The honest answer is that AI for a small business costs somewhere between $25 a month and $5,000 one time, and the number depends on a question most vendors skip: what is the AI actually for. This note lays out the real cost buckets with actual dollar figures, every one traceable to a published price or a public source, so you can build a budget without a discovery call first.

The three cost buckets

Every AI expense a small business touches falls into one of three buckets. Keeping them separate is what makes the math honest.

  • Subscription: a per-user, per-month fee for a cloud service. ChatGPT Team (now ChatGPT Business) is $25 per user per month billed monthly at OpenAI's published price — $20 per user billed annually. Five seats is $125 a month, or $1,500 a year.
  • Hardware: a one-time purchase if you run AI locally. A capable dedicated mini-PC costs about $1,400 one time, and it stays yours. Many local setups also run on the computer you already own, which makes this bucket zero.
  • Implementation: the one-time cost of getting the AI to do your actual work. This is the bucket vendors quote last and owners pay for most.

The model itself is often free. The model we run in our published local work, Qwen3.8-27B, is an open-weight model under the Apache-2.0 license. You download it once and run it with zero per-token cost, the same way you would run a spreadsheet, not the way you rent software.

Subscription math, straight

Subscriptions are not a scam. They are cheap for what they do, and the right first step for most small businesses. The math only turns against you in two situations.

First, at team scale. Five people on ChatGPT Team at $25 per user per month (billed monthly) is $1,500 a year, every year, forever. A one-time local setup on a $1,400 mini-PC with a $650 standard starter install is $2,050 once. The subscription passes that total at about 16 months. Over three years the subscription costs $4,500 and the local setup still costs $2,050 plus electricity.

Second, at workflow scale. A subscription bills per person whether the AI is doing real work or sitting idle in a browser tab. A local setup bills nothing per use, so a workflow that runs 200 times a week costs the same as one that runs twice.

Flip it around for honesty: for one or two people doing occasional writing and research, a subscription at $25 to $50 a month stays cheaper than any local setup for years. If that is your situation, keep the subscription and stop reading pricing posts.

Implementation math, published

These are the actual PuenteWorks prices, published and current, so you can place them in the budget without calling anyone:

  • Owner AI Setup Hour: $175. One hour, one written recommendation: local, cloud, hybrid, or stop. This is the cheapest way to avoid the most expensive mistake, which is buying the wrong bucket.
  • Local AI Starter Setup: $450 beta / $650 standard. One person, one machine, a working local AI app, one safe workflow, seven days of async support.
  • Private AI Workbench: $1,250. The setup connected to real repeated work, with workflow templates and a model and use-case matrix.
  • Small Team Local AI Kit: $2,500 to $4,500. A shared local AI surface for 2 to 5 people, with a written scope and a named admin.
  • AI operations readiness: $950 to $1,500 on the consulting path, before any larger build.

Full context for these numbers, including what each one is for, is in the forward-deployed AI note with real prices.

The hidden costs nobody quotes

Three costs show up after the invoice, and they are where budgets actually break.

  • Training. A tool nobody uses is a 100 percent loss. Budget real hours for the people who will run it, not a memo. Local setups include a walkthrough for exactly this reason.
  • Maintenance. Someone has to own updates, backups, and the answer when it breaks. A subscription bundles this into the fee. A local setup assigns it to a named person, which is cheaper and only works if the name exists.
  • Failed pilots. The biggest hidden cost in small business AI. A pilot that never touches real repeated work produces a demo, a shrug, and a canceled card. The fix is not a better tool, it is picking one repeated task with a clear definition of done and measuring before and after. That discipline costs nothing and is the entire difference between spending and investing.

Which bucket fits you

A short version of the decision rule:

  1. Occasional writing, research, and questions, one or two people: subscription. $25 to $50 a month, done.
  2. Repeated weekly work on private documents, one person: local starter. Roughly $2,050 one time including the machine, or less if your current computer qualifies.
  3. A team doing repeated work on private documents: the small team kit, $2,500 to $4,500 plus hardware, against a subscription bill that would run $3,000 to $6,000 over three years for the same seats.

If you cannot tell which line you are on, that is exactly what the $175 Setup Hour answers in writing before you spend anything larger. The services page has the full ladder with descriptions.

FAQ

Is the free model really free?

The model weights are. Qwen3.8-27B is Apache-2.0 licensed, downloadable, and has no per-token fee. What can cost money is the hardware to run it well and the time to set it up. That is the honest version of "free."

What does AI cost per year for a 10-person business?

Subscription path: 10 seats at $25 per user per month is $3,000 a year. Local path: a one-time $2,500 to $4,500 team kit plus about $1,400 in hardware, which is under $6,000 once and roughly $0 in year two and after. Estimates assume list prices and one shared machine.

What is the cheapest useful first step?

A $25 month of a subscription for general work, or the $175 Setup Hour if the question is which repeated workflow deserves the spend. Both cap your downside.