What an AI automation project actually costs

Quotes for AI automation are hard to compare because they bundle very different costs together. Some are paid once, some every month, and some grow with every message or minute. Separating them is the single most useful thing you can do before you sign anything. This guide lists the categories, shows published prices as of 21 October 2026 where they exist, and gives you questions to ask about the rest.

All prices below are in US dollars, are list prices published by the providers named, and were read on 21 October 2026. They change, differ by country and may differ again when a platform resells them. Use them to understand how costs are structured, not as a quote.

The five cost categories

Setup and build

This is the work of mapping your workflow, writing the conversation rules, connecting your calendar and CRM, and testing. It varies more than any other category, which is why there is no meaningful "typical" figure. A single acknowledgement-and-qualify workflow on one channel is a small job. Voice, several channels, custom integrations and approval steps each add to it. Ask for the scope in writing, including what testing is included and how many rounds of changes.

Software platform

Many automations run on an all-in-one platform. As one published example, HighLevel lists its Agency Starter plan at US$97 a month and Agency Unlimited at US$297 a month; the difference, according to its pricing FAQ, is the number of sub-accounts. Agencies often provide access through their own account instead, so the subscription may appear inside a monthly fee. Either way, find out who owns the account and your data.

Messaging and telephony

These are usage costs: they rise and fall with how many people contact you. Published US list prices from Twilio, as an example of the structure:

Two points catch people out. Platforms that resell phone service set their own rates, which can differ from the provider's list price. And rates differ by country, so a Canadian number is priced separately from a US one.

Model usage

Language models are billed per million tokens, with separate prices for what the model reads (input) and writes (output). Prices vary widely by model. On 21 October 2026 Anthropic listed input prices from US$0.10 per million tokens for Claude Haiku 5.5 to US$10 for Claude Fable 5.1. Other providers bill the same way; check their current price pages before you compare.

A worked example shows why model choice matters more than any other setting. Assume a text conversation where the model reads a combined 30,000 tokens over all its turns (the rules and history are re-sent each turn) and writes 2,000. These token counts are assumptions for illustration.

The same workload differs by a factor of a hundred depending on the model. Most enquiry conversations do not need the largest model, but some steps, such as summarising a complicated call, may benefit from a stronger one. Voice adds speech recognition and speech generation, which some platforms bill per minute on top of the call itself.

Support and changes

Your hours, prices, services and staff will change, and each change needs updating somewhere. Ask how changes are requested, how fast they are made, who monitors failures, and what happens if the provider is unavailable. A clear answer here is worth more than a lower setup price.

How to compare two quotes

• Put each quote into the five categories above, with one-off and monthly costs in separate columns.
• Estimate your monthly volume from real records: enquiries, texts sent, call minutes.
• Ask whether usage costs are passed through at cost, marked up, or bundled with a cap.
• Check what is included after launch, and for how long.
• Confirm the currency on every line. A mix of USD platform fees and CAD invoices is common.
• Ask what it costs to leave: exporting data, moving numbers, handing over workflows.

Limits

• Prices are the providers' published list prices on one date. Check them again before relying on them.
• The worked example uses assumed token counts. Ask your provider for measured usage from a test period.
• Taxes, currency conversion and payment fees are not included.
• This guide describes cost structure only. It is not a quote, and Manyhats prices are set per project after scoping.