Skip to content

All articles

How much does a custom automation cost? What the technology costs, and where the rest comes from

A text message sent by software costs about one to two US cents, and AI is billed per million tokens. The cost of a custom system lies elsewhere: in the work to build it, connect it to your tools and keep it up to date.

What do a text message, a server and AI really cost?

Providers publish their prices. Here is what their official pages showed on November 8, 2026. All amounts are in US dollars.

  • A text message in Canada. At Twilio, an outbound text costs US$0.0083 per segment, plus carrier fees of about US$0.0067 to US$0.0087 depending on the network. In total, about one to two US cents per text. A local number rents for US$1.15 per month.
  • A server. The smallest AWS Lightsail server costs US$5 per month. That is a floor. A system in production often adds backups, a database and monitoring.
  • AI. The Claude API is billed by usage, per million tokens. The Haiku 5.5 model costs US$0.10 for input and US$0.50 for output (requests of 100,000 tokens or less). Sonnet 5.5 costs US$2 and US$10. Opus 5.5 costs US$4 and US$20.

These prices change. Twilio itself states that its prices may change without notice. Taken one by one, these amounts look small. What makes the monthly bill is volume, and it is estimated project by project.

What is a token?

A token is the unit an AI model uses to read and write text. It is often a piece of a word, a short word or a punctuation mark.

Input is everything sent to the model: the request, the instructions, the documents to read. Output is what it answers. At Anthropic, for the three models listed, output costs five times more than input.

The monthly cost of AI therefore depends on volume: how many requests per month, how long, and with which model. Without that data, no honest amount is possible.

So where does the cost of a custom system come from?

From the work. Before the first line of code, you have to understand how the business really runs. What makes the cost to build vary:

  • the number of tools to connect together;
  • the state of your data;
  • what has to be built rather than connected;
  • the exceptions to plan for;
  • what must stay approved by a person.

Then the system runs every month. Are there monthly fees? Often, yes. They can be hosting, APIs (the connections between two pieces of software), AI usage depending on volume, texts and calls, or a subscription to a tool. The amount varies by project.

Finally, it has to be kept up to date. Your ways of working change. The connected tools also change on their side, and new features are added over time. A custom system lives with the business.

Why don't shortcuts really work?

Since AI costs pennies, it is tempting to plug in a quick tool and move on. The problem comes with the first exception: the client who cancels at the last minute, the invoice paid in two instalments, the request that fits no box. A system built without knowing them handles them poorly, and someone ends up redoing the work by hand.

IT budget overruns are well documented. Across 1,471 projects studied by Flyvbjerg and Budzier, the average overrun was 27%, and one project in six went over budget by 200% on average. These were mostly large projects, often public, far from the reality of a small business. The lesson still holds: what derails a budget is what was not planned for, rarely the server bill.

Why do context and memory matter more than the tool?

AI models are billed at the public prices posted by their providers. A small business has access to the same model as a multinational. What makes the difference is what the system knows about your business: your clients, your exceptions, how decisions are made, what happened last month.

That context builds up over time. Every month of use adds clean data and ways of working written down in black and white. The earlier you start, the stronger the system.

What support exists in 2026?

In APC.lab's conversations with business owners, the same observation keeps coming up: many are behind on automating their processes, often for lack of time and knowledge of artificial intelligence. Lack of funds plays a part too: a big agency is extremely expensive, and those agencies are mostly interested in large mandates.

Cost holds back many businesses. In Quebec, among those that had not invested in digital infrastructure in three years, high cost is the first obstacle named (27.2%), ahead of uncertain return (19.3%) and lack of knowledge (14.3%). These figures cover digital technology in general, not AI. For AI, in Canada, cost ranks second among the obstacles cited by surveyed businesses (10.6%), after cybersecurity and privacy (13.4%). Two in five surveyed businesses consider AI not relevant to them.

A few programs exist. They have conditions, to read closely.

  • ESSOR, components 1B and 1C (Investissement Québec). Up to 50% of a digital diagnostic and plan (maximum $20,000), then up to 50% of its implementation (maximum $50,000). It targets businesses with 250 employees or fewer and at least $2.5M in revenue, which excludes many small businesses. The program runs until March 31, 2027. The page does not say whether applications are open today.
  • LIFT (BDC). Financing for AI adoption, from $1M in annual revenue for the AI stream. A plan is required. It is a loan: it helps cash flow, it does not reduce the cost.
  • Offensive de transformation numérique (Quebec). It no longer accepts new projects.
  • SR&ED tax credit (federal). Work that gets around a problem with knowledge that is already available is not eligible. Connecting existing tools therefore generally is not. Since 2026, however, a business with gross revenue under $25M can ask the CRA before starting whether its project is eligible. The answer comes within 8 weeks.

Under $1M in revenue, neither ESSOR nor LIFT's AI stream applies.

This information is for general purposes. It is not tax advice. Before counting on a grant or a tax credit, check with your accountant.

How can you find out what your project would cost?

APC.lab publishes no general price, because two projects that look alike on paper can vary a lot depending on the tools, the data and the exceptions. To find out the cost of yours, describe the main lines of your project or the problem you would like to solve. A free call lets you discuss it in more detail. Then come a feasibility study, a proposal, and the work with you until the exact need is met. Depending on the contract, the code can belong to you or stay with APC.lab. For one-off needs and small integrations, the rate is $150 per hour plus taxes.

Sources

Talk about your project

Have a project in mind? APC.lab can help you.

Describe what you need