AI automation cost for a small business comes down to four things: the build, the software subscriptions that run it, the AI model usage, and the upkeep after launch. A single workflow that moves form leads into a CRM sits at the low end. A multi-step process that reads documents, makes decisions and updates three systems costs more to build and more to run.
There’s no honest flat number, because two workflows with the same name can differ widely in effort. What you can do is break the cost into parts, price each part against your own setup, and compare the total to the hours the workflow gives back. That’s the method this post walks through.
If you only want the payback math, skip to the section on working out return, or plug your numbers into our automation ROI calculator.
What you’re paying for when you automate a workflow
AI workflow automation is software that moves data and makes routine decisions between your business tools without a person doing each step. A typical example: a lead fills in a form, an AI model reads the message and scores it, the lead lands in your CRM with the right owner, and a reply goes out within minutes.
Every workflow like that has the same cost layers, whoever builds it.
- Build cost. The one-time work of mapping the process, building it, testing it with real data and handling the cases where things go wrong.
- Platform subscription. The automation tool that runs the workflow, such as Zapier, Make or n8n.
- AI model usage. Calls to models from OpenAI, Anthropic or Google, billed by the amount of text processed.
- Connected app costs. Some apps only open their API on higher plans, so connecting them can force an upgrade.
- Upkeep. Fixing breaks when an app changes its API, adjusting prompts and adding new steps as the business changes.
AI automation cost: the build
The build is usually the largest single line, and it’s driven by complexity more than by the tool. Four questions decide most of it.
How many steps and systems are involved
A workflow that touches two apps is quick to build and test. One that pulls from email, writes to a CRM, updates a spreadsheet and posts to Slack has four places to fail, and each one needs testing and error handling.
How clean your data is
Automation exposes messy data fast. If your CRM has duplicate contacts, free-text fields where there should be dropdowns, or three spellings of the same product, part of the build becomes cleanup. That’s work worth doing anyway, but it belongs in the estimate.
How many exceptions there are
The happy path is often a fraction of the build. The rest is what happens when a form arrives half empty, an invoice is a photo instead of a PDF, or the AI isn’t confident in its answer. Good builds route those cases to a person instead of guessing.
Whether a person reviews the output
Adding a human approval step, such as a Slack message with approve and reject buttons, adds some build time. It also cuts the risk of an AI mistake reaching a customer, which is why we add one to most customer-facing workflows.
Running costs: platforms and AI usage
Running costs are monthly and scale with volume. They’re easy to underestimate at the start and easy to control once you know what drives them.
Each automation platform bills differently. Zapier charges by tasks, where each successful action step counts. Make charges by usage per module run. n8n can be self-hosted for free under its community license, which trades a subscription for server and maintenance costs. Our comparison of n8n, Make and Zapier covers which suits which kind of business.
AI model costs come from API calls. OpenAI, Anthropic and Google all price their APIs per token, which roughly tracks the amount of text sent in and generated. Classifying a short email costs very little per run. Summarizing long contracts or reading scanned documents costs more, so the volume of long documents matters more than the number of runs.
A few habits keep these costs down:
- Filter early, so the workflow stops before the AI step when a record doesn’t need it.
- Use a smaller, cheaper model for sorting and tagging, and save the larger model for drafting and reasoning.
- Send only the fields the model needs, not the whole record.
- Batch low-priority jobs to run once an hour instead of on every event.
- Set spending limits in your AI provider’s dashboard so a loop can’t run up a bill overnight.
Upkeep: the cost people forget
Automations break. An app renames a field, a login expires, or an API version is retired. None of this is a sign of a bad build. It’s the normal life of software that depends on other software.
Plan for upkeep in one of two ways. Either someone on your team learns the platform well enough to fix small breaks, or you keep a support arrangement with whoever built it. Either way, ask for error alerts that go to a named person, plus a short written run sheet for each workflow that explains what it does and where it can fail.
Three sizes of workflow
It helps to think in sizes instead of chasing a single price. The figures will depend on your tools and your volume, but the shape of the cost is predictable.
Simple: one trigger, a few steps
Form to CRM, new customer to welcome email, invoice paid to a Slack message. These are quick to build, cheap to run, and usually the best place to start. Our list of seven workflows worth automating first is mostly made of these.
Mid-size: AI reads and decides
Inbox triage, lead scoring, pulling data out of PDFs into accounting software. The AI step adds model costs and the need for review queues and confidence checks. Build effort goes up because testing has to cover many real examples.
Large: agents across several systems
An AI agent that answers customer questions using your help docs, order data and CRM history, then hands off to a person when needed. These need the most testing, the most monitoring and the clearest rules about what the agent may and may not do.
How to work out whether it pays back
The return on a workflow is the time it gives back, plus any errors or missed leads it prevents, minus what it costs to build and run. You can get a useful estimate in about fifteen minutes.
- Pick one process and count how many times it runs in a typical week.
- Time how long a person takes per run, including switching between apps.
- Multiply to get weekly hours, then multiply by the loaded hourly cost of the person doing it.
- Estimate the share of runs the automation can handle without a person. Be conservative here.
- Compare the yearly saving to the build cost plus a year of running costs and upkeep.
As a worked example with made-up numbers: if a coordinator spends six hours a week copying web leads into a CRM, and the automation handles most of them, you get most of those six hours back every week. Put your own figures into the automation ROI calculator and it does the math for you.
Speed counts too, and it’s harder to put in a spreadsheet. A lead that gets a reply in two minutes instead of the next morning is more likely to book. That benefit doesn’t show up in hours saved, so treat it as upside.
Questions to ask before you sign off on a quote
- Which platform will this run on, and who owns the account?
- Which AI model is used at each step, and why that one?
- What happens when the AI isn’t confident, or an app is down?
- Who gets the error alerts, and how fast are breaks fixed?
- Is there a written run sheet I can hand to a new hire?
- What are the expected monthly running costs at my current volume?
A builder who can answer all six clearly has thought about how your business runs day to day.
Getting a real number for your business
The fastest way to get an accurate figure is to describe one process in detail and have it scoped. At Intelligent Solutions Tech we start with a free audit, then send a written proposal with a fixed monthly scope and price, so you know the full cost before any build starts. Work is staged and you review it before anything goes live.
See how we approach builds on our AI workflow automation services page, or send us one workflow you’d like to automate first.