You Already Know How to Buy AI Agents

When a company needs more hands, it can hire, bring in a temp, or outsource the whole function. AI agents arrive through the same three doors:¹

1. Client-owned digital labor.
   You license or build the agent and run it in-house.
2. Leased digital labor.
   You rent agents from a third party, “in a manner akin to traditional temp staffing.”
3. Fully outsourced AI subdepartments.
   A vendor runs an entire process, such as order fulfillment or a call center, with AI and a small team of human experts.

At least one staffing group is reorganizing around the idea. In 2025 the Adecco Group launched r.Potential with Salesforce to help companies “manage human and digital labour at scale.”² By August 2026, Adecco said half the revenue in its Adecco unit was already “agent-enabled,” in the company’s term, and it raised the goal to 70 percent by year-end.³

The comparison has a limit inside the company. Treating an agent as an employee blurs who is accountable. As a way to decide how to buy agents, it is more useful.

What the staffing playbook already answers

Procurement and HR teams already answer these questions for people. In our reading, many of the answers carry over.

Who trains it. Our framing: an employee learns from you, a temp’s specifics come from your team, and an outsourcer trains its own people. Agents can split the same way, depending on which door they came through.

Who handles the exceptions. The authors give the example of an AI customer-service agent that gets a complaint about a billing dispute above a set dollar amount. A rule might route it to a human specialist. It is the kind of rule a manager might write for a new temp.

What you pay for. The authors’ point is that with agents, “You’re buying an outcome that might come from a combination of people and AI.” In our reading, the comparison procurement makes shifts from hours of work to results.

When each one fits. The authors suggest leasing for seasonal spikes, and full outsourcing for functions that are repetitive but critical.

• • •

Where the comparison breaks

A temp who spends six months on your team learns where the work gets stuck. When the contract ends, that know-how leaves with the person.

The authors list the agent version among the questions to resolve: “When AI is trained on your proprietary data, who owns the resulting capabilities, you or the owner of the AI model or agent?” They also note that “AI’s value compounds over time as it learns from interactions.”

In our reading, an agent’s know-how lives in software: the instructions, examples, and corrections your team writes over months. Each door puts that software in a different place, with a gain and a cost.

1. Owned.
   The know-how stays with you. So do the cost of building it and the work of keeping it current.
2. Leased.
   Fast to start and easy to scale down. The instructions and corrections may sit in the vendor’s system.
3. Outsourced.
   The vendor brings what it learned running the same process elsewhere. What it learns running yours accumulates on its side too.

A vendor’s promise not to train its model on your data is about the model. A separate question for the vendor is who keeps the instructions and corrections stored in its system.

The authors’ advice is to negotiate contracts that let you capture the improvements while respecting the vendor’s intellectual property or your data sovereignty. One test is export: can what your team taught the agent leave with you, in a form another system can use?

What’s hard

Export is easier to write into a contract than to use. Instructions written for one vendor’s platform may not run on another’s without rework.

Owning has its own cost. The know-how stays in-house only if someone in-house maintains it.

Liability strains the analogy too. The authors ask who bears it “if the AI makes a mistake,” and a staffing contract has a person and an agency to point to. With agents, the contract has to name who answers.

What a temp learns leaves with the temp. What an agent learns stays with whoever holds the system, and it can be copied.