Buy an AI support agent such as Intercom Fin (from $0.99 per outcome) or Zendesk AI agents ($1.50 to $2.00 per automated resolution beyond plan allowances) when answers live in a help centre and volume is moderate. Build a custom retrieval-augmented (RAG) chatbot when answers depend on your own systems, data must stay under your control, or per-resolution fees outgrow engineering cost. Published 2026 guides price custom builds from about $15,000 to $150,000, with enterprise platforms above that.
An AI chatbot in 2026 is a language model plus four things: retrieval over your content, tools that read your systems, rules about what it may say and do, and a way to hand the conversation to a person. The model is the easy part. The rest decides whether customers trust it, and the law treats its answers as yours: in Moffatt v. Air Canada (February 2024), British Columbia's Civil Resolution Tribunal held the airline liable for a bereavement-fare policy its chatbot misstated and ordered it to pay about C$812.
Build vs buy: the comparison
| Option | Pricing | Time to launch | Control of data and behaviour | Best for |
|---|---|---|---|---|
| Intercom Fin | From $0.99 per Fin outcome, plus seats from $29 per seat a month (billed annually) | Days to weeks if your help centre is current | Configured within Intercom's product and policies | Teams already on Intercom with help-centre answers |
| Zendesk AI agents | 5 automated resolutions per agent a month on Suite Team and 10 on Professional, then $1.50 committed or $2.00 pay-as-you-go; Suite from $55 per agent a month | Days to weeks inside an existing Zendesk | Configured within Zendesk's product and policies | Support teams standardised on Zendesk |
| Custom RAG chatbot | Build cost plus model, hosting and maintenance; model cost often cents per conversation | 4–12 weeks, depending on integrations | Full: your models, prompts, data, logs and evaluation set | Answers that need your own systems, strict data rules, high volume or an in-product experience |
When buying is the right call
Buy when your answers already live in a maintained help centre, your conversations are mostly questions rather than actions, and your volume keeps per-resolution fees below what an engineer would cost. You get a tested widget, handoff into the helpdesk your agents already use, and reporting on day one. When the No Yelling driving-school app, which our team has maintained since 2020, needed in-app support in 2024, we integrated Intercom's chat rather than building one; a small business needs a working support queue more than a custom platform.
When building is the right call
- Answers depend on live data, such as order status, bookings, balances or entitlements that only your APIs know.
- The bot must act, not just answer: rescheduling, refunding within limits or updating records, with permission checks.
- Data rules are strict: residency, retention or contractual limits on which providers may see conversations.
- Volume makes per-resolution pricing expensive: 10,000 resolutions a month at $0.99 is $9,900 a month on Fin, before seats.
- The assistant is part of your product, with your interface, tone and analytics, rather than a support widget.
Build model-agnostic. Spyra Beauty's internal document tool, a Next.js app we built that turns supplier PDF, Word and Excel sheets into branded PDFs, uses both the Claude and OpenAI APIs, and SlimAI uses Gemini. Keeping the model behind an interface means a price or quality change at one provider is a configuration change and an evaluation run, not a rewrite.
What a good production chatbot looks like
Retrieval over your documents
Retrieval-augmented generation (RAG) means the bot searches your content first and answers from what it finds, with citations. Split documents into passages, embed them, and store the vectors in pgvector inside your existing Postgres or in a managed service such as Pinecone, whose Standard plan has a $50 monthly minimum. Embedding is cheap: OpenAI lists text-embedding-3-small at $0.02 per million tokens. Combine vector search with keyword search for product names and codes, re-sync content when it changes, and filter by customer and permission before retrieval, not after.
Handoff to humans
Define when the bot stops: low retrieval confidence, a request for a person, account disputes, safety topics, or two failed attempts. The handoff should carry the transcript, the retrieved sources and the customer's identity into your helpdesk, so nobody asks the customer to repeat themselves.
An evaluation set
Before launch, collect a few hundred real questions with known-good answers and sources, including questions the bot should refuse or escalate. Run the set on every prompt, model or content change, and track answer correctness, citation accuracy and correct handoffs. Add every production failure to the set.
Cost per conversation
Model cost is usually small next to per-resolution fees. A six-turn support conversation that sends about 3,000 input tokens per turn, covering instructions, retrieved passages and history, and returns about 250 output tokens per turn uses roughly 18,000 input and 1,500 output tokens. At a typical mid-tier model price of around $2 per million input tokens and $10 per million output, that is about 5 cents; a small, cheap model brings it under one cent. Hosting, monitoring and engineering time come on top, so compare total cost per resolved conversation.
