How Much Does an AI Agent Cost a Small Business?

An AI agent can cost a small business nothing extra, if it's the agent mode built into a $20-$25 plan your team is already on, such as ChatGPT Work or Claude Cowork. Customer-service agents on outcome pricing cost about $0.50 to $1 per resolved conversation. Custom agents on automation platforms or APIs run from about $30 to a few hundred dollars a month.

Software is usually the smaller part of the bill. An agent is software that takes actions on its own, such as sending an email, updating a booking or placing an order, so someone has to set its permissions, check its work, and deal with its mistakes. For most small businesses, that supervision time costs more than the subscription in the first few months. The number to compare across options is cost per task done correctly, not the monthly fee.

Follow me on Instagram@sagnikteaches

The five ways agents are priced

Vendors price agents in very different units, which is why they're hard to compare. As of September 2026:

Connect on LinkedInSagnik Bhattacharya

Two notes on the outcome-priced agents. HubSpot's Customer Agent needs any Professional or Enterprise hub (usually Service Hub); it isn't on Starter. Professional includes a monthly pool of about 3,000 credits, which covers roughly 60 resolved conversations at 50 credits each before you pay more at $10 per 1,000 credits. And both vendors charge their normal platform fees on top, so the per-outcome price is only part of the cost.

Subscribe on YouTube@codingliquids

The cost-per-task sum

To compare an agent with the way you do the job now, or two agents with each other, turn everything into a cost per task:

Cost per task =
    (monthly platform fee / tasks per month)
  + usage cost per task (credits, outcomes or tokens)
  + (minutes of human checking per task x hourly cost / 60)
  + (error rate x average cost of fixing an error)

The last two lines are the ones vendors leave out. An agent that costs $0.20 a task in software but needs two minutes of checking by a $25-an-hour employee really costs about $1.03, and if 5% of its tasks go wrong at $20 each to put right, add another $1.

A worked example: quote follow-ups at a car repair garage

Picture an illustrative independent garage that sends about 120 estimates a month for repairs beyond the original job. Around a third go unanswered, and nobody has time to chase. The owner wants an agent that, two days after an estimate goes out, drafts a friendly follow-up, sends it once a person approves it, and logs the reply in a spreadsheet.

Option A: ChatGPT Work, run by the service adviser

The adviser already has a ChatGPT Business Standard seat ($25 a month, monthly billing). Once a day, she asks ChatGPT Work to go through the estimates folder and the sent emails, draft follow-ups for anything two days old with no reply, and add them to a sheet for her to approve and send. No extra software cost.

Software:            $0 extra (existing seat)
Human time:          10 minutes a day to run and approve
                     = about 3.5 hours a month x $25 = $88
Tasks:               about 80 follow-ups a month
Cost per follow-up:  $88 / 80 = about $1.10

Option B: a Zapier agent that runs on its own

A Zapier agent watches for new estimates, waits two days, checks for a reply, drafts the follow-up and puts it in a queue for one-click approval. Say each run uses about five activities. For 120 estimates, that's 600 activities a month, over the free tier's 400, so the paid tier at about $33 a month (annual billing) is needed.

Software:            about $33 a month
Human time:          1 minute per approval x 80 = 80 minutes
                     = about $33 a month
Setup:               about 6 hours ($150), spread over a year
                     = about $13 a month
Cost per follow-up:  ($33 + $33 + $13) / 80 = about $0.99

What decides it

The two options cost about the same per follow-up. Option B saves the adviser about three hours a month, but it's another system to maintain, and it breaks quietly if the estimates folder moves. For 80 follow-ups a month, Option A is a sensible start; Option B earns its place if volume doubles or the adviser leaves. And the payoff is on the other side of the sum: if follow-ups turn even four extra estimates a month into approved work at an illustrative $300 each, either option pays for itself many times over. That's the number to track, not the software fee.

The cheapest agent, tested on a real job

Before paying for anything, it's worth seeing what an included agent mode does with one of your own tasks. Here's an illustrative ChatGPT Work request from a home-care provider's office, and the kind of result it returns.

Request:
Open the file "Training matrix.xlsx" in our shared drive.
List every carer whose moving-and-handling or medication
training expires in the next 60 days. For each, give the
expiry date and the date of their next rostered day off
from "Rota October.xlsx". Put the result in a new sheet
called "Renewals due" and don't change anything else.

Illustrative result:
Created sheet "Renewals due" with 7 rows.
- Carer 14: medication, expires 3 Nov, next day off 29 Oct
- Carer 22: moving and handling, expires 9 Nov, next day
  off 1 Nov
- ...
Note: 2 carers in the training matrix don't appear in the
October rota; I've listed them with "not on rota".

