatbot has two cost layers: the one-time build and the ongoing monthly run cost. Build cost covers design, development, testing, and launch. Run cost covers model usage or platform fees, infrastructure, monitoring, human escalation, and maintenance. Ignoring either side leads to a budget that looks reasonable on a slide deck and falls apart in production.

Chatbot development cost at a glance

The table below gives planning ranges across five common chatbot types. Use it to frame internal conversations, not as a quote.

These ranges assume a custom build by an experienced team. Off-the-shelf SaaS platforms compress the build cost but shift spending into per-resolution or per-conversation fees that can grow quickly at volume.

Rule-based chatbot vs. LLM chatbot cost

A rule-based bot follows decision trees. You script every path, wire up buttons and quick replies, and the bot never says anything you did not write. That makes it cheap to build and almost free to run. It also makes it brittle: any question outside the tree gets a dead-end fallback.

An LLM chatbot generates responses from a model, usually grounded in your own content through retrieval-augmented generation (RAG). It handles a wider range of questions, but the engineering surface is larger. You need data preparation, chunking and embedding of your knowledge base, prompt and system message design, retrieval tuning, output evaluation, guardrails to prevent hallucination or policy violations, observability to catch failures, and a human handoff path for cases the model should not resolve alone.

Each of those layers adds design time, testing, and ongoing cost. That is why the gap between a $15k rule-based bot and a $150k LLM bot is not about the model API call itself; it is about everything around it. If you are weighing whether your use case needs a chatbot or a more autonomous agent, the distinction between AI agents vs. chatbots is worth reading separately.

What changes the chatbot development cost

Scope and conversation count. A bot that handles three FAQ topics is a different project from one that covers order status, returns, product recommendations, and account changes. Every workflow needs its own conversation design, test cases, and edge-case handling.

Channels. Web widget, mobile SDK, WhatsApp, SMS, Slack, voice: each channel has its own API, UI constraints, and testing surface. Adding a second channel does not double the cost, but it is not free either. Expect 15-30% more build effort per channel after the first.

Conversation design. Good conversation design is the difference between a bot people use and one they immediately try to bypass. It includes tone, error recovery, disambiguation, and graceful handoff. Skipping it saves a few thousand dollars and costs you in deflection rate.

Knowledge and data work. If your knowledge base is a pile of outdated PDFs and tribal knowledge in Slack threads, someone has to clean, structure, and chunk that content before retrieval works well. Data work is often 20-40% of an AI chatbot build.

Integrations. Connecting to a CRM, order management system, ticketing platform, or payment gateway requires API work, auth handling, error management, and security review. Each integration adds $5k-$20k depending on the system's API quality and your security requirements.

AI model strategy. Using a single large model for every request is the simplest architecture and the most expensive to run. Routing simple queries to a smaller or cached model and reserving a frontier model for complex reasoning reduces token spend but adds engineering work upfront.

Compliance and security. Healthcare, finance, and legal use cases need data residency controls, audit logging, PII redaction, and sometimes human approval before the bot sends certain responses. These requirements can add 20-50% to the build.

Analytics and QA. You need to know what the bot is getting wrong. That means logging conversations, tagging outcomes, tracking resolution rates, and reviewing low-confidence responses. Building this instrumentation into the first release is cheaper than retrofitting it later.

Support and maintenance. Models change, knowledge bases drift, and user behavior shifts. Budget for monthly maintenance: prompt updates, knowledge refresh, model version upgrades, and bug fixes. A typical maintenance retainer runs $1k-$5k per month for a mid-complexity bot.

How monthly chatbot costs are calculated

Monthly run cost breaks down into a simple formula:

Monthly cost = platform/license fees + model or request usage + retrieval and storage + monitoring + human review and escalation + maintenance

For rule-based bots on platforms like Dialogflow CX (now part of Google Conversational Agents), you pay per request or per session. Costs stay low and predictable.

For LLM-powered bots, the model usage line is where surprises live. The math is:

Token cost = number of requests x average tokens per request (input + output) x provider rate per token

Providers bill input and output tokens at different rates. OpenAI's API pricing shows tiered rates across model families, with cached input tokens priced lower than fresh input tokens. Anthropic's Claude pricing follows a similar structure and offers prompt caching and batch processing discounts.

A support bot handling 50,000 conversations per month with an average of 800 input tokens and 300 output tokens per turn, running two turns per conversation, can generate 110 million tokens monthly. At frontier model rates, that is a meaningful line item. Routing straightforward queries to a smaller model or using cached system prompts can cut token spend by 40-60% when the architecture supports it.