Logging
Log every conversation with the retrieved passages, model and prompt version, latency, cost and the customer's rating, with personal data redacted and a retention period you can defend. Logs are how you find bad answers, and they become the next evaluation cases.
Cost drivers
| Driver | Lower cost | Higher cost |
|---|---|---|
| Knowledge sources | One help centre or document set | Many systems with different formats and owners |
| Actions and integrations | Answers only | Authenticated actions across several APIs with permission checks |
| Channels | One web widget | Web, in-app, WhatsApp and voice |
| Languages | One | Several, each with its own evaluation set |
| Compliance | Public information only | Personal, health or financial data with audit and residency requirements |
| Evaluation depth | A few hundred test questions | Large graded sets, red-teaming and regression gates in CI |
| Volume | Hundreds of conversations a month | Hundreds of thousands, where model choice and caching decide margins |
Published 2026 guides show how much definitions vary. Codiant's July 2026 guide prices an AI chatbot MVP at $15,000 to $50,000 over 6 to 12 weeks and a custom RAG chatbot at $35,000 to $150,000 over three to six months, with maintenance at 15 to 25% of the build each year. Salt Technologies' February 2026 guide prices a RAG support bot at $15,000 to $35,000 over two to four weeks. Ask any vendor which of the drivers above their number includes.
Security and data handling
- Treat prompt injection as the top risk, as OWASP's 2025 Top 10 for LLM applications does. Retrieved documents and user messages are untrusted input and must never change the bot's permissions.
- Limit agency. The bot calls tools with the customer's permissions, not an admin key, and anything irreversible needs confirmation.
- Isolate tenants in retrieval. Filter vector search by customer before ranking, so one account's documents can never appear in another's answer.
- Know your provider's terms. OpenAI does not train on API data by default and keeps abuse-monitoring logs for up to 30 days; zero data retention and regional data residency require its approval.
- Tell users they are talking to AI. Article 50 of the EU AI Act requires this for systems that interact with people, from 2 August 2026.
- Cap spend and rate. OWASP lists unbounded consumption as a risk; per-user rate limits and a monthly budget stop a loop or an abuser from running up costs.
Launch checklist
- The evaluation set passes at the threshold agreed before the build, including refusals and handoffs.
- Handoff works in and out of business hours, with a clear message when nobody is available.
- The chat window says the customer is talking to AI and how to reach a person.
- Logs capture sources, cost and ratings, with personal data redacted.
- Rate limits, a spend cap and alerts are live.
- A kill switch can route every conversation to people within minutes.
How long it takes
In our experience, a production RAG support bot over one knowledge base, with handoff, logging and an evaluation set, takes four to six weeks. Add authenticated actions across several systems and plan for eight to twelve. The long pole is rarely the model. It is cleaning the content, agreeing the handoff rules and building the evaluation set with the people who answer customers today. Our AI chatbot team scopes this under NDA on request, sends an estimate within two business days, and hands over code, prompts and evaluation data that you own.
Frequently asked questions
- Which company developed the AI chatbot ChatGPT, and can my business have its own?
- OpenAI developed ChatGPT and released it on 30 November 2022. Your business can have its own assistant in three ways: ChatGPT business plans with custom GPTs for internal use, a bought support agent such as Intercom Fin or Zendesk AI, or a custom chatbot built on OpenAI, Gemini or Anthropic models with your data, branding and controls.
- How much does it cost to build an AI chatbot?
- Published 2026 guides range from about $15,000 for a focused RAG support bot to $150,000 or more for custom builds with integrations, with maintenance at 15 to 25% of the build each year. Model cost per conversation is often cents. Bought agents charge per outcome instead: from $0.99 on Intercom Fin and $1.50 to $2.00 on Zendesk.
- What are the top AI chatbot development companies?
- Lists vary by source and month, and directories such as Clutch note they may earn fees for some placements, so treat rankings as a starting point. Evaluate shortlisted firms on an evaluation set from a past project, their handoff design, a cost-per-conversation model, written data handling, a live reference and a contract that assigns you the code and prompts.
- How long does chatbot development take?
- A focused RAG chatbot over one knowledge base, with human handoff and an evaluation set, typically takes four to six weeks. Adding authenticated actions across several systems pushes that to eight to twelve weeks. Codiant's July 2026 guide gives 6 to 12 weeks for an AI chatbot MVP and three to six months for a custom RAG build.
About the author
Zoraiz Ejaz
Co-founder, Techparser
Zoraiz Ejaz is a co-founder of Techparser and leads its engineering and product practice. He has spent close to a decade designing, building and scaling mobile, web and AI products for startups and enterprise teams across health, fintech, payments, social and education, from first architecture and release pipelines through to launch and years of production support. He writes about how to scope, cost and ship software that lasts.
