What AI automation actually costs in Canadian dollars

There are two lies told about the cost of AI automation, and they point in opposite directions. The first is that it's basically free now — tokens are cheap, look how cheap. The second is that it's an enterprise line item requiring a six-figure budget. Both survive because almost nobody publishes the arithmetic.

So here is the arithmetic, for a workload we see constantly in the North American mid-market: 5,000 documents a month — invoices, purchase orders, claims, or intake forms — extracted, validated against business rules, and written into a system of record, with a human reviewing the exceptions.

Step one: the model call

A single document of the kind above runs roughly 3,000 input tokens (the document text plus your extraction instructions) and about 800 output tokens (structured fields coming back). Call it 3,800 tokens per document, and assume a realistic 15% retry rate for malformed output, timeouts, and the genuinely weird documents.

• 5,000 documents × 1.15 retries = 5,750 model calls/month
• Input: 5,750 × 3,000 = 17.25M input tokens
• Output: 5,750 × 800 = 4.6M output tokens

At current frontier-model API pricing, that lands in the low hundreds of dollars per month. Not thousands. On a mid-tier model — which is usually the right choice for structured extraction — it can fall under a hundred.

Two levers move this number more than model choice does: prompt caching, where the repeated instruction block is billed at a steep discount rather than full price on every call, and model routing, where the easy 80% of documents go to a cheap model and only the hard ones escalate.

One of these is a published number and two are not, so we will separate them. Provider pricing bills cached input reads at a fraction of the base rate — up to 90% off repeated input tokens. Routing and prompt tightening compound on top, but their size is a property of your own traffic mix, not an industry constant, so we will not quote you a percentage for them. Which is the real headline here: a badly engineered pipeline can cost five to ten times a well-engineered one for identical output. The bill is an engineering artifact, not a market price.

Step two: the parts nobody quotes

Now the line items that don't appear in anyone's "AI is cheap" post.

Infrastructure lands somewhere around $200–600/month for a workload this size. Still not the expensive part.

Step three: the number that decides everything

The exception queue is the real cost centre, and it's a function of accuracy.

An Ontario document clerk at a $52,000–58,000 base costs roughly $65,000–72,500 a year fully loaded once you add CPP, EI, Employer Health Tax, WSIB, benefits and paid vacation. Over about 1,800 working hours, that's $36–40 per hour.

Now run the accuracy scenarios on 5,000 documents, assuming an exception takes 4 minutes to resolve:

Read that table again, because it's the whole argument. Moving from 80% to 95% accuracy saves about $1,950 a month — roughly five times the entire model API bill.

This is why we sell an evaluation harness rather than a demo. The accuracy number isn't a bragging right; it's the largest variable line in the operating cost. And it's why a vendor's self-graded accuracy score should worry you: if the number is wrong by ten points, your real monthly cost is wrong by four figures.

So what does the whole thing cost?

For 5,000 documents a month at a realistic 95% accuracy, running in your chosen region:

• Build: a one-time engineering cost — ours starts at CAD $26,000 for a single document type
• Run: roughly $1,000–1,700/month all-in, of which model API is $100–400 and human review is the largest slice

Compare that to the manual baseline. Five thousand documents at 4 minutes each of full manual handling is 333 hours a month — about $12,000–13,300/month in loaded clerk time, or most of two full-time people. That's the comparison that matters, and it's why this maths usually works.

It's also why it sometimes doesn't. At 500 documents a month instead of 5,000, the manual baseline drops to roughly $1,200/month, the build cost doesn't change, and the payback period stretches past the point where you should bother. Volume is the first thing we check, and it's the most common reason we tell someone not to hire us.

The honest caveats

• Token pricing moves. The per-token figures above are directionally right at time of writing, but check the provider's current pricing page before you build a business case on them.
• Token counts vary a lot by document. A one-page invoice and a forty-page contract are not the same workload. Measure yours on a real sample.
• These are our estimates, not your invoice. The only way to know your number is to run a labelled sample through a real pipeline — which is most of what a readiness assessment does.