How Much Does an AI Chatbot Cost? A Builder's Real Numbers

Every pricing page for this stuff is either "contact sales" or a number that has nothing to do with what you will actually pay. I build and operate these systems for small businesses, so what follows is the cost structure as I see it from the invoice side: what the platforms bill me, what I bill clients, and where the money quietly leaks.

The short version: a working AI chatbot for a small business lands somewhere between $0 and $500 a month depending on whether you are doing it yourself, buying a product, or hiring someone to run it. The spread is enormous because "chatbot" covers four very different things with four very different cost curves.

The four cost buckets

Almost every quote you receive decomposes into these:

1. Build cost. One-time. Someone designs the conversation, writes the prompt, loads the knowledge base, wires the integrations, tests it.
2. Platform cost. Recurring. A seat, a plan tier, or a per-message allowance from whatever tool hosts the agent.
3. Usage cost. Variable. Model tokens, voice minutes, telephony minutes, SMS segments, transcription.
4. Operating cost. Recurring and invisible in most quotes. Someone reads transcripts, fixes wrong answers, updates hours and pricing, and gets paged when it breaks.

Bucket 4 is the one that kills DIY projects. A chatbot is not a website. It drifts. Your prices change, you add a service, a model provider deprecates something, and nobody notices for six weeks that the bot has been quoting last year's rates.

DIY: $0 to $80 a month, plus your weekends

If you are technical and patient, you can genuinely run a decent chat widget on a hobby budget. As of mid-2026 the pattern that works is: a low-cost hosted chatbot product for the widget and knowledge base, a model API key, and your own time.

What you actually pay:

Line item	Typical shape	Realistic small-business range
Hosted chat widget product	Per-message allowance or seat, monthly	$0 to $50
Model API (text chat)	Per token, both directions	$2 to $20
Site hosting for the widget	Usually already paid for	$0
Your time, first build	10 to 30 hours	The real cost

The token bill for text chat is smaller than people expect. A typical support conversation on a small site is a handful of turns with a few thousand tokens of retrieved context each. At mid-2026 prices for a mid-tier model, hundreds of conversations a month is single-digit dollars. If your model bill for a text chatbot is over $50 a month, you are almost certainly stuffing your entire knowledge base into every prompt instead of retrieving the relevant chunk.

DIY breaks down on two things. First, voice. Once you add phone, the plumbing (numbers, SIP, forwarding, transfer logic) turns into a real project. Second, maintenance. If you cannot commit to reading transcripts weekly, you will end up with a bot that confidently makes things up and you will not find out from a customer, you will find out from a review.

Platform: $50 to $500 a month

This is buying a product rather than assembling one. You pick a platform, do the setup in a visual builder, and pay a plan fee plus usage. As of mid-2026 the pricing models cluster into three shapes:

• Per-seat plus message allowance. Common in support-desk style products. Costs scale with your team size and ticket volume, not really with bot quality.
• Per-conversation or per-resolution. You pay when the bot handles something. Predictable per unit, brutal if you get a traffic spike.
• Per-minute. Standard for voice agents. You pay for connected talk time.

Voice per-minute pricing is the number most people underestimate and then overestimate. As of mid-2026, all-in conversational voice on the main agent platforms runs in the range of a few cents to roughly a dime and a half per minute depending on tier and model, and telephony (the actual phone number and carrier minutes) adds about a cent per minute on top. Do the arithmetic before you panic: a business taking 200 calls a month at 3 minutes average is 600 minutes. At the expensive end that is under $100. At the cheap end it is a nice lunch.

The cost anchor I trust most from my own stack: I run an OpenAI Realtime plus Twilio voice agent for an energy company where the fixed infrastructure is about $14 a month plus usage. Fourteen dollars. That agent replaced an older speech-stack setup that cost more and did less. The lesson is that the recurring platform floor for voice is low, and almost all real cost is minutes and labor.

For a shared telephony account across several small clients, my Twilio bill runs around $130 a month total. Split across those businesses, the phone plumbing is a rounding error compared to what any of them pay a human to answer calls.

If you are choosing between hosting it yourself and buying the managed version, build vs buy goes through the tradeoff in more detail.

Agency: setup fee plus retainer

What I see across this market, including my own pricing, is consistent: a setup fee in the hundreds, and a monthly retainer from about $99 up to a few hundred for a small business. Larger builds with real system integration go higher, but for a lawn-care company, a clinic, or a remodeler, that is the band.