What you'd check: the two carers "not on rota" (have they left, or are they on leave?), and that the expiry dates match the certificates for a sample of three people. This took the coordinator about 15 minutes including checking, against about an hour by hand. Software cost: nothing beyond the existing seat. That's the benchmark any paid agent has to beat.

Why custom agents cost more than their token price suggests

If you or a developer build an agent directly on a model's API, you pay per token (a token is a short chunk of text; a million tokens is roughly 750,000 words). An agent works in steps, and each step usually re-sends the growing conversation, so token use per task climbs quickly.

A quick sum for an illustrative agent that takes 12 steps per task, averaging 8,000 input tokens and 500 output tokens per step:

Per task:  12 x 8,000 = 96,000 input tokens
           12 x   500 =  6,000 output tokens

Claude Sonnet 5 ($2 / $10 per million):
  0.096 x $2 + 0.006 x $10 = $0.19 + $0.06 = about $0.25 a task
Claude Haiku 4.5 ($1 / $5):
  0.096 x $1 + 0.006 x $5  = $0.10 + $0.03 = about $0.13 a task

At 500 tasks a month: about $125 (Sonnet 5) or $63 (Haiku 4.5)

Now imagine a bug that makes the agent retry the same failing step 40 times before giving up. That task costs more than three times as much, and if it happens on every task overnight, the month's bill arrives several times over budget. Set a spending limit on the API account and a maximum number of steps per task before anything runs unattended; our tutorial on spending limits and alerts on pay-as-you-go AI tools shows where the settings are.

The bigger cost of a custom agent is building it. Connecting it safely to your systems, handling failures, and testing it properly is developer time measured in weeks. Unless no off-the-shelf agent can do the job, start with one of the other four models.

Matching the pricing model to your volume

The cheapest model changes as volume grows, so estimate your monthly task count before comparing prices.

• Under about 50 tasks a month: use an agent mode included in a seat you already have. Paying per outcome is also reasonable at this level, because 50 resolutions at $0.99 is about $50, and there's no fixed fee to waste in a quiet month.
• 50 to 500 tasks a month: compare outcome pricing with an activities- or credits-based platform. At 300 resolutions, Fin is about $297 a month; a Zapier agent doing a simpler job at 300 runs of five activities (1,500) sits within the paid tier at about $33. They're doing different jobs, though, so compare cost per task done correctly, not headline prices.
• Over 500 tasks a month: per-outcome costs grow in a straight line. At 1,000 resolutions, Fin is about $990 a month, or nearly $12,000 a year. At that scale, it's worth pricing a custom agent on a low-cost model, or cutting the volume first with a clearer FAQ, better order-confirmation emails and delivery tracking links, so fewer customers need to ask.

The last point is the one most often missed. The cheapest agent conversation is the one that never has to happen.

Three budgets for three kinds of agent

Look at the mid scenario's software line again: at 250 resolutions a month, Fin's $0.99 each adds up to about $3,000 a year before seat fees. That's fine if each resolved conversation saves a staff member five minutes and keeps an order, and poor value if most of those questions could be answered by a better FAQ page. Our tutorial on outcome-based AI pricing covers how to check what counts as "resolved" before you sign.

Costs that don't appear on any price page

• Supervision. Every agent needs someone to check its work, at least at first. Budget one to two minutes per task for the first month, then cut it back as trust is earned.
• Mistakes. An illustrative home-care provider lets an agent send rota-change confirmations without review. It misreads one message and confirms a visit on the wrong day. The cost is a missed visit, an apology and a complaint to handle, far more than a month of the software. Read what can go wrong when AI agents take actions for you before letting any agent send anything unchecked.
• Permissions work. Agents act with the access you give them. Setting up a separate account with only the permissions it needs takes an hour or two and is worth every minute.
• Product churn. Agent products change fast. OpenAI retired ChatGPT agent in July 2026 in favour of ChatGPT Work. HubSpot stopped new installs of five Breeze agents on 23 July 2026. Build around your own process notes so you can move if a product changes.
• The platform underneath. Outcome-priced agents sit on help desks and CRMs with their own seat fees, which often cost more than the agent.

Reading an agent quote: what's usually missing

Agent vendors and agencies often quote a headline figure. Here's an illustrative quote a florist might receive for a website support agent, and the questions it should prompt.

