Build vs Buy an AI Support Chatbot: What DIY Really Costs in 2026

Every developer who has wired a chat box to an embedding index knows the feeling. The demo answers the first ten questions well, and finishing it looks like a few more evenings. Then the list grows. The bot has to stay in sync with the website, hand off to a person, live on someone else's page without breaking it, ignore instructions hidden in a PDF, and keep transcripts you can delete on request.

This guide is for the technical founder, CTO or developer weighing that decision for a customer support chatbot: the kind that answers visitors on your website from your own content. An AI support agent is a chatbot that answers from your content, can act through tools and hands off to a human when it cannot help; the full definition covers how it differs from a scripted bot. Below you will find what a production agent needs, rough effort for each part, the model bill at OpenAI's published prices as of October 2026, and a side-by-side monthly cost with a hosted tool at 2,000 and 10,000 conversations.

Effort figures are our estimates; prices come from the vendors' own pages. If your question is whether to use ChatGPT itself rather than the API, that is a different decision, because custom GPTs live inside ChatGPT rather than on your site; the guide to ChatGPT for customer service covers it.

What can you build in a weekend?

A weekend gets you a convincing retrieval prototype that answers the questions you thought of from your own pages. Retrieval-augmented generation (RAG) is the pattern under every grounded support bot: find the passages in your content that match the question, then ask the model to answer from those passages only. The core loop fits on one screen.

// The weekend prototype, in outline (helper functions omitted)
const pages = await crawl("https://example.com/sitemap.xml");
const chunks = split(pages, { size: 800, overlap: 100 });
await db.insert(chunks, await embed(chunks, "text-embedding-3-small")); // pgvector

async function answer(question: string) {
  const passages = await db.nearest(await embed([question]), { limit: 5 });
  return chat("gpt-4o-mini", {
    system: "Answer only from the passages. If they do not cover the question, say so.",
    passages,
    question,
  });
}

Fetch your help pages, split them into chunks, embed them, store the vectors in Postgres with the open-source pgvector extension, retrieve the five closest chunks for each question and send them to a small chat model with an instruction to answer only from them. Put a text box on a page and you have a demo. The RAG explainer walks through each step in plain language.

The prototype is still a useful test of your content. If it cannot answer your top questions from your own pages, no product will either, and the fix is better pages rather than better code.

What the prototype lacks is everything a customer touches after the first answer: the question your content does not cover, the visitor who asks for a person, the Tuesday page update that makes Monday's index wrong, and the stranger who copies your widget code onto their own site and spends your token budget.

What does a production AI support chatbot need?

A production support chatbot needs about a dozen components around the model call, and most of the effort goes into ingestion, the widget, the human handoff and the team inbox. The table lists each one with a rough DIY effort, assuming one experienced full-stack developer who has used an LLM API before and leans on managed services where possible.

Added up, the rows come to roughly 60 to 110 person-days before the first customer sees the bot, or three to five months for one developer. Your number will differ, but it is rarely the two weeks the prototype suggested.

Handoff and the team inbox are a second application

The rows people underestimate most are the human handoff and the team inbox, because together they are a second product. Human handoff is the moment the bot stops and a person takes over the same conversation. Doing it well takes statuses so the bot and the human never talk over each other, an alert that reaches someone, a live transcript for the operator, replies that appear under a person's name, and a rule for what the visitor sees when nobody is online. The human handoff guide lists the five triggers and the context the person must receive.

Ingestion is where answers go stale

Ingestion is the second trap. A crawler has to respect robots.txt, follow sitemaps, skip oversized pages and read scanned PDFs, then do it all again every week while re-indexing only what changed, or the bot keeps quoting last quarter's prices. Stale and contradictory pages produce confident wrong answers, which the guide to chatbot hallucinations traces back to their source.

A public endpoint needs security from day one

A support bot is an unauthenticated endpoint that spends money on every message. It needs rate limits per session and per account, an allowed-domains check so another site cannot embed it, defences against instructions planted in crawled pages, and, once it calls your APIs, protection against server-side request forgery and leaked credentials. The AI chatbot security guide covers each risk and the control that closes it.

How much does a DIY support chatbot cost to run?

The model bill for a support chatbot is small: about $3 a month at 2,000 conversations on gpt-4o-mini, and about $16 at 10,000. Infrastructure adds tens of dollars. The costs that decide the question sit elsewhere.

The arithmetic uses these assumptions, so you can swap in your own:

• Two AI replies per conversation: one answer and one follow-up.
• About 4,000 input tokens per reply: roughly 1,000 for the system rules and your instructions, 2,000 for five retrieved passages and 1,000 for the conversation so far.
• About 300 output tokens per reply.
• Standard prices from OpenAI's API pricing page as of October 2026, without the discount prompt caching gives on repeated input.

