What does an AI workflow actually cost?

Write down the unit of work

Start with a business unit you can count: one document processed, one review classified, or one report generated. Monthly token totals become meaningful only when they are connected to the work you need to finish. Define both expected monthly volume and the share of results that must pass human review.

This guide uses illustrative arithmetic rather than a current provider price list. For an actual budget, copy rates from your provider's pricing page, record the model identifier and date, and check whether the rate applies to your account and processing mode.

Estimate the text generation component

For providers that charge separate rates per million input and output tokens, multiply each token quantity by its matching rate and divide by one million. Add those two costs, then multiply by the expected number of requests. Measure token counts on a sample; character counts are only a rough proxy and vary by language and tokenizer.

As a hypothetical example, 2,000 input tokens at $1 per million and 500 output tokens at $4 per million cost $0.004 per request. At 10,000 identical requests, that component is $40. These example rates do not describe any named product.

• Include repeated system instructions and retrieved context in input estimates.
• Account for conversation history if you resend it on each turn.
• Model cached-input pricing separately when it applies.
• Check how the provider bills reasoning, images, audio, and tool usage.

Add the costs outside tokens

Real workflows may also pay for search, data collection, embeddings, vector storage, code execution, and network transfer. Some providers charge per tool call or session. These components need their own line items because they do not necessarily scale in proportion to text tokens.

Human review is another cost. If a result needs two minutes of correction, the time spent correcting thousands of results may outweigh a small difference in API rates. Track handling time for a representative sample rather than assuming every output is ready to use.

Budget for retries and growth

A failed request may lead to another generation, additional context, or a switch to a more expensive model. Estimate retries explicitly and check your provider's billing treatment of failed calls. A simple planning allowance is useful, but it should be replaced by measured retry behavior after a pilot.

Our calculator compares two user-defined token-price scenarios. It includes a configurable retry allowance; it excludes caching discounts, tool charges, tax, and review time. It runs locally in your browser and sends no calculation inputs to a server. Use it to explore assumptions, then verify those assumptions against a small live run.

A related practical decision

• Compare automation platform costs: n8n vs Make ↗