AI Automation Cost: What Drives the Price

AI automation cost has two parts that behave very differently. There is the cost to build the automation, which is mostly people's time, and the cost to run it, which is mostly usage fees that rise and fall with volume. Most surprises come from confusing the two, or from forgetting the third part: keeping it working.

This post breaks down each part so you can read a quote, estimate running costs yourself, and spot what is missing. You will not find a typical project price here. What you will find is the arithmetic for the running cost, which you can do yourself with a provider's public price list.

The three layers of AI automation cost

1. Build: designing the workflow, writing prompts and logic, connecting your systems, testing with real examples.
2. Run: model usage fees from the AI provider, plus hosting, plus any other software the automation relies on.
3. Upkeep: monitoring, fixing what breaks when a connected app changes, updating prompts as your business changes, and reviewing outputs.

A quote that covers only the first layer is incomplete.

What drives the build cost

Number of systems involved. An automation that reads emails and drafts replies touches one system. One that reads an intake form, checks your scheduling software, updates the CRM, drafts a care plan summary and notifies a coordinator touches five. Each connection needs authentication, error handling and testing. Integrations are usually the largest share of build time.

How messy the inputs are. Clean, structured form data is cheap to work with. Scanned PDFs, handwritten notes, forwarded email chains and photos of receipts need extraction and checking, and more testing to reach acceptable accuracy.

How costly a mistake is. Drafting a social post can tolerate occasional misses. Classifying a tax document, routing a patient message or quoting a price cannot. Higher stakes mean more test cases, more guardrails and a human review step designed into the flow.

Where data is allowed to go. If you handle health or financial data, the build may need specific vendor agreements, data residency settings, logging and access controls. Our post on an AI policy for small business covers the decisions to make first.

Off-the-shelf versus custom. Many simple automations can be assembled in existing tools. Our guide to workflow automation for small business covers that end. Custom builds make sense when the logic is specific to your business or the volume is high.

What drives the running cost

AI providers charge by usage, measured in tokens, which are small chunks of text. Anthropic and OpenAI both publish prices per million tokens, with separate rates for input (what you send the model) and output (what it writes back). On their pricing pages, output tokens cost more than input tokens for the same model, and the larger, more capable models cost much more per token than the smaller ones.

The levers that move the monthly bill:

• Model choice. Using a smaller model for simple steps like sorting or extraction, and a larger one only where reasoning matters, is often the biggest saving.
• How much text goes in each time. Long instructions, whole documents and conversation history all count as input on every request.
• Caching repeated context. Anthropic's pricing documentation says a prompt cache hit costs 10% of the standard input price as the standard rate, and less on some of its newest models, so instructions or reference documents reused across many requests become much cheaper. Writing to the cache costs a little more than normal input. OpenAI lists a lower cached-input rate too.
• Batch processing. Anthropic's Batch API offers a 50% discount on input and output tokens for work that does not need an instant answer, and OpenAI's Batch API guide describes a 50% cost discount compared with its synchronous APIs, with a 24-hour completion window. Overnight document processing is a good fit.
• Tools the model uses. Some features carry their own fees. Anthropic, for example, lists web search at $10 per 1,000 searches plus token costs.
• Volume. Usage costs scale with the number of emails, calls, documents or chats. Estimate from your real monthly numbers.

Prices change often, so check the provider's current pricing page when you estimate, and ask your builder to show the math.

A worked example: estimating before you build

Say a CPA firm wants an automation that reads each incoming client email, labels it (document submission, question, deadline, billing), extracts any dates and drafts a short reply for staff to approve.

To estimate running cost:

1. Count volume. How many client emails a month? Peak season might be several times the quiet months.
2. Measure size. Take ten typical emails and the instructions the model will need. Estimate input tokens per email, and output tokens for the label and draft reply.
3. Pick models per step. A small model for labeling and extraction; a stronger one only for the draft.
4. Apply discounts. Cache the fixed instructions. Batch anything that can wait.
5. Multiply and add a margin for retries and growth.

Here is the math with hypothetical inputs. Suppose the firm gets 2,000 client emails a month, each request sends about 1,500 input tokens (the email plus instructions) and the model writes about 300 output tokens. That is 3 million input tokens and 0.6 million output tokens a month. Anthropic's pricing page, checked in October 2026, lists Claude Haiku 4.5 at $1 per million input tokens and $5 per million output tokens. So the month costs about $3 for input and $3 for output, roughly $6 before caching or batching. Run the drafting step on a model that costs five times as much per token and the same volume is roughly $30. Your volumes and the current prices will differ; the method stays the same.

For many small business workflows, this math shows usage fees are modest compared with the build and the staff time saved. For high-volume work like transcribing and summarizing every phone call, usage can become the main cost, and the model choice matters a lot. Our breakdown of AI chatbot cost walks through a similar estimate for chat.

Questions to ask any vendor

1. What is included in the build, and which integrations are covered?
2. Which models will run each step, and why?
3. What is your estimate of monthly usage cost at our volume, and at three times our volume?
4. Whose account are the AI and hosting fees billed to? Billing to your own account keeps costs visible and portable.
5. How do you measure accuracy before launch, and what is the review step?
6. What happens when a connected app changes its API? Is maintenance included, and for how long?
7. Can we see the logs of what the automation did?

How Pinn.Media prices AI automation

We scope automations by the three layers above, estimate running costs from your real volumes before we build, and set up model usage in accounts you own. Our case study for an AI automation studio shows the kind of live demos and service pages that come out of that thinking. If you have a workflow you want to automate, book a discovery call and we will walk through the build, run and upkeep costs with you.