At 2,000 conversations that is 4,000 replies: 16 million input tokens and 1.2 million output tokens a month. On gpt-4o-mini, at $0.15 per million input tokens and $0.60 per million output tokens, the bill is $3.12. On the newer gpt-6-luna, at $0.10 and $0.50, it is $2.20. On a larger model such as gpt-6.1-sol, at $2 and $10, it is $44. At 10,000 conversations, multiply by five: about $16, $11 and $220.

Embeddings barely register. text-embedding-3-small costs $0.02 per million tokens, so embedding 4,000 questions costs well under a cent, and re-embedding a 500-page site every week costs a few cents a month. Labelling each conversation with a topic for your analytics adds pennies on a small model.

The vector store is cheap or free at this size. pgvector runs inside a Postgres database you may already operate. A managed service such as Pinecone lists a free Starter plan capped at 2 GB in a single AWS region, a $20-a-month Builder plan and a Standard plan with a $50 monthly minimum on its pricing page as of October 2026. Add hosting for the API and widget, email delivery for one-time codes and alerts, and monitoring; we assume $25 to $100 a month, and it can be close to zero if it rides on infrastructure you already pay for.

So the software itself costs somewhere between about $30 and $200 a month at 2,000 conversations, depending mostly on the model and whether you pay for a managed vector database. That is the number that makes building look cheap. It leaves out the person.

The bigger cost: engineering and maintenance time

Engineering time is the largest line in a DIY support chatbot, both to build it and to keep it working. At the US median software developer wage, the build costs about $31,000 to $58,000 and maintenance about $1,050 to $2,100 a month.

The US Bureau of Labor Statistics puts the median annual wage for software developers at $135,980 in May 2025. Spread over 260 weekdays, that is about $523 a day in salary alone, before benefits, payroll tax, equipment and the time of whoever manages the work. Substitute your own day rate if yours differs.

At $523 a day, the 60 to 110 person-days from the component table come to roughly $31,000 to $58,000. Then the bot has to be run. Pages change and syncs fail, someone has to read the escalated conversations and fix the content behind them, integrations break when a partner rotates a key, model versions are retired, and abuse arrives in waves. We assume two to four developer-days a month for all of that, or $1,050 to $2,100 a month. Skipping the work does not remove the cost; it moves it to customers, who get older and less accurate answers.

There is also a cost no invoice shows. The three to five months your developer spends on a support widget are months not spent on the product your customers pay for.

DIY vs hosted: what each costs at 2,000 and 10,000 conversations

Over a first year, a DIY support chatbot costs roughly $44,000 to $85,000 including the build, against $828 for a hosted flat plan at 2,000 conversations a month and about $2,750 to $3,000 at 10,000. The table shows where the difference comes from.

The hosted column uses PepoChat's published prices. Starter costs $49 a month and includes 2,000 AI replies; replies past the allowance cost $10 per 1,000. At two replies per conversation, 2,000 conversations is 4,000 replies, or $69 a month. At 10,000 conversations, Starter plus 18,000 overage replies is $229, and Growth, which raises the team to ten members and removes the PepoChat branding, is $249. The DIY column uses gpt-4o-mini and the assumptions above.

The model line, the one people worry about, is the smallest number in the DIY columns. The line that decides the comparison is maintenance, and it does not shrink with volume: a bot that handles 300 conversations a month needs its content synced and its escalations read just as much as one that handles 3,000.

Volume does eventually flip the result. With these assumptions, the hosted bill only reaches the DIY running cost somewhere between about 55,000 and 120,000 conversations a month, and that is before the build has been paid back. Far below that range, the case for building has to rest on something other than price.

The comparison changes if your alternative charges per resolution. Intercom lists its Fin AI agent at $0.99 per outcome as of October 2026. At 10,000 conversations with half resolved by the AI, that is $4,950 a month for the AI line alone, more than double the high end of the DIY running cost, and the build would pay for itself in roughly 8 to 21 months. The Intercom Fin pricing breakdown and the AI chatbot cost guide work through those bills in detail.

Why model and API deprecations are a hidden cost

When you build on a model API, every retirement on the vendor's schedule becomes a migration on yours. OpenAI's own deprecation history shows how often that happens.

