“How much does AI cost?” depends almost entirely on what you are trying to do. Turning on an AI feature in your CRM is a very different purchase from building an agent that reads your inbox, checks customer history, updates records, and follows up automatically. For a small business, the useful comparison is the total cost of each realistic way to get the job done.

Small Business AI Cost Tiers: Native, Standalone, and Custom

There are three basic cost levels, and we would check them in this order.

The first is AI already inside software you use. Your CRM, accounting platform, email provider, payroll system, or project management tool may already include AI for drafting, summarizing, categorizing, forecasting, or follow-up. Sometimes it is included in your current plan; sometimes it sits one tier up. Either way, this is usually the cheapest place to start because you are not introducing another system.

That does not mean the cost is literally $0. Someone still has to find the feature, configure it, test it on real work, and decide whether it is good enough. But if a tool you already pay for can solve 90% of the problem, that is usually a better first move than buying or building something new.

The second level is standalone software: a purpose-built tool for a specific job, like an AI support platform, scheduling product, bookkeeping tool, meeting assistant, or automation platform. Many small-business products start in the tens of dollars per month and move up with seats, volume, or functionality.

A $50 or $100 monthly subscription can be a great deal if it replaces five hours of work every week. The important comparison is not “is $100 expensive?” It is “what are we paying today, in software + employee time, to get this job done?”

The third level is custom AI. For a small business, “custom” usually means configuring an agent or workflow around the systems you already use, not building a new software product from scratch. The main cost is typically the setup: mapping the workflow, connecting systems, encoding your rules, setting permissions, and testing it on real data. Ongoing model and infrastructure costs are often much smaller.

Complexity drives the price. Connecting one inbox to one CRM with three clear rules is a very different project from a workflow that reads several data sources, handles exceptions, remembers prior interactions, and writes back into multiple systems.

One nuance that gets missed in the buy-vs.-build conversation: sometimes a simple DIY build is materially cheaper than another SaaS subscription. If a narrow $200/month tool can be replicated with a lightweight automation that costs very little to run and does not create a maintenance headache, building can be the more economical option. The decision is not “buy good, build bad.” It is whether the extra flexibility or savings justify owning the setup yourself.

How Token and Usage-Based AI Pricing Works

If you use an AI model through an API, part of the cost is based on how much text the model processes. That text is measured in tokens, which are small chunks of words. Providers usually charge separately for what the model reads and what it writes.

Using the pricing example in the original research for this article, Anthropic’s Claude Sonnet 5 costs $2 per million input tokens and $10 per million output tokens. On a normal small-business workflow, those numbers can be surprisingly small.

Say an agent reads a customer email plus some account history — roughly 2,000 tokens — and writes a 400-token draft response. Even at a few hundred runs per month, the pure model cost may only be a few dollars.

Where token costs start to matter is when the workflow runs at very high volume or repeatedly feeds the model far more context than it needs. An agent that re-reads a huge knowledge base every time it answers a simple question will cost more than one that only pulls the relevant information.

For most small businesses, though, we would not let token pricing dominate the buying decision. A workflow that costs $12/month in model usage but requires 20 hours of employee review is expensive. A workflow that costs $50/month and gives 30 hours back is not.

Hidden Implementation and Maintenance Costs

This is where most AI budgets get understated.

Setup is the obvious one. Someone has to understand the current process, decide where AI belongs, connect the systems, write the rules, set permissions, and test the awkward cases. Off-the-shelf software has setup costs too; they just tend to show up as your team’s time instead of an implementation invoice.

Integrations can add cost quickly if one of your existing tools requires a higher plan for API access, a connector subscription, or some custom work to expose the right data.

Human review matters more than people expect. If a person spends two minutes reviewing each draft and the workflow handles 500 items per month, that is more than 16 hours of employee time. It may still be a huge improvement over the old process, but those hours belong in the calculation.

Maintenance is the ongoing piece: logins expire, software changes, processes change, and exceptions show up that nobody saw in testing. A workflow that touches several systems will need some amount of upkeep.

And there is an adoption cost in the first few weeks. Teams often run the old process and the new one in parallel while they build trust. That is reasonable, but it temporarily increases the labor cost.

This is why we prefer an all-in cost rather than a software-price comparison. If you are evaluating two options, put subscription fees, setup, internal hours, review time, and expected maintenance on the same page.

Predictive ROI Modeling: Calculate Payback Before You Spend

You do not need a complicated financial model. You need a few assumptions that are explicit enough to challenge.

Monthly value = hours expected back × real hourly cost

Use fully loaded employee cost when you can. The original research for this article cites March 2026 BLS data showing private-industry compensation averaging $46.60 per hour worked, with a median of $34.78.

Then calculate the recurring cost:

Monthly cost = subscriptions + model usage + review time + maintenance

Keep the setup separate:

Payback period = one-time setup cost ÷ monthly net value

A simple example: a workflow takes about 43 employee hours per month today. At $34.78 per hour, that is roughly $1,495 in monthly labor.

If the new workflow costs $150 per month to operate and $4,000 to set up, the projected monthly net value is about $1,345. That implies a payback period of roughly three months.

Then make the assumptions worse on purpose. Cut the expected time savings in half. Increase review time. Add a software upgrade you forgot about. If the project still makes sense, you have a much stronger case.

The forecast tells you whether something is worth trying. It does not tell you whether it worked. Once the workflow is live, measure the realized ROI against the baseline you wrote down before launch.

Questions to Ask Before Approving an AI Budget

Before you approve a tool, vendor, or custom project, get clear answers to these:

1. What is the total monthly cost at our actual volume, including overages?
2. What setup work is included, and what work falls on our team?
3. Do any of our existing tools need a plan upgrade or paid connector?
4. How much human review does this assume in month one and once the workflow is stable?
5. Who maintains it when our process or software changes?
6. What do we own if we stop working with the vendor?
7. What would make you tell us not to buy this?

That last question is one we like because it forces the provider to show some judgment. If the answer is “nothing,” you are probably listening to a sales pitch.

If you want help pricing a workflow this way, send Unprompted the process you are considering. We can help separate the software cost from the implementation cost & the employee time that usually gets left out.

The Bottom Line

• Compare the all-in workflow cost, not just the software fee.
• Native features are usually the cheapest place to start; standalone software is often best for standard jobs; custom can make sense for cross-system work or when a simple DIY build is materially cheaper.
• Model usage is often much cheaper than the setup, review, and maintenance around it.
• Run a payback calculation before spending, then replace the forecast with real numbers after launch.
• Next: how to choose AI tools without adding unnecessary software, buy vs. build, or return to the full AI for small business guide.