SaaS chatbot platforms price differently. Intercom's Fin AI agent charges $0.99 per resolution. Salesforce Agentforce charges $2 per conversation. At 10,000 resolutions per month, those models produce $9,900 or $20,000 in monthly fees respectively. Verify current rates directly with each vendor; these figures change.

Three budget examples

Ecommerce product recommendation and order-status bot. A mid-size retailer wants a chatbot on their website and mobile app that recommends products based on browsing history and answers order-status questions by pulling from their OMS API. Build range: $45k-$90k. The lower end assumes a clean product catalog and a well-documented OMS API. The upper end accounts for messy product data, a second channel (mobile SDK), and conversation design across multiple languages. Monthly run cost: $1k-$6k, driven by model usage, OMS API calls, and a part-time conversation analyst reviewing escalated sessions.

B2B lead qualification bot connected to CRM. A SaaS company wants a chatbot on their marketing site that qualifies inbound leads, books meetings, and pushes data into HubSpot or Salesforce. Build range: $25k-$70k. Rule-based qualification flows sit at the low end. Adding an LLM layer to handle freeform questions about pricing and features, plus calendar integration and lead scoring logic, pushes toward the upper range. Monthly run cost: $500-$3k, mostly CRM platform fees and light model usage.

Regulated support bot with knowledge retrieval and approval gates. A financial services firm needs a chatbot that answers customer questions using approved regulatory content, flags responses that touch compliance-sensitive topics for human review before sending, and logs every interaction for audit. Build range: $120k-$260k. The spread reflects the depth of compliance engineering: PII handling, approval workflows, audit trail architecture, and security review. Monthly run cost: $8k-$40k, with the high end driven by human reviewer staffing, model usage on long regulatory documents, and infrastructure for data residency.

How to keep the budget under control

Start with one workflow. Pick the single highest-volume, lowest-risk conversation type. Build, measure, and expand from there. Trying to automate everything in v1 inflates scope and delays launch.

Cap automation scope explicitly. Define what the bot should not do. A clear boundary prevents scope creep during build and reduces the guardrail engineering needed at launch.

Clean the knowledge base before build starts. Every week your team spends cleaning content during development is a week the project timeline extends. Do this work in parallel with vendor selection or before the engagement begins.

Design fallback and escalation from day one. A bot that fails silently damages trust. A bot that says "Let me connect you with a person who can help" and does so quickly is a bot people will use again.

Instrument unresolved intents. Log every question the bot cannot answer. This data tells you what to build next and whether your knowledge base has gaps.

Choose model tiers by task. Use a smaller, faster model for intent classification and a larger model for generation. This is standard practice in production AI chatbot development services and can meaningfully reduce token spend.

Estimate peak usage honestly. If your support volume doubles during holiday season, your model costs double too. Build that into the annual budget, not just the monthly average.

Agree on acceptance tests before build. Define what "done" looks like in measurable terms: resolution rate targets, response latency, escalation rate, and accuracy on a test set. This protects both sides.

If your project spans multiple AI capabilities beyond a chatbot, such as document processing, recommendation engines, or workflow automation, a team offering custom AI solutions can help you plan a shared architecture that avoids redundant infrastructure. Similarly, if the main challenge is connecting an LLM to your existing systems rather than building from scratch, scoping the project as AI integration services often produces a tighter budget. For a broader look at how AI project budgets work across application types, the guide on AI app development cost covers adjacent territory.

Vendor questions before you approve the estimate

Use this checklist when reviewing proposals:

• Which channels are included in the build estimate, and what does adding another channel cost?
• Who is responsible for knowledge base preparation: your team, the vendor, or both?
• Which integrations are in scope, and what assumptions has the vendor made about API quality and access?
• How are model updates and prompt changes handled after launch? Is there a retainer, or are changes billed hourly?
• What usage volume does the monthly cost estimate assume? What happens if volume doubles?
• How is PII handled in transit and at rest? Where are conversations stored?
• What analytics and reporting are included? Can you export raw conversation logs?
• How does escalation to a human agent work, and is the escalation integration included?
• What is the post-launch support SLA, and what does it cover?
• Who owns the prompts, conversation data, and any fine-tuned model artifacts?

Getting clear answers to these questions before signing prevents the most common budget overruns.

Vladimir Terekhov

Co-founder and CEO at Attract Group