What an AI Agent Costs Per Month: The Full Bill, Itemized

What does an AI agent cost to run each month?

About $47 a month in pure infrastructure, for a small WhatsApp agent handling 1,000 conversations — at July 2026 published prices, before anyone gets paid to watch it. That number surprises people in both directions: lower than feared on tokens, higher than expected once hosting and monitoring show up as real lines instead of afterthoughts.

A production AI agent is software that holds a real conversation with a customer, checks or writes to a live system such as a calendar, and replies without a person typing every message. Running one is not a single bill; it is a stack of separate line items — model tokens, message delivery, always-on hosting, observability, and the APIs it calls — each metered by a different vendor.

This article itemizes that stack with a live price and a retrieval date on every line, then works the arithmetic so you can swap in your own volume. It does not price the one-time build — that comparison, and the market bands for hiring someone to do it, live in our cost guide for small businesses. Two sibling teardowns already priced the marginal cost of one conversation, cent by cent: what one AI booking costs and what one AI-answered call costs. This page is the monthly rollup those two deliberately leave out — the fixed costs that show up whether you handle one conversation a month or ten thousand.

What are the monthly line items for a production AI agent?

Six line items, and two of them are usually zero. Modeled assumptions, stated plainly: 1,000 conversations a month over WhatsApp, each shaped like the conversation priced in our booking-agent teardown — five session messages plus one next-day reminder, four model calls — running on Claude Haiku 4.5. At July 2026 published prices, retrieved July 8, 2026, that stack totals $46.55 a month.

The token line is the one everyone worries about, and it is the smallest real cost on the sheet. Claude Haiku 4.5 costs $1 per million input tokens and $5 per million output tokens; our own architecture teardown prices one booking conversation on Haiku at about $0.007 in model fees. At 1,000 conversations a month, that is $7. Swap in OpenAI's GPT-5.4-mini at $0.75/$4.50 per million tokens and the same conversation drops to about $0.004 — roughly $4 a month.

Messaging costs more than the model. Twilio charges $0.005 per WhatsApp message, inbound or outbound; Meta adds $0.0034 per template message sent outside the 24-hour customer-service window, and nothing for free-form replies inside it. A conversation shaped like the one in our booking teardown — five session messages plus one next-day reminder template — runs about $0.0334 in messaging fees. At 1,000 conversations, that is $33.40, the single largest line on the sheet. The number carries its own flat fee too: $1.15 a month for a US local Twilio number, the one your WhatsApp sender is registered on — charged whether you send one message or a hundred thousand.

Hosting is the line most back-of-envelope estimates skip entirely. Something has to receive the webhook, call the model, and stay awake to do it again — a free tier that sleeps after a few idle minutes will silently drop a customer's message. DigitalOcean's App Platform prices its smallest always-on instance — 1 vCPU, 512 MiB of memory, 50 GiB of bandwidth — at $5 a month, plenty for a webhook receiver handling a few conversations an hour.

Observability is the other line nobody quotes, and at this volume it is free. LangSmith's Developer plan costs $0 a seat and includes up to 5,000 base traces a month, with pay-as-you-go beyond that at $2.50 per 1,000 traces; the Plus plan is $39 a seat for up to 10,000 traces. A 1,000-conversation month logs somewhere between 1,000 and 4,000 traces, depending on whether you log per conversation or per model call — inside the free tier either way. Langfuse's free Hobby tier is more generous still, 50,000 units a month, with its Core tier at $29 a month for 100,000 units if you outgrow LangSmith's structure. Free now does not mean free forever: cross 5,000 traces on the Developer plan and every trace after that is billed.

The calendar line is the easiest to verify wrong, because everyone assumes it costs something. Google's own documentation states that "all standard use of the Google Calendar API is available at no additional cost," with a default quota of 10,000 requests a minute per project — a small business will not get near that ceiling. Most small-business CRM APIs follow the same pattern: HubSpot's free tier allows 100 requests per 10 seconds and 250,000 requests a day per account, free unless you are already paying for seats for other reasons.

What does a representative agent cost per month, all in?

$46.55 a month, all in, for the modeled scenario above — a WhatsApp booking-and-support agent on Claude Haiku 4.5, handling 1,000 conversations, hosted on one small instance with free-tier observability. The arithmetic: $7.00 in tokens, $33.40 in messaging, $1.15 for the number, $5.00 for hosting, and $0 for observability and the calendar API.

That $46.55 is a modeled assumption stacked on published unit prices, not a measurement of anyone's real traffic — the volume is ours to assume, the prices are not. Halve the volume to 500 conversations and the bill does not halve to $23. It only drops to about $26.35, because $6.15 of it — the number, hosting, observability's free-tier headroom — is fixed and does not move with volume. Double the volume to 2,000 and those same fixed lines barely register against $81 of tokens and messaging.

