How to budget for an AI customer support bot

Every AI support bot has three cost lines. Most agencies quote you one of them.

Line 1: Build

This is the one-time fee to design the conversation flow, connect to your knowledge base or documentation, wire up the widget on your site, and test it. For a focused FAQ bot — one product, one team, defined edge cases — this is 10–20 hours of work. At our rates, that's in the $1,500–$3,500 range. A more complex bot (order lookups, CRM integration, handoff to live agent) runs $5,000–$12,000. The build fee is real, but it's not where things go wrong.

Line 2: Monthly API cost

This is the line agencies either don't mention or bury in footnotes.

Here's the actual math for a support bot doing 1,000 conversations per month at roughly 3,000 tokens per conversation (system prompt + conversation history + user message + response):

• •Claude Haiku: $0.80 per million input tokens, $4 per million output tokens. At 3K tokens average, assuming 2,400 in / 600 out: ~$1.92 input + $1.44 output = $3.36/month per 1,000 conversations.
• •Claude Sonnet: $3 per million input tokens, $15 per million output tokens. Same token split: ~$7.20 input + $5.40 output = $12.60/month per 1,000 conversations.

That's the favorable end of the range. Real conversations are messier. Longer context windows, RAG retrieval chunks added to the prompt, multi-turn conversations with full history — a realistic Sonnet deployment doing 1,000 conversations/month often lands $25–$60/month in API spend, sometimes higher.

At 5,000 conversations/month, you're looking at $125–$300/month on Sonnet. At 15,000 conversations/month, it's a real budget line.

Haiku keeps costs low but handles nuance worse. Sonnet handles nuance well but costs roughly 4x more per conversation. The right answer depends on how often customers ask edge-case questions the bot needs to think through. We model this in discovery before recommending either.

Line 3: Care (your monthly maintenance)

A support bot is not a set-and-forget deployment. Knowledge bases change. Products get updated. New edge cases appear. Customers find failure modes the test suite didn't cover. Something drifts.

The question isn't whether the bot needs ongoing attention — it does. The question is who's paying for it and whether that's explicit.

The "support retainer" trap: agencies bundle their time into a "$200/mo retainer" that sounds like it covers everything. Read the contract. In most cases, that covers their availability to respond to tickets, not the API spend, not database hosting, not any proactive monitoring. The API bill still lands on your card.

We separate these explicitly. Our Care tier is our time — $199/month for a bot in the Essential tier, $399 for a production workflow getting regular updates. API spend is always pass-through: you use your own keys, or we manage and bill at cost. There's no markup hidden in a bundled retainer.

What to ask before you sign

1. What model does the build use, and why? 2. What's the per-conversation token estimate at my expected volume? 3. Who pays the API bill after launch, and how? 4. What does "ongoing support" actually cover — line by line?

If the agency can't answer all four, they've either built the bot in a cost-blind way or they're planning to absorb a surprise and bill you for it later.

Run the numbers for your volume in 90 seconds at our cost estimator. Bring the output to whatever vendor you're evaluating.