Quote: AI support agent for your online shop
- Setup and training on your website:        $750 one-off
- Platform subscription:                     $99 a month
- Includes 500 AI conversations a month
- Additional conversations:                  $0.40 each

Missing from it: what counts as a "conversation" (one customer message, or a whole exchange?), what happens at 501 (does the agent stop, or keep billing?), whether it can hand over to staff and where those handovers go, who updates the answers when your delivery prices change, and how you get your content out if you leave. At the florist's busiest time, say 1,200 conversations in the week before Mother's Day, the monthly bill would be $99 plus 700 x $0.40, which is $379. Ask for the quote at your peak month, in writing.

Signs an agent is costing more than it saves

• Checking time isn't falling. If staff still spend two minutes on every task after a month, either the agent isn't reliable enough or the checking habit hasn't been reviewed. Sample instead of checking everything once error rates are low.
• Customers ask for a person straight away. An illustrative bookshop finds that most chat conversations start with "can I talk to someone?" Every one of those is billed and resolves nothing.
• The unit count keeps surprising you. Activities, credits or outcomes consistently above your estimate usually mean the agent is taking more steps than planned, or counting partial conversations.
• Nobody can say what it achieved. If the agent's value can't be stated as a number (estimates approved, hours saved, orders kept), you can't tell whether its cost is justified.

How to keep agent costs predictable

1. Start with the agent included in your current plan. ChatGPT Work or Claude Cowork inside an existing seat is the cheapest way to learn what an agent can do for you.
2. Run a two-week pilot with a person approving every action. Measure the cost per task and the error rate before removing the approval step. Piloting your first AI agent without risking customers sets out how.
3. Count units at your volume. Outcomes, conversations, activities or credits: multiply the unit by your real monthly volume, and by the busiest month, not the average.
4. Cap everything. Spending caps, step limits and budget alerts, set before launch.
5. Review monthly for the first quarter. Cost per task, error rate and the business result you wanted. If the cost per task isn't falling as supervision drops, something is wrong.

If you're still working out whether you need an agent at all, rather than a simpler automation or a chatbot, read AI agent vs chatbot vs automation first. Plenty of jobs sold as "agent" work are cheaper as a fixed automation with one AI step.

AI agent costs: what owners ask next

What happened to ChatGPT agent?

OpenAI retired ChatGPT agent, also called Agent Mode, on 9 July 2026 and replaced it with ChatGPT Work, an agent mode for longer multi-step tasks across apps and files. It's available on Plus, Pro, Business and Enterprise at existing plan prices, so for many small businesses the first agent costs nothing beyond an existing seat.

Is outcome-based pricing cheaper than a fixed fee?

It's cheaper when volumes are low or uneven, because you pay only for conversations the agent resolves. It gets expensive as volume grows, since every resolution costs the same. Work out your monthly resolutions, multiply by the per-outcome price, and compare with the fixed-fee alternatives at your actual volume, not the vendor's example.

How do I stop an agent running up a large bill?

Set limits before it goes live: a monthly spending cap on the API account or credit pool, a per-run step limit where the tool offers one, and an alert at 50% and 80% of budget. Run it on a small batch first and check the cost per task before switching it on for everything.

Further reads

• AI Agent Use Cases for Small Businesses: 10 Worked Examples — Ten worked agent examples to price against.
• Can ChatGPT Agent Handle Business Admin Tasks Unsupervised? — What ChatGPT's agent mode can handle unsupervised.
• How to Prepare Your Small Business for AI Agents — Get systems and permissions ready before spending.
• How Much Does a Custom AI Assistant Cost a Small Business? — Costs for assistants that answer rather than act.
• How Much Does an AI Receptionist Cost for a Small Business? — The phone-answering agent priced in detail.
• Zapier vs Make vs n8n for AI Automation: Which Fits Your Business? — Compare the platforms many small-business agents run on.
• What Are AI Marketing Agents and Should a Small Business Use One? — What separates a marketing agent from a chatbot or automation, the agents small businesses can switch on now, what they cost and the guardrails to set.
• What Is Copilot Studio and Does a Small Business Need It? — What Microsoft's agent-building tool actually does, how its credit billing adds up, and the four situations where a small firm genuinely needs it.
• What Is an AI Browser Agent, and Should Staff Use One? — What an AI browser agent does, which jobs it suits, how prompt injection can hijack it, and the settings and staff rules that keep mistakes small.
• What Does an AI SDR Cost, and What Does It Actually Do? — Published AI SDR prices from about $1 a lead to $3,750 a month, the mailbox and staff costs around them, and how to work out cost per booked meeting.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI help centre (ChatGPT Work, custom GPT retirement), Zapier pricing page (Agents activities), HubSpot knowledge base and pricing (Customer Agent credits), Intercom Fin pricing, Jobber Receptionist pricing, Microsoft Copilot Studio pricing, Anthropic and OpenAI API pricing; checked September 2026. Business figures are illustrative.