A client called us last quarter in a mild panic. They'd spent $40,000 on a chatbot platform subscription, and after three months of configuration, their bot was giving customers answers like "I'm sorry, I cannot help with that. Please contact support." For 70% of queries. They were paying premium pricing for a fancy error message.
On the flip side, we've seen startups burn $80,000 building a custom chatbot from scratch when a $200/month platform would have done the job perfectly. The build vs. buy decision for AI chatbots is genuinely hard, and most businesses get it wrong because they evaluate based on demos, not deployment reality.
This article shares what we've learned from building and deploying both custom and platform-based chatbots across e-commerce, SaaS, healthcare, and professional services.
The AI Chatbot Landscape in 2026
The chatbot market has split into three clear tiers:
Tier 1: No-code platforms — Intercom Fin, Drift AI, Zendesk AI, Ada. These plug into your existing support stack, train on your help docs, and start answering questions within days. Great for standard customer support. Limited customization.
Tier 2: AI-first platforms — Voiceflow, Botpress, Kore.ai, Yellow.ai. More flexible than Tier 1 — you can design complex conversation flows, integrate with multiple systems, and customize the AI behavior. Takes weeks to configure properly, but handles sophisticated use cases.
Tier 3: Custom-built — Your own chatbot built on LLM APIs (Claude, GPT-4) with RAG for knowledge retrieval, custom tool integrations, and tailored conversation logic. Maximum flexibility, maximum effort. Takes 4-12 weeks to build.
The right tier depends on three factors: how unique your use case is, how much system integration you need, and your budget.
The Buy Path: What Platforms Actually Deliver
What Works Well
- FAQ deflection: Platforms excel at answering questions that exist in your help documentation. "What are your shipping rates?" "How do I reset my password?" If your support volume is dominated by repetitive questions with documented answers, platforms deliver fast ROI.
- Chat-to-human handoff: Built-in escalation flows, queue management, and context passing to human agents. This is table stakes for any support chatbot and platforms handle it well.
- Analytics: Pre-built dashboards for resolution rates, satisfaction scores, common topics, and gap analysis. Getting this right in a custom build takes weeks.
Where Platforms Break Down
- Complex integrations: When the chatbot needs to pull data from your CRM, check inventory in your ERP, process a refund through your payment system, and update a shipping record — all in one conversation — platforms struggle. Most support 10-20 pre-built integrations; anything beyond that requires custom development.
- Domain-specific knowledge: Medical advice, legal guidance, financial recommendations — areas requiring deep domain expertise and careful guardrails. Generic platforms produce dangerously generic answers in these contexts.
- Multi-step transactions: Booking a complex service, configuring a custom product, processing an insurance claim. These require conditional logic, state management, and system coordination that most platforms handle poorly.
Real Costs (What They Don't Show in the Demo)
| Cost Component | Monthly Range |
|---|---|
| Platform subscription | $500 - $5,000 |
| AI resolution fees (per-conversation pricing) | $200 - $3,000 |
| Setup and configuration (agency or internal) | $5,000 - $20,000 (one-time) |
| Ongoing tuning and content updates | $500 - $2,000 |
The per-conversation AI pricing catches people off guard. At $0.50-2.00 per AI-resolved conversation, a chatbot handling 5,000 conversations/month adds $2,500-10,000 to your bill on top of the subscription fee.
The Build Path: Custom AI Chatbots
Architecture
A custom chatbot typically involves:
- LLM API — Claude or GPT-4 for conversation and reasoning
- RAG system — Vector database with your knowledge base for accurate answers
- Tool integrations — Direct API connections to your systems (CRM, ERP, payment, shipping)
- Conversation manager — State tracking, context memory, escalation logic
- UI widget — Chat interface embedded on your site or app
- Admin dashboard — Conversation history, analytics, content management
Real Build Costs
| Scope | Build Cost | Monthly Running |
|---|---|---|
| Basic (FAQ + handoff) | $8,000 - $15,000 | $100 - $500 |
| Standard (RAG + 2-3 integrations) | $20,000 - $40,000 | $200 - $1,000 |
| Advanced (multi-agent, deep integrations) | $50,000 - $120,000 | $500 - $3,000 |
The monthly running costs for custom builds are dramatically lower than platforms because you're paying LLM API costs directly (pennies per conversation) instead of the platform's marked-up per-conversation pricing.
