When an SMB (small and medium-sized business) owner asks how much an AI agent costs, the answer they get nine times out of ten is "it depends." Followed by a useless range: between $5,000 and $500,000.
That's the worst possible answer. It dodges the conversation instead of having it.
Here's the real answer, in three layers: the pricing models you'll encounter, the variables that actually move the price, and a breakeven calculation you can run tonight on the back of a napkin.
A heads-up before we start: everything that follows is about custom AI agents, not chatbots. The distinction isn't jargon. It's what separates software with a price listed online from a project priced off your own systems.
Chatbot and AI agent: two products, two pricing grids
A lot of people hear "AI agent" and picture a slightly smarter chatbot. Common mistake, and one that throws off every price calculation that follows.
A chatbot answers. It does one thing, in one channel (your site, Messenger, WhatsApp), with one decision at a time. Botnation, Voiceflow, Intercom Fin: these are SaaS (software as a service) platforms sold as standard subscriptions. Several list their prices online: Botnation offers plans from €39 a month, and Intercom charges $0.99 USD for each conversation its Fin agent resolves.
An AI agent accomplishes a multi-step goal. It reads an email, fetches data from your CRM (customer relationship management software), generates a quote, sends it for approval. The chatbot replies. The agent executes. If you want the technical version of that distinction, we cover it in What is the difference between an AI agent and ChatGPT?.
For the rest of this article, we're talking about agents.
The three pricing models you'll encounter
1. The "build": one consultant, one big check
A consulting firm delivers your agent in 2 to 6 months, hands you the code, bills you for the project in one shot. After delivery, you're on your own. If Anthropic ships a new model, you pay for the migration.
Upside: you own everything. Downside: an agent built in 2026 will be worth 60% of its value by 2027. The pace of model improvement makes maintenance non-optional. It's the bulk of the work, not a footnote.
2. The chatbot SaaS platform
As mentioned above, this isn't really the same product category, but it's what most SMBs compare against by default. Practical for automating an FAQ. Not fit for purpose the moment you need to integrate three systems or make a decision that falls outside the script.
3. The custom subscription
This is the Peich model, and the one that's becoming standard among specialist studios in 2026. You pay a monthly fee that covers initial development, hosting, ongoing iteration, and migration to the best models as they ship. No big upfront check, no technological orphaning six months after delivery.
How much that monthly fee comes to depends mostly on the five variables below. That's why it gets worked out from your case rather than read off a rate card.
The five variables that move the price of a custom agent
Once you're in subscription territory, two agents can cost three times one another based on these five factors.
Number of integrations. An agent that touches a single system (your inbox, for instance) is trivial to wire up. The same agent that has to read your CRM, write to QuickBooks, drop a file in Google Drive and post in Slack is five points of friction and five surfaces that can break when a vendor changes their API (application programming interface, the technical layer through which two software systems talk to each other).
Monthly action volume. An agent processing 200 emails a month lives in a workshop economy. The same agent at 50,000 actions a month enters infrastructure economy. You have to think about caching, queuing, resilience. The code isn't the same, and neither are the costs.
Judgment complexity. Sorting emails into three clean categories is a simple agent. Deciding whether a vendor invoice should be auto-approved or escalated to a human based on twelve criteria is an agent that requires chained evaluations, more tests, and more rounds of post-launch tuning.
Desired autonomy level. The more the agent has to handle ambiguous cases on its own without human intervention, the more you have to invest in escalation logic, guardrails, and decision auditing. An agent that escalates at the slightest doubt costs less to build than one that has to guess right 95% of the time.
Delivery speed. An agent shipped in 2 to 4 weeks for a simple case is a normal pace. The same agent in one week is doable, but it means sequencing technical debt to pay it off later. Speed isn't free. It's just accounted for differently.
The hidden cost everyone gets wrong
Ask ten SMB owners: what share of an agent's total cost do you think the AI model API calls represent? Many will answer 30 to 50%. That was true in 2023.
In 2026, it's not even close to true.
At the prices published in September 2026, Claude Haiku 4.5 (Anthropic's economy model) costs $1 USD per million tokens in and $5 USD per million out. A token is roughly three quarters of a word. Claude Sonnet 5 runs $2/$10 and Claude Opus 5.5 $4/$20. On the OpenAI side, GPT-6 Luna costs $0.10/$0.50 per million tokens, and GPT-6 Sol $2/$10, all in US dollars.
