A chatbot that answers from your documents runs $8,000–$15,000. Retrieval over live private data — actual RAG — runs $15,000–$40,000. An agent that owns a real workflow end to end runs $40,000–$80,000. The spread inside each band is integrations and evaluation, almost never the model.
Quotes for "an AI chatbot" vary by 10x because the phrase covers three structurally different builds. Founders comparing a $6,000 quote against a $60,000 quote are usually comparing a wrapper against a system. Here is how we scope the three tiers and what moves the number inside each.
Tier one: the docs-grounded chatbot — $8,000–$15,000
A bot that answers questions from a fixed body of content: your help articles, policies, product docs. Ingestion, chunking, a retrieval layer, a refuse-to-answer threshold so it says "I don't know" instead of inventing, and a simple review loop.
This tier sits at or below our usual minimum, and we'll say so on the call. If your content is clean and the use case is support deflection, a good template or a competent freelancer can get you most of the way. Where it earns a studio build: messy content, compliance constraints, or a hard requirement that wrong answers never ship.
Tier two: RAG over live private data — $15,000–$40,000
Retrieval-augmented generation over data that changes: a product catalog, a document pipeline, customer records. Now the build includes sync from your systems of record, permissioning (who is allowed to retrieve what), an evaluation set that catches quality regressions, and monitoring.
The number moves with data messiness more than data volume. One clean Postgres source sits at the bottom of the band. Four sources, two of them PDFs with tables, sits at the top. For the LaunchProd build we ran a section-level retrieval pipeline on pgvector with a refuse-to-answer threshold — and the most valuable engineering decision was removing complexity: cutting multi-tier chunking and reranking dropped latency roughly 35 percent with no quality loss. You are paying for that judgment, not for the vector database.
Tier three: the workflow agent — $40,000–$80,000
An agent that executes multi-step work — triages the ticket, drafts the response, updates the CRM, escalates the exception — rather than chatting about it. The model is one component. The build is the state machine around it: tool integrations, failure handling, human approval gates, an audit trail, and an evaluation harness that measures task completion, not vibes.
The upper half of this band is where our usual $15,000–$60,000 range tops out; past that we split the build into scoped phases so each milestone ships something operable on its own.
What moves the number
- Integration count. Every system the AI reads from or writes to adds a connector, auth, and failure modes. This is the single biggest cost driver across all three tiers.
- Evaluation infrastructure. Two to four engineering weeks to build the harness that tells you whether outputs are correct. Skipping it is how AI features quietly degrade after launch — we build it first, and we put it in the quote as a line item.
- The escape hatch. Anything customer-facing needs a clean handoff to a human. Designing that path well is product work, not prompt work.
- Volume economics. Inference runs $0.50–$5 per thousand calls depending on model and prompt length. At 50,000 monthly users that's $25–$250 a month — fine. At millions of calls, model choice and caching become architecture decisions, and the build reflects it.
What doesn't move the number
Model choice, mostly — providers are swappable behind an interface, and we pick per evaluation results, not brand. UI polish matters less than founders expect; the chat interface is the cheap part. And "GPT-5 versus everyone else" debates move the quote by approximately zero dollars.
Red flags in the quotes you're comparing
A $3,000 "AI agent" is a wrapper — one API call in a trench coat. A quote with no evaluation line item is a demo with an invoice. Per-seat pricing on a custom build means you're renting your own product. And any proposal that can't tell you what happens when the model is wrong hasn't thought about the only question that matters in production.
Ask each bidder one question: "how will we know the outputs are correct?" The quality of the answer prices the quote for you.