---
title: "AI Agent as a Service: Buyers Want the Job Done, Not an Agent"
publishedAt: "2026-11-17"
summary: "AI agent as a service usually means renting an agent you still set up and supervise. Buyers want the job done, run by the company that owns the product."
author: "Ben Skoog"
authorImage: "/ben_skoog.jpeg"
---

AI agent as a service (often shortened to AaaS) is a delivery model where a provider hosts and runs AI agents and a business pays for access, usually by subscription or usage, instead of building the agents itself. That is the standard definition, and it is accurate. What it describes, though, is renting an agent. The buyer still decides what the agent should do, connects it to their tools, sets its permissions, and checks its work. Most buyers did not want an agent. They wanted a job done.

We learned that by selling one. Until October 2026, Aimdoc sold an agent. On October 8 we stopped, and started selling the work instead: the jobs a company's customers need done, delivered through the company's own product. The reason was simple. The hard part of every deal was never the agent. It was everything the buyer had to do after buying it.

So the argument of this post is narrower than "agents are overhyped." Agents work. The question is who should run them and what should be sold. Our view: the right seller of an AI agent as a service is usually not an agent vendor. It is the company whose product the work happens in, and the thing it should sell is the job, not the agent.

## What "AI agent as a service" means on the pages that rank

The definitions agree with each other. [Hostinger's guide](https://www.hostinger.com/tutorials/ai-agents-as-a-service) calls it "a cloud delivery model that lets you use hosted or managed AI agents" without building or maintaining what sits underneath. [xcube LABS](https://xcubelabs.com/blog/agent-as-a-service-aaas-the-emerging-business-model-replacing-traditional-saas) describes "a delivery model in which organizations access AI agents through a subscription or usage-based model." Sierra's founders, introducing their own product under the name, put it as "[prompts, not clicks](https://sierra.ai/blog/agents-as-a-service)": customers describe the outcome they want and agents deliver it.

The case for the model is borrowed from a bigger idea. [Sequoia's Sonya Huang and Pat Grady](https://www.sequoiacap.com/article/generative-ais-act-o1/) wrote that "the AI transition is service-as-a-software," and that "Cloud companies sold software ($ / seat). AI companies sell work ($ / outcome)." AaaS pages take that line and conclude that an agent is how you buy work.

That is where the definitions and the reality part ways. Read Hostinger's own setup steps for an AI agent as a service:

1. Choose or configure an agent.
2. Connect the agent to tools and data.
3. Define goals, permissions, and guardrails.
4. Let the agent plan and execute the task.
5. Monitor results and approve sensitive actions.

The provider runs the hosting. The buyer does steps 1, 2, 3 and 5. The same guide is candid that the business still has to review results and investigate failures, and that guardrails reduce wrong actions but cannot eliminate them. That is not buying work. That is hiring a new kind of employee and managing it yourself.

<img src="/blog/ai-agent-as-a-service/who-does-the-setup.svg" alt="Table of the five setup steps Hostinger lists for an AI agent as a service. For a rented agent the buyer does configuration, connecting tools, defining goals and permissions, and monitoring. For a job run by the product's owner, the vendor defines the job once, the product is already connected, the job acts with the customer's permissions, and the vendor owns the result." />

*The five setup steps from Hostinger's guide, and who does each. The right column is our view.*

## Why renting an agent so often stalls

The numbers on agent projects point the same way. [Gartner predicted in June 2025](https://www.outlookbusiness.com/artificial-intelligence/over-40-of-agentic-ai-projects-will-be-scrapped-by-2027-says-gartner): "More than 40% of agentic AI projects will be cancelled by the end of 2027," due to "escalating costs, unclear business value or inadequate risk controls." Its analyst Anushree Verma put it bluntly: "Most agentic AI propositions lack significant value or return on investment (ROI)." Gartner also estimated that only about 130 of the thousands of vendors calling their products agentic are the real thing.

Costs, unclear value and risk controls are not model problems. They are ownership problems. A rented agent has no defined job, so nobody can say what value it was supposed to produce. It acts in someone else's systems, so the risk controls are the buyer's to design. And it runs until someone stops it, so the cost has no natural limit.

Anthropic's own engineering guidance on [building effective agents](https://www.anthropic.com/engineering/building-effective-agents) makes the technical version of the point. It separates workflows, "where LLMs and tools are orchestrated through predefined code paths," from agents, which "dynamically direct their own processes and tool usage." It warns that "the autonomous nature of agents means higher costs, and the potential for compounding errors," and recommends "finding the simplest solution possible, and only increasing complexity when needed." A general agent pointed at a business and told to be useful is the most complex option there is.

## What we saw when we sold an agent

This section is our own experience, not a survey.

When Aimdoc sold an agent, the conversation after the sale was always the same. What should it do first? Which tools should it have? Who checks its work? Each buyer had to answer those questions for their own business, and the ones who answered them well were the ones who already knew exactly which job they wanted done. The agent was the easy part to buy and the hard part to own.

