What it costs to run an AI product each month

The three costs that actually matter

When you build an AI product—whether it's a tool that reads invoices for your accounting firm, suggests replies for your service team, or searches customer documents—you'll pay for three things: the AI service itself, the infrastructure to run it, and the labour to keep it working.

The AI service is usually the biggest line item. Every time a customer uses your tool, you send a request to an AI company like OpenAI, Anthropic, or Google. They charge you for that request. The cost per request is tiny—often between $0.001 and $0.10—but it adds up fast if thousands of customers use your product daily.

Infrastructure is what it costs to run the servers, databases, and connections that hold your product together. For most small products, this runs $50 to $500 a month. A property manager in Surrey with a tool that reads lease documents might spend $100 monthly on infrastructure. A distributor in Calgary with a product that searches inventory across ten locations might spend $300.

Labour is the cost of someone checking on the product, fixing bugs, and answering customer questions. You might do this yourself at first, or hire someone part-time. Budget 5 to 15 hours a week, depending on how many customers you have and how complex the tool is.

What a real product costs

Let's walk through three examples.

A Burnaby accounting firm builds a tool that reads invoices and files them automatically. The firm has 40 clients. Each client uploads 10 to 20 invoices a month. That's 400 to 800 API calls monthly. At $0.02 per call, that's $8 to $16 in AI costs. Infrastructure is $150 a month. A bookkeeper spends 4 hours a week checking the results and fixing errors. At $30 an hour, that's $480 a month. Total: about $650 a month. The firm charges clients $50 a month for the service, so with 30 active users, they bring in $1,500 monthly and break even after 13 months.

A 12-person software company in Gastown adds a feature that writes first drafts of support replies. Their 200 customers send 2,000 support emails a month. The tool drafts a reply for each one. At $0.01 per call, that's $20 in AI costs. Infrastructure is $200. One engineer spends 6 hours a week monitoring quality and retraining the tool. At $80 an hour, that's $480 a month. Total: about $700 a month. They charge $10 per month per customer for the feature, so 100 customers use it. Revenue is $1,000 monthly. They're cash-positive after two months.

A plumber in Richmond builds a tool that answers customer texts about booking appointments. He has 300 customers on his text list. About 100 texts come in per week asking about availability. The tool answers 70 of them without human help. That's 280 API calls a month at $0.015 per call, or $4.20. Infrastructure is $80. The plumber spends 3 hours a week reviewing conversations and handling complex bookings. At $40 an hour, that's $120 a month. Total: about $204 a month. He doesn't charge for this feature—it's a competitive advantage. The time savings alone (10 hours a month answering routine texts) is worth $400 to him.

How to keep costs predictable

The first rule is to track usage. Before you launch, build a simple counter that logs every API call. You'll see exactly what's costing money. A dental clinic in Kelowna built a tool that summarizes patient notes. After one month, they saw that 60% of their calls were wasted on empty or duplicate notes. They fixed the data pipeline and cut their bill by 40%.

The second rule is to choose the right AI service for your use case. OpenAI's GPT-4 costs more per call but makes fewer mistakes. Google's Gemini costs less but is slower. For a task like summarizing a patient note or drafting a reply, the cheaper option often works fine. For a task like reading a legal contract or making a recommendation that affects a customer's safety, you need the better model. Your choice here determines whether you spend $200 or $2,000 a month.

The third rule is to set a monthly budget and stick to it. Tell your AI service provider what you're willing to spend. Most offer alerts when you're approaching your limit. This prevents a $10,000 surprise bill if something goes wrong.

The fourth rule is to batch work when you can. Instead of answering one customer question at a time, collect ten and send them together. Instead of reading one invoice, read a folder of them. Batch processing is cheaper and faster.

Where to start

• Pick one small process to automate. Choose something you do 10 to 50 times a month, not something you do once a year or 1,000 times a day. This gives you a manageable bill and clear feedback on whether the tool works.

• Estimate your API calls for three months. Count how many times per month the process happens, then multiply by three. Call your chosen AI service provider and ask what that volume would cost. Write the number down. If it's more than you want to spend, pick a cheaper model or a simpler process.

• Build usage tracking before you launch. Add a simple log that counts API calls, errors, and time taken. You'll use this to spot problems and find savings. A spreadsheet is fine for the first month.