All posts
7 min read

What AI agents actually cost a business

Almost every answer to this question is "it depends," which is true and useless. So here are the actual numbers for our own platform: four ways to pay, what each one really costs, what counts against your plan, and where the money goes when an agent works overnight. If you're comparing vendors, the useful part isn't our prices — it's the list of questions at the end that any vendor should be able to answer this precisely.

First, a benchmark that resets the question

Before any of our numbers: Ramp's AI Index, built from aggregated card and bill-pay spend across more than 70,000 US businesses, put the median firm at $11.38 per employee per month on AI in June 2026. The top 10% spent $611, the top 1% $7,450 — per employee, per month.

That median is worth sitting with, because most content on this question is written by people quoting five-figure implementation budgets. The typical business is spending about twelve dollars a head a month. If a proposal you're reading is two orders of magnitude above that, it isn't describing your situation.

The cost has two halves, and vendors blur them

Running agents costs you a platform fee (the software) and model usage (the AI doing the thinking). Blurring those two is how "$X per month" ends up meaning something different in month two. They're separate here, and the second one has four routes — including one that costs nothing.

Route one: free, with a real catch

There is a genuinely free tier of models, and the catch is stated rather than buried. The daily allowances: 50,000 tokens per user per day on the flagship free class, 500,000 on the smaller/faster one, and 25 requests per day on the experimental class when one is offered (none is at the moment). Those reset at midnight UTC.

The catch, verbatim from what you see in the app when you pick one: "Free because your prompts AND the model's outputs are shared with OpenAI and used to train their models. Do not send anything confidential." For a business that is a real constraint, and it is the reason free is a way to evaluate rather than a way to run your support inbox. There is also a platform-wide daily pool: heavy use by other people can exhaust free models for everyone until midnight UTC, which is why you set a fallback model. The full breakdown of what free costs you in data terms is its own post.

Route two: your own API key — the provider's rate, nothing added

If your business already pays OpenAI, Anthropic or Google, paste that key in and agent usage bills to your existing account at your provider's published rate. We add nothing on top — the calls skip our billing entirely rather than passing through it. For a company with existing AI spend and a procurement process already pointed at one vendor, this is usually the cheapest and the simplest route, and it means model costs land on a bill you already reconcile.

Route three: a subscription you already pay for

If someone at your company pays for Claude Code or ChatGPT, an agent can run on that subscription — sign in once and there is no per-token cost at all, because the plan you already bought is doing the work. The limit is the subscription's own: hit its cap and the agent waits or switches to another model you allowed. For a small team with one or two power users already paying for a plan, this is the most underrated line in the whole pricing conversation.

Route four: platform credits

The convenience route — one bill, any model, nothing to set up. Every paid plan includes a monthly allowance: $8 on Solo ($20/mo), $30 on Team ($50/mo), and $180 on Scale ($200/mo). Signing up gets you $2 of credits once you confirm your email, with no card at all, and paid plans start with a 7-day trial.

Two things worth stating plainly, because they're the sort of detail people discover at checkout. Credits bought beyond your allowance carry a 5% processing fee (minimum $0.80) — card processing costs more than 5% on small amounts, so a bare percentage would be sold at a loss. And platform credits include our margin on model usage, where your own API key does not: that difference is the trade you're making between convenience and cost. Annual billing is exactly ten months charged for twelve — "two months free" literally rather than approximately.

What counts against your plan (and what doesn't)

This trips people up, so: a run is a scheduled or triggered job — the 6am digest, the webhook that fires when a form is submitted. Conversations do not consume runs. Messaging your agent on WhatsApp, a customer emailing your support desk, someone calling your phone line — none of those count against the monthly run limit. The free plan's 100 runs a month is therefore about three scheduled jobs a day, not a cap on how much your agent talks to people.

Two other costs are worth naming because nobody mentions them. Transcribing a voice note runs at the provider's rate of $0.006 per minute — six tenths of a cent. And a phone line runs on your own Twilio number at Twilio's rates, billed by Twilio directly to you; we don't resell minutes, so the call itself never appears on our invoice.

You can see the actual number

Every model call, tool call and connector request is recorded append-only with its cost, so "what did that overnight run cost" has an answer to the cent rather than an estimate at the end of the month. That matters more than any headline price: it's the difference between budgeting and hoping. It also distinguishes what was actually billed from what was merely estimated — usage on your own key or a subscription shows up on the same page without pretending it was spend.

The questions to ask any vendor, including us

  • Can I use my own API key, and do you add a margin when I do?
  • What exactly counts against the plan's limits — every message, or only scheduled jobs?
  • Can I see what a single run cost, or only a monthly total?
  • If your free tier trains on my data, where do you say so — in the product, or in the terms?
  • What happens when I hit a limit: does it stop, queue, or silently bill me?

Ours are answered above, which is the point of publishing them. The cheapest way to find your own number is to run one real workflow for a week and read the usage page — the $2 of credits needs no card, only a confirmed email, and if you'd rather have the first workflow set up with you, that call is free too.

See it for yourself — free, $2 of credits once you confirm your email, no card.

Start free