How much does it cost to build a custom AI sales agent?

A custom AI sales agent costs two things: a one-off build (US$40,000 in our labelled assumption) and a per-conversation meter that, at vendor list prices read on 15 October 2026, runs about US$0.31 for a four-minute US voice call, US$0.20 for an SMS thread and US$0.10 for a web chat. LeadsNow sees booked-call costs of $30–$400, depending on industry, offer, price and many other variables.

• The question this page answers: “how much will it cost me to build my own AI sales agent, and when is it cheaper than buying one?”
• The meter: telephony per minute, a voice session per minute, and language-model tokens. Prices from OpenAI, Anthropic and Twilio’s own pricing pages.
• The fixed cost: build plus upkeep. At US$40,000 over 12 months plus US$3,000 a month upkeep (both assumptions), US$6,333 a month before the first conversation.
• The decision rule: the break-even conversation count. Fixed monthly cost ÷ (what buying costs per conversation − what your own meter costs). At US$2 bought, building pays above 3,739 conversations a month.
• Not on this page: agency line items in Australian dollars, or a whole-company AI budget. This page is one sales agent, priced per conversation.

What does each conversation cost a custom AI sales agent to run?

A custom AI sales agent’s variable cost per conversation ranges from US$0.005 for a web chat on a small model to US$0.55 for a four-minute call to an Australian mobile.

Each row adds the vendor list prices for one typical conversation. Conversation length and token counts are assumptions; carrier surcharges are not included. OpenAI’s gpt-6.1-sol and Anthropic’s Claude Sonnet 5.5 both list US$2.00 per million input tokens and US$10.00 per million output tokens, so the mid-size model column holds for either.

Sources, all read 15 October 2026: OpenAI API pricing (gpt-live-1 at US$0.05 a minute, with backend model use charged separately), Anthropic’s Claude pricing, and Twilio’s US voice, Australian voice and US SMS pages. The quotable version: on a custom AI sales agent, the voice minute costs about five times the thinking.

The break-even conversation count: when does building beat buying?

The break-even conversation count is fixed monthly cost divided by the gap between a bought conversation’s price and your own meter; at US$2 a bought conversation, it is 3,739.

The fixed side uses our assumptions: a US$40,000 build spread over 12 months (US$3,333.33) plus US$3,000 a month of upkeep for prompt changes, integration fixes and transcript review, so US$6,333.33 a month. The variable side is the US$0.306 US voice conversation above. The “bought” prices are assumption bands for what a vendor or platform charges you, all-in, per conversation; replace them with your quotes.

Two cautions. The count assumes the agent you build performs as well as the one you would buy; a weaker agent burns leads, and that cost is not on the meter. And it assumes upkeep stays at US$3,000 a month after year one, which holds only if someone owns the agent. The capacity gates on our build vs buy decision for AI sales systems decide whether you can build at all; this count decides whether it pays.

Worked calculation: what does a booked call cost from your own AI sales agent?

A self-built AI sales agent’s cost per booked call falls from US$86.48 at 500 voice conversations a month to US$10.48 at 5,000, at an assumed 15% booking rate.

What this leaves out, honestly: the cost of the leads themselves, any human who takes the hand-off, and the leads a poorly tuned agent loses in its first months. Add your lead cost before comparing the result with any quoted cost per booked call. Our breakdown of US AI voice agent per-minute prices shows how failed dials and voicemail inflate the voice meter beyond this clean four-minute call.

What goes into the build of a custom AI sales agent?

A custom AI sales agent build covers six sales-specific components, and the evaluation set is the one that keeps costing money after launch.

• Qualification logic: the questions, the order, and the rule for when a lead is ready to book.
• Objection and question library: pricing, timing, “send me info”, competitor mentions, with approved answers.
• Booking: live calendar reads, time zones, rescheduling, confirmations and reminders.
• CRM write-back: outcome, transcript and next step on the right record, so a rep never re-asks a question.
• Voice tuning: latency, interruptions, voicemail detection and a clean transfer to a human.
• Evaluation set: a few hundred scripted test conversations re-run after every prompt or model change.

Our US$40,000 assumption is 250 hours at US$160 an hour across those six; a chat-only agent with no voice tuning can sit well below it. Model versions change under a custom agent, so the evaluation set is a running cost, not a launch task.

Does the choice of model change what an AI sales agent costs?

Model choice changes a chat-based AI sales agent’s meter twentyfold but a voice agent’s by only about 16%, because telephony and the voice session dominate calls.

The small models on both vendors’ pages, OpenAI’s gpt-6-luna and Anthropic’s Claude Haiku 5.5 (for prompts up to 100,000 tokens), list US$0.10 per million input tokens and US$0.50 per million output tokens. On the web chat row, that cuts US$0.10 to US$0.005. On the US voice row it cuts US$0.306 to US$0.2585, because US$0.256 of the call is telephony and voice session. Pick the model on conversation quality for voice, and on cost for high-volume chat.

Build, use a platform, or buy booked calls: who is each option wrong for?

Three ways exist to get a working AI sales agent, and each is wrong for a specific buyer, regardless of what it costs.

LeadsNow builds and runs AI sales agents across calls, SMS and DMs as part of its AI appointment setting service; no-shows aren’t charged. If your volume is under the break-even count, the cheapest agent is often the one someone else already maintains.

Frequently asked questions

How much does it cost to build a custom AI sales agent?

In our labelled assumption, about US$40,000 to build (250 hours at US$160 an hour) plus about US$3,000 a month to maintain, then a per-conversation meter of roughly US$0.31 for a four-minute US call, US$0.20 for an SMS thread and US$0.10 for a web chat at 15 October 2026 list prices.

Why is my SMS bill higher than the per-message price?

Twilio’s US SMS pricing page lists US$0.0083 per segment but states that messages are charged per segment, that additional carrier fees may apply, and that US A2P 10DLC is subject to registration onboarding fees. A long AI-written reply can be several segments.

Is a voice AI sales agent more expensive than an SMS one?

Per conversation, yes: about US$0.31 for a four-minute US call against US$0.20 for an eight-message SMS thread in our model. Per booked call it can go either way, because the booking rate of each channel matters more than the meter.

How many conversations a month before building my own AI sales agent pays?

Divide your fixed monthly cost by the gap between a bought conversation and your own meter. With US$6,333 a month fixed and US$0.306 per call, that is about 3,739 conversations at US$2 bought, or 1,715 at US$4.

Does a cheaper model make an AI sales agent much cheaper?

For chat, yes: a small model cuts a US$0.10 chat to US$0.005. For voice, only about 16%, from US$0.306 to US$0.2585, because telephony and the voice session make up US$0.256 of the call.