What AI Really Costs a Small Business: Seats, Tokens, Integrations, and the Hours Nobody Counts
The invoice looks harmless. Ten seats at twenty-something dollars each, plus an API bill that came in under fifty bucks last month. Leadership signs off, and AI goes in the budget as "cheap."
Then the office manager spends Friday afternoons fixing CRM fields the automation mangled. A developer loses two days rebuilding a broken Zapier chain. Someone has to read every AI-drafted client email before it goes out. None of that is on an invoice, and for most small businesses it's the biggest line item.
This guide is for founders, operators, service businesses, and small tech teams who want the full number before committing. We'll go through the four cost layers in order: seats, tokens, integrations, and hours. Prices come from official vendor pages as of October 4, 2026. Vendors change prices often, so check the linked pages before you budget.
The short version
- Seats are predictable and easy to over-buy. Budget for the people who will use AI daily, and leave out the people who might try it someday.
- Tokens are usually smaller than people fear for normal business workloads. They grow quickly with long prompts, premium models, and agents that loop.
- Integrations are where one-time costs pile up, and they keep costing you every time a vendor changes an API.
- Hours (review, cleanup, maintenance, governance) are the cost nobody tracks, and they decide whether your ROI is real.
Layer 1: Seats — the subscription you can see
Seat pricing is the familiar SaaS model: a fixed fee per user per month. It's the easiest layer to forecast, and it's where most small businesses start.
| Plan (as of Oct 4, 2026) | Listed price | What to watch |
|---|---|---|
| Claude Team, standard seat | $20/seat/month billed annually, $25 monthly (Anthropic) | Built for teams of 2–150. Premium seats cost $100 annually or $125 monthly and include more usage. |
| Claude Enterprise | $20/seat/month billed annually, plus usage at API rates (Anthropic) | A hybrid model: the seat fee is fixed, and the usage portion grows with how much your team runs. |
| Microsoft 365 Copilot Business (add-on) | $21/user/month paid yearly, discounted to $18 for the first year on a promotion running through Dec 31, 2026; $25.20 month-to-month (Microsoft) | Requires a separate qualifying Microsoft 365 license, so the true per-user cost is the add-on plus the base plan. |
| ChatGPT Business | Priced per user per month, with monthly and annual options (OpenAI) | Starts at 2 users. Check the live page for current seat pricing and credit options. |
The trap: a ChatGPT or Claude subscription does not include API access. OpenAI states that API usage is billed separately from ChatGPT Plus, Business, Enterprise and Edu. If you plan to build automations on top of the model, that's a second bill.
The decision: count daily users, not curious users. A 12-person firm where 4 people use AI every day should probably buy 4–6 seats and let everyone else use free tiers for low-risk experiments, without client data. Revisit after 60 days.
Layer 2: Tokens — the meter that runs while you sleep
When you connect an AI model to your own systems, such as a support inbox, an intake form, or a document pipeline, you usually pay per token. A token is a chunk of text. Anthropic's rough guide is that one token is about four characters, or three-quarters of an English word. Input tokens (what you send) and output tokens (what the model writes) are priced separately, and output almost always costs more.
What the meter looks like today
| Model (standard rates, Oct 4, 2026) | Input / 1M tokens | Output / 1M tokens | Source |
|---|---|---|---|
| OpenAI GPT-5.6 Luna (fast, affordable) | $0.20 | $1.20 | OpenAI |
| OpenAI GPT-5.6 Sol (flagship) | $5.00 | $30.00 | OpenAI |
| Claude Haiku 4.5 | $1 | $5 | Anthropic |
| Claude Sonnet 5.5 | $2 | $10 | Anthropic |
| Gemini 3.8 Flash | $0.75 through Dec 31, 2026; $1.50 from Jan 1, 2027 | $3.75 through Dec 31, 2026; $7.50 from Jan 1, 2027 |
The Gemini row matters. Google publishes a scheduled price increase on its pricing page: the standard Flash rates double on January 1, 2027. If you budget a 2027 project at today's rate, your forecast will be off by 2x on day one. Always check for dated price changes.