To estimate this line yourself, multiply three numbers: tasks per month, tokens per task, and the model's price. Take volumes made up for the example: 300 tasks per business day, about 6,300 a month, with 6,000 tokens read and 1,000 written per task. On Haiku 4.5, the model bill lands around $70 USD a month. On Sonnet 5, around $140 USD. On Opus 5.5, around $280 USD. Tens of dollars for a simple case, a few hundred when the judgment calls for a stronger model.
In the TCO (total cost of ownership, the sum of everything you pay to have and keep the agent alive) of an agent in 2026, this line weighs little. The rest is human work: design, integrations, monitoring, ongoing iteration, and the time your team spends reviewing what the agent produces.
Concrete implication: if a vendor sells you an agent and justifies their price by "AI costs," that's a tell that they haven't read the 2026 market. The cost isn't in the model. It's in the human work around the model.
The breakeven calculation, in one minute
Here's the formula you can run mentally tonight, with the quote you have in hand.
Take the employee who currently does the task an agent would replace. Calculate their fully loaded hourly cost: salary ÷ 1,800 hours per year, times roughly 1.3 for payroll taxes and benefits. For an employee at $60,000/year, that's about $43/hour fully loaded.
Divide the agent's monthly fee by that hourly cost. The result is the number of hours per month the agent has to free up just to pay for itself.
Compare it to what the task consumes today (hours per week × 52 ÷ 12), minus the time someone will spend reviewing the agent's work and handling its escalations. Those hours stay on your bill; they just move to someone else's desk.
An example, with a $1,000 monthly fee picked because it's round, and which is nobody's rate. At $43/hour, the agent has to free up about 23 hours a month. A receptionist spends 8 hours a week sorting and rerouting emails, about 35 hours a month. Count one hour a week to review the cases the agent escalates, about 4 hours a month: that leaves 30 hours freed against 23 to pay for. The math holds, and the 7-hour gap goes back to higher-value work.
If the hours to free up exceed what the task actually consumes, the math doesn't work, however good the agent is. Above the threshold, it improves fast.
What an AI agent subscription should include
A monthly fee means nothing until you know what it covers. Whoever the vendor is, here's what a custom agent subscription should include, and the question to ask about each line.
Design and initial development. Mapping the task, writing the logic, building the tests that show the agent decides the way your best employee would. In a subscription, this work is spread across the monthly fee instead of billed upfront. Ask whether there's a separate setup fee and what it covers.
Integrations. Every system plugged in (CRM, accounting, email, storage) costs something to build and can break later. Ask who fixes it when a vendor changes their API, and whether that's covered by the monthly fee.
Inference. The model fees worked out above, plus the hosting around them. Ask up to what volume they're included and how overage is billed. The answer should fit in one sentence.
Monitoring and iteration. Tracking errors, tuning the agent on the edge cases of the first months, migrating it to a better model when one ships. That's what separates a subscription from a project delivered and forgotten. Ask how often the agent gets reviewed and who decides on a migration.
What stays on your side: review time. No subscription includes the hours your team will spend validating the agent's decisions and receiving its escalations. They're real, especially in the first months. Put them in the breakeven calculation, not in the savings column.
Peich doesn't publish prices: each agent is quoted after a first call, once we've seen your systems and your volume.
When an agent isn't worth it
Three situations where you should say no, even if the budget exists.
If the task takes fewer hours than the threshold you worked out above, the breakeven math doesn't hold. Wait until you've saturated a team on a specific activity before automating.
If the task lives in a system that exposes no API and requires manual clicks at every step, integration cost explodes. An honest audit will tell you "not before that system is replaced."
If nobody on the team is ready to handle the agent's escalations when it hesitates, it's the agent that ends up either blocking or making bad decisions silently. Technology doesn't fix a human governance problem.
Where to start
The first conversation is free, no commitment. 30 minutes to map your repetitive tasks, identify the one with the best cost-benefit ratio, and estimate a real budget for your case. If we conclude the moment isn't right, or that another solution would be a better fit, we'll tell you frankly.
Written byXavier Peich