This total is deliberately not the number in our per-conversation teardowns. Our booking-agent architecture teardown and our AI appointment booking cost breakdown price the marginal cost of one conversation — $0.04 to $0.06, depending on the model — and explicitly exclude hosting and monitoring. This page adds those back in and rolls the whole thing up to a monthly figure, because "what does the next conversation cost" and "what do I owe this month" are different questions with different answers.

How does that compare to a subscription AI agent platform?

Dramatically more, at the volumes a subscription is built for. Fin, a usage-based AI support agent sold independently of any specific helpdesk, prices at $0.99 per billable outcome — a resolution or a configured handoff — with a published base plan of $49 a month covering the first 50 outcomes. That base plan alone already costs more than the entire $46.55 DIY stack itemized above.

Fin's pricing page and help documentation set the mechanics: $0.99 for every billable outcome beyond the included 50, no extra platform fee if you pair Fin with a non-Intercom helpdesk such as Salesforce or HubSpot, and $29 per seat a month if you route it through Intercom's own helpdesk instead. Fin's help center defines a Resolution as an outcome the customer confirms, or does not ask further about, after Fin's answer, and a Procedure handoff as Fin completing a configured handoff to a person or workflow.

Neither Fin nor Intercom publishes what share of conversations become a billable outcome, so treat the scaling math below as our own modeled range, not a vendor figure. At 1,000 monthly conversations, a conservative one-in-five conversion — 200 outcomes — puts Fin at $197.50 a month: 150 outcomes beyond the included 50, at $0.99 each, plus the $49 base. A strong seven-in-ten conversion — 700 outcomes — puts it at $692.50. Either end of that range runs four to fifteen times the DIY infrastructure total, and neither end includes the setup and tuning work a subscription still expects you to do.

When does a subscription actually beat self-hosting?

When you price in the people, not the infrastructure. The $46.55 DIY total above is server bills only — it assumes the system already works and nobody is watching it. Add professional monitoring at the published $500–$2,000-a-month band our cost guide cites for the simplest custom-built agents, and the self-run total becomes $547 to $2,047 a month — which meets or beats Fin's modeled $198–$693 range, not loses to it.

That is the honest scoping this page owes you: the token bill is the cheap part; hosting and babysitting are the line items nobody quotes. A subscription is not paying for infrastructure you could build cheaper yourself — infrastructure was never the expensive part. It is paying someone else to own the monitoring, the prompt fixes when a model updates, and the on-call pager, bundled into one price per outcome instead of a separate line item for a contractor's time.

A subscription wins outright when you have nobody in-house to own the babysitting, when your volume is low enough that $49 a month easily beats hiring anyone at all, or when you need a capability — CRM-native routing, a shared support inbox, sales qualification — that would otherwise be its own build. Self-hosting wins when volume is high enough that per-outcome pricing dominates the bill, when you already have engineering capacity that would otherwise sit idle, or when your requirements — data residency, a specific calendar system, a home-grown CRM — do not fit any vendor's flow builder. Build vs. buy has its own full crossover math if you want to run your exact numbers rather than these illustrative ones.

Who should not run their own AI agent infrastructure?

Anyone without a name attached to "who watches this." Self-hosting turns you into the vendor — you own the on-call, the security patches, the day a WhatsApp template gets rejected, and the day a model deprecation quietly breaks a prompt overnight.

Concretely, do not self-host if your monthly volume is low enough that Fin's own $49-a-month floor already undercuts our $46.55 modeled infrastructure total once you value your own setup time; if nobody on the team can own monitoring and incident response as a real, assigned job; if you need compliance guarantees — HIPAA, SOC 2, data residency — that a self-run stack on a $5 instance does not remotely provide; or if you already pay for a helpdesk that includes an AI agent you have simply not turned on yet.

Gartner's read on this is blunt and worth repeating exactly. Its June 25, 2025 press release states: "Over 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls." Escalating costs come first in that list. A $46.55-a-month infrastructure bill that quietly needs a $500–$2,000-a-month human behind it is exactly the kind of cost escalation that stat describes — not because the tokens got expensive, but because nobody itemized the rest of the bill before signing up to run it.

If you want that itemization done against your real numbers — your volume, your channel, your calendar and CRM — instead of our modeled 1,000 conversations, that is a scoping conversation, not a sales pitch. Book a 30-minute call and we will tell you honestly whether self-hosting or a subscription fits your numbers, including when the honest answer is neither one of us.