Where Custom Wins
- Deep system integration: Your chatbot can do anything your API can do — process refunds, modify orders, update records, trigger workflows.
- Brand voice: Full control over tone, personality, guardrails, and conversation style.
- Cost at scale: Custom chatbots cost $0.02-0.10 per conversation in LLM fees vs. $0.50-2.00 on platforms. At 10,000+ monthly conversations, custom saves $5,000-20,000/month.
- Data ownership: All conversation data stays in your infrastructure. No third-party access.
Head-to-Head: Build vs Buy
| Factor | Buy (Platform) | Build (Custom) |
|---|---|---|
| Time to deploy | 1-4 weeks | 4-12 weeks |
| Upfront cost | $5K-20K setup | $8K-120K build |
| Monthly cost (5K conversations) | $3K-15K | $200-1K |
| Customization depth | Limited | Unlimited |
| Integration flexibility | Pre-built connectors | Any API |
| Maintenance burden | Vendor handles | Your team / agency |
Our Decision Framework
After dozens of chatbot projects, here's how we advise clients:
Buy a platform when:
- Your primary use case is FAQ deflection with documented answers
- Volume is under 5,000 conversations/month (platform pricing is reasonable)
- You need it deployed within 2 weeks
- Your support stack is already Zendesk, Intercom, or Salesforce (native integration)
- You don't have engineering resources to maintain a custom system
Build custom when:
- You need deep integration with internal systems (CRM, ERP, custom databases)
- Volume exceeds 10,000 conversations/month (custom saves $5K+/month)
- Your domain requires specialized knowledge and strict guardrails (healthcare, finance, legal)
- Brand voice and conversation design are competitive differentiators
- Data privacy requirements prohibit third-party conversation access
The Hybrid Approach (What We Often Recommend)
For many clients, the best answer is neither pure build nor pure buy — it's hybrid. Start with a platform for quick wins, then build custom components for the complex use cases the platform can't handle.
Example: Use Intercom Fin for standard FAQ deflection (deployed in a week), then build a custom AI agent for complex transactions (order modifications, refund processing, account changes) that the platform routes to when it detects a complex query.
This gives you 80% of the value in 20% of the time, with a clear path to full customization as your needs grow.
Need help deciding? Our AI consulting team provides a free assessment of your chatbot needs — we'll recommend build, buy, or hybrid based on your specific use case, volume, and integration requirements.
Frequently Asked Questions
How long does it take to train a chatbot on our knowledge base?
For platforms: 1-3 days to ingest your help docs and start answering. Accuracy improves over 2-4 weeks of tuning. For custom RAG-based chatbots: 1-2 weeks to build the ingestion pipeline, embed your documents, and tune retrieval quality.
What resolution rate should we expect?
For well-scoped FAQ use cases: 50-70% automated resolution. For custom chatbots with system integrations: 40-60% full resolution, 20-30% partial resolution (agent gets pre-filled context). Anything claiming 90%+ resolution is either narrowly scoped or measuring incorrectly.
Can a chatbot handle multiple languages?
Modern LLM-based chatbots handle 50+ languages natively. Claude and GPT-4 are strong in European languages, Mandarin, Japanese, Hindi, and Arabic. Quality drops for less-common languages. Platform support varies — check language support before committing.
How do we measure chatbot ROI?
Track: tickets deflected (cost saved), average handle time reduction (for escalated tickets with AI context), CSAT for bot-resolved vs. human-resolved, and containment rate (% of conversations fully resolved without human). Compare monthly bot cost against equivalent support headcount cost.