Agent as a service

Most companies that use an AI agent never build one. They rent one, the way they rent email hosting or payroll software: a vendor runs the model, the tools, the safety rails, and the updates, and you plug the agent into your help desk or website. The industry name for this is agent as a service. The interesting part is how you pay — in our view, agents are the first mainstream software priced less like a product and more like an employee.

The three pricing shapes

Almost every offer you'll see is one of three shapes:

• Per seat. A flat monthly price per user, like classic software. Salesforce sells Agentforce this way at $125-plus per user per month [2].
• Per action. You buy credits and the agent spends them as it works. Salesforce's Flex Credits cost $500 per 100,000; each action the agent takes — updating a record, answering a question — burns 20 credits, about $0.10 [1].
• Per outcome. You pay only when the agent finishes the job. Intercom's Fin support agent charges $0.99 per resolved conversation [3]; Sierra prices custom enterprise deals the same way — in most cases, no resolution means no charge [4].

The trade-offs, in plain terms: seats are predictable, but you pay even when the agent idles. Actions track real work, but bills are hard to predict. Outcomes sound fairest — until you read what counts as an outcome. Fin, for example, counts a chat as resolved when the customer simply stops replying after an answer [3]. That isn't a scandal; it's a contract detail. But it's the detail to read first.

One agent can use all three at once. The shapes are not rivals, and some products combine them. Somi, the social media agent from Kadres, is a worked example. Its seat is a brand — one seat buys the agent for one brand's accounts, so a company running three brands takes three seats. Its actions are the real work of making the content and posting it, and they are priced by type, because a reel, a carousel and a video are not equal amounts of work. Its outcome is the post going live: the balance is only drawn down once a post has published successfully to the account it was meant for, so a post that never makes it costs nothing.

Read the definition of "outcome" before you admire the price.

What can go wrong

The analyst firm Gartner predicted in June 2025 that more than 40% of agentic-AI projects will be canceled by the end of 2027, as costs climb, business value stays unclear, and risk controls lag [5]. It also warns about "agent washing" — vendors relabeling ordinary chatbots and automation scripts as agents — and estimates that of the thousands of self-described agent vendors, only around 130 offer the real thing [5].

Four questions that separate real offers from washed ones:

1. What exactly counts as a billable action or outcome — and who measures it?
2. What data does the agent see, and what does the vendor keep?
3. What happens when it's wrong? (The failure modes guide lists what "wrong" looks like.)
4. What would doing it yourself cost? (The cost guide is the baseline for judging any vendor's price.)

Rent or build?

Rent when the job is common — support, lead qualification — and vendors compete hard on it. Build when the agent is your product, or the volume is so high that per-outcome fees dwarf running your own. If you're not sure you need an agent at all, the five-question self-check is the place to start. And our opinion, labeled as such: for a first agent, renting a well-defined outcome is the cheaper lesson.

AiAgentNook editorial · corrections are noted here, not silently edited.