A worked example: a support assistant
Suppose a service business routes 2,000 customer conversations a month through an AI assistant. Each conversation sends about 3,000 tokens of context (instructions, the customer's message, a knowledge-base excerpt) and gets back about 500 tokens. That's 6 million input tokens and 1 million output tokens a month. At the standard rates above:
- GPT-5.6 Luna: about $2.40/month
- Gemini 3.8 Flash: about $8.25/month at 2026 rates, about $16.50 at 2027 rates
- Claude Haiku 4.5: about $11/month
- Claude Sonnet 5.5: about $22/month
- GPT-5.6 Sol: about $60/month
For this kind of workload, the token bill is close to a rounding error. Anthropic's own example puts 10,000 support tickets on Haiku 4.5 at roughly $37. The real question is whether the cheap model answers well enough that a human doesn't redo the work, because human rework costs far more than the model upgrade.
Where token bills actually blow up
- Agents that loop. An agent that searches, reads, reasons, and retries can burn through tokens on every step. Each tool definition also adds tokens to every request. Anthropic documents a browser toolset that adds roughly 6,600 input tokens per request before the agent does anything.
- Paid tools on top of tokens. Web search runs $10 per 1,000 calls on OpenAI and $10 per 1,000 searches on Claude. Gemini includes 5,000 grounded searches a month free, then charges $14 per 1,000.
- Stuffing whole documents into every call. Sending a 40-page policy manual with every question multiplies your input bill for no reason.
- Defaulting to the flagship model. In the example above, the gap between the cheapest and most expensive option is 25x.
Tokens aren't apples-to-apples across vendors
Different models break the same text into different numbers of tokens. Anthropic notes that its newer models use a tokenizer that produces about 30% more tokens for the same text. A lower per-token price doesn't guarantee a lower bill. Run your actual prompts through each candidate and compare the invoices, not the rate cards.
Three levers that cut the bill in half or better
- Batch processing. OpenAI, Anthropic, and Google all offer about 50% off for work that can wait for results, such as overnight document tagging, weekly report summaries, or bulk lead scoring (Anthropic, Google).
- Prompt caching. If every request starts with the same long instructions, caching lets you pay a fraction of the input price on repeats. Anthropic prices a standard cache hit at 10% of the base input rate, and OpenAI lists cached input at one-tenth of normal input on its current models (OpenAI).
- Model routing. Send routine classification and short drafts to a small model, and send only the hard cases to a premium one.
Set a hard ceiling on day one. OpenAI lets you set a monthly budget that stops requests once it's reached, but warns that enforcement may lag and you're responsible for any overage. Set the cap lower than your real limit, and turn on email alerts.
Layer 3: Integrations — where the one-time costs hide
An AI model that can't see your CRM, calendar, inbox, or job-management system is a chat window. Connecting it is where real money and time go:
- Build time. Someone has to wire the model to your systems, whether that's your developer, a contractor, or a no-code platform. This is usually the largest single upfront cost.
- Middleware subscriptions. Automation platforms, vector databases, and hosting add their own monthly fees on top of the AI vendor.
- Data cleanup. If your CRM has duplicate contacts and inconsistent fields, AI will confidently produce output based on that mess. Cleaning data before launch costs less than cleaning up AI mistakes after.
- Breakage. APIs change, models get retired, and connectors break. An integration needs a maintenance budget, not just a launch budget.
- Security and data residency. If you need data processed in a specific region, expect a premium. Anthropic charges 1.1x for US-only inference, and OpenAI lists data residency at +10%.
The decision: before building anything custom, check whether a tool you already pay for has the AI feature built in. Microsoft markets Copilot as starting inside the apps teams already use, with 100+ connectors for search across business systems. A built-in feature that's 80% as good with zero build time often beats a custom integration that's 100% as good and takes six weeks.
Layer 4: The hours nobody counts
This is the layer that turns a "cheap" AI rollout into an expensive one. The hours are real, they recur, and they rarely show up in a budget spreadsheet.
Where the hours go
- Review. Someone has to check AI output before it reaches a client, a contract, or a payroll file. For anything customer-facing or regulated, that review is permanent.
- Prompt and workflow upkeep. Business rules change, so the instructions behind your automations need updating too.
- Exception handling. The 5% of cases the AI gets wrong usually take the most human time to untangle.
- Training and adoption. Showing people what AI is good for, and what it shouldn't touch, takes repeated effort beyond a single lunch-and-learn.
- Governance. Deciding what data can go into which tool, who approves new use cases, and what happens when something goes wrong.