What the setup fee buys, when it is honest work:

• Conversation design and the system prompt, including an explicit list of things the agent must never claim it can do
• Knowledge base construction from your real pages, pricing, and policies
• Phone number provisioning and forwarding from your existing system
• Booking, CRM, or email integrations
• Real test calls and texts from real handsets, not just simulator runs
• A handoff path to a human that actually works

What the retainer buys: transcript review, prompt fixes, hours and price updates, model and platform migrations when the vendor changes something, and someone to call at 8am when the phone is answering with dead air.

That last one is not hypothetical. I have shipped an assistant whose configured dynamic variables were empty, and the result was total silence when it picked up. It looked exactly like a carrier fault. It was a config bug, and it took experience with that specific failure to find it fast. I wrote up that class of problem in voice agent dead air debugging. Paying a retainer is largely paying for someone who has already made those mistakes on somebody else's dime.

If you are weighing doing it in-house versus hiring out, DIY vs agency is the direct comparison.

Cost by channel

The channel drives the bill far more than the platform does.

Web chat

Cheapest by a wide margin. No telephony, no speech models, text tokens only. A busy small-business site running a few hundred conversations a month costs single to low double digits in model usage. The cost here is design time, not runtime.

SMS

Two components: the carrier fee per message segment (fractions of a cent to a couple of cents in the US, as of mid-2026) plus messaging service and campaign registration fees. Registration is the surprise. US A2P registration involves one-time and recurring fees per brand and campaign, and it is not optional. Budget for it, and budget time for it, because approval is not instant.

SMS also carries compliance cost that is not measured in dollars until it is. I built an SMS assistant inside a sports-facility company's CRM with a hard consent gate: the bot physically cannot text a contact who has not opted in, no exceptions, no override. That gate is the cheapest insurance in the entire system. See TCPA compliance for AI SMS.

Voice

The expensive channel, and the one worth paying for. Per-minute agent cost plus per-minute telephony plus the number rental (about a dollar a month for a local number). The real variable is call length. A well-designed agent that qualifies and books in 90 seconds costs half what a chatty one costs, and converts better. Prompt tightening is a cost-reduction lever, not just a quality lever.

What actually blows up the bill

In order of how often I see it:

1. Long calls. An agent with no ending discipline will happily talk for eight minutes. Give it explicit wrap-up behavior.
2. Whole-document context. Retrieval exists so you do not pay to resend your FAQ on every turn.
3. Retries and loops. A broken transfer that makes the agent retry three times bills three times. Transfer bugs are cost bugs.
4. Unmonitored spam. Bots and scrapers hitting a public chat widget. Rate limit it.
5. Paying for seats you do not use. Support-desk pricing models charge for humans. If your bot is deflecting, you may be able to shrink the human plan.

A sane budget for a small business

If you want one number to plan against, here is how I would budget a first year for a service business that wants both a site widget and an AI answering the phone:

• One-time build: several hundred to low four figures depending on integrations
• Monthly platform and retainer: $99 to $400
• Monthly usage at typical small-business volume: $20 to $120
• Contingency for the first 60 days: assume you will change the scope once

Then measure it against the thing it replaces. A part-time person answering phones costs more than all of that in a week. The honest comparison is not "chatbot versus nothing," it is "chatbot versus the calls you are currently missing." Which is exactly why missed-call text-back tends to pay for itself faster than anything else on this list, and why some businesses should not buy any of it at all, which I cover in when not to use a chatbot.

FAQ

Is a per-minute or per-conversation price better?

Per-minute is more honest for voice because it tracks the underlying cost, and it rewards you for building a tighter agent. Per-conversation is easier to forecast but you pay the same for a 20-second wrong number as for a booked job. If your call mix is heavy on quick hangups, per-minute usually wins.

Why do agency quotes vary so much for the same thing?

Because "the same thing" usually is not. A quote that includes phone forwarding, booking integration, real test calls, and monthly transcript review is a different product from a quote that includes a prompt and a widget embed. Ask specifically who fixes it when it answers wrong, and how fast.

Can I start cheap and upgrade later?

Yes, and I would. Start with web chat on a low-cost platform, get the knowledge base and the prompt right where mistakes are cheap and invisible, then move to voice once the content is proven. The conversation design work transfers; the plumbing does not.

What is the hidden cost nobody quotes?

Testing and monitoring. A simulator cannot inject things like caller ID, so a class of bugs only appears on real calls. Every rollout I do includes calling and texting my own cell first. Budget for that, and for someone reading transcripts weekly, or you are buying a system that degrades silently.

Does the model I choose change the bill much?

For text chat, less than you think, because volumes are small. For voice, yes, because you are billed per minute of a speech-capable model and the tiers differ meaningfully. Pick on latency and interruption handling first, then price, because a laggy agent loses callers and that costs more than the model ever will.