We also looked hard at the opposite of a narrow agent: give a capable general agent every tool a product exposes and ask it to deliver a result for a customer. Our conclusion over the summer was that it would not work well enough to promise anyone anything. Broad agents are hard to tune: every new tool and every new kind of request is another way to drift. What did work was a bounded job with a clear goal, a fixed set of tools, structured inputs and outputs, and a budget. Controllable, testable, and something we could put a name and a price on. We wrote about the tool side of this in [too many MCP tools: why we built services](/blog/too-many-mcp-tools-why-we-built-services).

So on October 8 we changed what we sell. The agent became the way customers ask for work, in the app or over email. The product became the work itself: [services](/services), each one a job the AI owns for each customer and keeps working, like getting every new hire through onboarding or keeping every location compliant. We run our own business this way too. When a company signs up for Aimdoc, its setup runs as a service on Aimdoc, and our own customer success work for our customers runs as services rather than as a person with a checklist.

## Who should run the agent

The objection we heard most in the last few months, from other founders and advisors, was some version of "software companies will just build this themselves." It is a fair objection. Our answer was the same each time: most teams also stopped racking their own servers, not because they could not, but because running that layer well is a specialty. Running agents that act for thousands of customers, each with their own permissions, budget and history, is turning into the same kind of specialty.

The second argument, which we hear from builders, is that everything goes headless: every product becomes a set of tools, and each customer brings their own agent to drive them. Some of that will happen, and it is good. But we do not expect a complex business system to hand every internal operation to an outside agent and hope it orchestrates them correctly. A slice of software will keep the agent on its own side, because the company that built the product knows how the work should be done, holds the data the work depends on, and is answerable to the customer when it goes wrong.

That is the core difference between an agent vendor and the product's owner:

| | Agent vendor (rented agent) | The company that owns the product |
|---|---|---|
| Knows how the work should be done | The buyer has to teach it | Built the product the work happens in |
| Has the data and permissions | Through whatever the buyer connects | Already holds them, per customer |
| Can see whether the result held | Only what it is shown | Sees the state in its own product |
| Who answers when it fails | The buyer, mostly | The vendor |

[Bain's 2025 technology report](https://www.bain.com/insights/will-agentic-ai-disrupt-saas-technology-report-2025/) reaches a similar conclusion from the market side. "If an agent replaces a human task, customers will expect to pay based on outcomes, not log-ons," it says. In categories where "companies hold exclusive data and rules, giving them a head start on full automation," Bain adds that "incumbents should have the advantage, but to keep it they will need to proactively replace SaaS activity with AI." Its advice is to "stop charging for access and start charging for work done."

The table above is also why a separate agent vendor is in an awkward spot. It sells the agent to the buyer, but the buyer's work happens in someone else's product. We covered the general version of this argument in [outcome as a service](/blog/outcome-as-a-service); the short form is that the party who owns the product, the data and the customer is the only one that does not have to ask permission to deliver the job.

## What to sell instead of an agent

If you run a company whose customers need ongoing work done in your product, here is what we would tell you over coffee.

- **Do not sell the agent.** Your customers will not configure it, and you will spend your time helping them do steps 1 to 5 above. Use an agent as the way customers reach the work, not as the product.
- **Sell one job.** Pick the work your best customers already do by hand every week, or already ask your support team to do for them. Define it as a state you can see in your product, like every account set up, every certificate current, every new hire onboarded.
- **Keep it bounded.** A fixed set of your product's actions, a clear goal, a budget, and a person looped in for the decisions that need one. That is what makes it something you can promise.
- **Price the job, not the agent.** Usually a subscription for the job, set by you and charged to your customers. We compared selling a feature, an agent, or finished work in [AI monetization: sell a feature, an agent, or finished work?](/blog/how-software-companies-monetize-ai), so we will not repeat the pricing menu here.

Keep two kinds of money apart while you do this. What you charge your customers for a job is your price, through your own Stripe account. What you pay to have the AI do the work is a cost, and it should be capped and visible per customer.

## FAQ

### What is AI agent as a service?

AI agent as a service is a model where a provider hosts and runs AI agents and a business pays for access, typically by subscription or usage, instead of building its own. In most offerings the buyer still configures the agent, connects its tools, sets its permissions and monitors its results.

### What is the difference between AaaS and SaaS?

With SaaS, people use the software to do the work. With AaaS, an agent does some of the work. In both, the buyer usually still owns the result. Selling a job as a service is the step past both: the provider owns the result for each customer.

### Is agents as a service the same as service as software?

They overlap. Service as software is the broader idea that AI lets companies sell work instead of tools. Agents as a service is one way to package it, by renting the agent. Selling a job, delivered through the product the work happens in, is another, and in our experience the one buyers actually want.

### Who should own the agent: the agent vendor or the software company?

For work that happens inside a product, usually the company that owns the product. It knows how the work should be done, holds the data and permissions, and can see whether the result held.

---

Your customers want the job done, not an agent to manage. Aimdoc connects securely to your product and your systems and delivers that work for each customer, metered, monitored and cost controlled. See how [services](/services) work and how to [sell them through your own Stripe account](/services/sell), or [start free](https://app.aimdoc.ai/signup).