Put a number on it
Use this formula for each AI workflow:
Monthly hidden cost = (review hours + maintenance hours + exception hours) × fully loaded hourly cost
Here's a hypothetical. If an operations lead with a fully loaded cost of $50/hour spends 3 hours a week reviewing and fixing AI-drafted proposals, that's about $650 a month. That's more than ten Claude Team seats and over 200 times the token bill in the support example above. You're not tracking the wrong expenses; you're tracking the smallest ones.
Use a real framework for the governance hours
You don't need a compliance department, but you do need a system. The NIST AI Risk Management Framework is free, voluntary, and was released January 26, 2023. Its four functions are Govern, Map, Measure, and Manage. NIST also published a Generative AI Profile (NIST-AI-600-1) in July 2024 covering risks specific to generative AI, and NIST notes the core framework is being revised under the White House AI Action Plan.
A practical small-business version:
- Govern: one named owner for AI decisions, plus a one-page policy on what data is off-limits.
- Map: a list of every AI tool in use, who uses it, and what data it touches, including the free tools people signed up for on their own.
- Measure: a monthly check of error rates and review time per workflow.
- Manage: a decision to keep, fix, or kill each workflow based on the numbers.
Calculating AI ROI without fooling yourself
Most AI ROI math compares the subscription fee to the hours "saved" and stops there. Here's a more honest version:
Net monthly value = (hours saved × loaded hourly cost) + measurable revenue impact − (seats + tokens + middleware + amortized integration build + review and maintenance hours)
Ground rules that keep the math honest:
- Measure before you launch. Time the current process for two weeks. Without a baseline, "time saved" is a guess.
- Count saved hours only if they get redeployed. If the 6 hours a week saved on scheduling turn into longer lunches, that's a perk, not ROI. Decide in advance what those hours will go toward.
- Amortize the build. Spread integration costs over the realistic life of the workflow, typically 12–24 months. Don't count it once and forget it.
- Track rework. If an AI-drafted document needs 20 minutes of fixing, subtract those 20 minutes.
- Pilot one workflow at a time. Pick a high-volume, low-risk, easy-to-measure process (intake triage, meeting summaries, first-draft quotes) before anything client-critical.
A 30-minute budgeting worksheet
- List the workflows. Name three specific tasks you want AI to handle. "Use AI more" doesn't count.
- Count daily users. That number is your seat count.
- Estimate volume. Requests per month × average input and output length. Price it on two models: a small one and a premium one.
- Check the calendar. Look for dated price changes on each vendor's pricing page, and promotions that expire after year one.
- Scope integrations. For each workflow, decide whether it's built-in, no-code, or custom, and get a time estimate.
- Price the hours. Review plus maintenance plus exceptions, multiplied by loaded cost.
- Set the caps. API budget limits, alert thresholds, and a 60-day review date.
Comparing notes with people who have already run this math is often the fastest way to find the line items you missed. That kind of operator-to-operator conversation is what the Tech Lab Miami community is built around.
FAQs
What are the ongoing costs of AI for a small business?
Four categories: per-seat subscriptions, usage-based API charges (tokens plus paid tools like web search), integration maintenance and middleware, and the staff hours spent reviewing, maintaining, and governing AI output. The last category is usually the largest and the least tracked.
Are LLM API costs expensive?
For typical business workloads on small models, usually not. A few thousand support conversations a month can cost single or low double digits at current rates (OpenAI, Anthropic, Google). Costs rise sharply with premium models, long prompts, agent loops, and paid tool calls, so set budget caps before launch.
Does a ChatGPT or Claude subscription include API access?
No. OpenAI bills the API separately from ChatGPT subscriptions, and Claude's API is likewise pay-as-you-go through its developer platform (Anthropic). Budget them as two separate lines.
How do I keep AI tool pricing from surprising me?
Date every price you put in a budget, check vendor pages for scheduled changes and expiring promotions, set hard monthly limits on API accounts, and review actual spend against forecast every month for the first quarter.
Is AI worth the investment for a small business?
It depends on the workflow. AI tends to pay off on high-volume, repetitive tasks where output is easy to check. It tends to disappoint where every output needs heavy human review. Measure a baseline, pilot one workflow, and let the numbers decide.
The bottom line
The subscription is the smallest part of what AI costs you. Tokens are usually manageable if you pick the right model and set caps. Integrations cost more than expected and keep costing as vendors change things. The hours spent reviewing, fixing, and governing AI output are what separate a rollout that pays for itself from one that only looks cheap. Put all four layers in the spreadsheet before you sign anything.
Visuals