According to OpenAI's deprecations page, developers using the Assistants API were notified on August 26, 2025 that it would be removed from the API a year later, on August 26, 2026, with the Responses API and Conversations API as the replacement. For early adopters it was the second forced move: access to the Assistants API's first beta had already been discontinued on December 18, 2024, after the second beta shipped. The same page lists more shutdowns announced in June 2026: reusable prompt objects, the Evals platform and Agent Builder on November 30, 2026, and older GPT-5 and o3 model snapshots on December 11, 2026.

This is not a reason to avoid OpenAI; any provider you build on can retire versions on its own schedule. It is a reason to budget two to five developer-days per migration and to keep an evaluation set: 50 or more real customer questions with the answers you expect, so that when you switch models you can see whether the bot got better or worse instead of hearing it from customers.

When should you build your own support chatbot?

Build when the assistant is part of your product rather than a support channel, or when a hard constraint rules out a hosted tool. Building pays back in control, not in savings, unless your volume is very high. If you want custom logic without writing the whole stack, agent builders sit in between; our Botpress vs Voiceflow comparison covers the two most common.

The agent is a feature, not a channel. An in-app assistant that changes settings, drafts content or explains a user's own data is product work. It needs your permission model and your interface, and no support widget will fit it as well as code you write.

You need data controls a vendor cannot give. If procurement requires a specific hosting region, a signed data processing agreement, a SOC 2 report or a model running in your own cloud account, check what a vendor actually provides before you assume. PepoChat, for example, does not currently claim a hosting region, a DPA template, or SOC 2 or ISO 27001 certification, so a buyer who needs those today should ask about current status or build. The GDPR guide for chatbots lists the questions to put to any vendor.

Your volume is very large. Past the crossover above, per-reply pricing adds up to more than an engineer's time, and a dedicated team can tune the cost of each conversation.

You already run LLM infrastructure. If your team maintains retrieval for another product, much of the ingestion, evaluation and deployment work already exists, so the table overstates your cost.

When is buying the better choice?

Buy when support is a channel you need working this month, the team is small, and nobody wants to own a second application.

A hosted tool turns the component table into configuration. In PepoChat, a site crawl, a sitemap import or a file upload builds the knowledge base, URL sources are re-fetched weekly with only changed pages re-indexed, and every reply follows the same route: retrieve, answer or act, and escalate when neither works, as How the agent answers describes. One script tag installs the widget. The team inbox, escalation emails, assignment and internal notes are already there, visitors can verify an email with a one-time code, and every feature is on the free plan, so you can test all of it before you pay.

Buying has real limits, and they deserve the same scrutiny as the build estimate:

• Channels. PepoChat's widget is web only. There is no WhatsApp, Messenger or Instagram channel today, and the optional voice assistant runs on your own Vapi account without sharing the chat agent's knowledge base or inbox.
• Actions. Integrations and custom actions run at least once, so a rare retry can repeat a write. That is fine for looking up an order and wrong for refunds, which should never be wired to a bot. There is also a cap of 20 actions per workspace.
• Analytics. You get satisfaction votes and auto-labelled topics, not resolution-rate charts; rates come from inbox counts or the CSV export.
• Models. PepoChat runs on OpenAI models. If your policy requires a different provider, that is a point for building.

If one of those limits blocks you, building or choosing a different vendor is the right call.

The hybrid: buy the agent, connect your own API

The middle path is to buy the support agent and expose your own business logic through an API it can call. You keep control of the data and the rules, and you skip building the widget, the inbox and the ingestion pipeline.

Most of what makes a support bot specific to your business is not the chat. It is the lookup: the order status in your database, the plan in your billing system, the booking in your scheduler. In PepoChat that lookup is a custom action, an HTTPS request the agent makes on a visitor's behalf. You set the method, a URL with placeholders in the path and query, headers whose values are encrypted and never shown again, and a JSON, form or XML body with the visitor's inputs escaped for that format. The agent collects up to 10 inputs from the visitor and reads only the response fields you list, up to 20. Private and internal hosts are blocked, redirects are not followed and calls time out after 10 seconds. The custom actions documentation has the full rules.

The split works because your API stays the source of truth. It checks permissions, applies your business rules and returns only what the visitor should see, while the bought agent handles conversation, retrieval, handoff and the inbox. If you later decide to build after all, the time was not wasted: the inbox's CSV export gives you real transcripts, up to 5,000 conversations per export, which make a better evaluation set than any you could invent.

What to do next

Run the cheapest experiment first: the weekend prototype or a free hosted workspace, pointed at the same content and asked the same 20 questions from the AI chatbot launch checklist. If the answers are wrong in both, fix the content before anything else. If the hosted tool answers well and nothing on the build list is a feature of your product, you have your decision. Compare plans on the pricing page, or create a free workspace and import your site as one knowledge source.