How to Write an AI Agent Co-Employment Risk Clause Before You Scale

Startups swapping contractors for AI agents are walking into a joint-employer exposure nobody has priced yet, and the fix is a specific clause, not a legal footnote.

Quick Summary - TLDR:

• Joint-employer liability can attach when a startup directs the day-to-day work of a staffing vendor's AI agents the same way it would a contractor
• California's AB 5 ABC test and the federal Department of Labor's economic-realities test both turn on control, and an agent you configure and supervise looks a lot like a worker you control
• Delaware General Corporation Law doesn't require a co-employment clause, so the liability only shows up in the contract you sign with your agent vendor or staffing partner
• A usable clause needs four parts: scope of control, indemnification trigger, data and output ownership, and a termination-for-reclassification right
• Uber and Lyft spent nine years and over $200 million in California litigation fighting the same control-test argument that now applies to agent vendors

Here's the thing nobody selling you an AI agent platform tells you: the contract you sign to deploy it probably doesn't protect you from a misclassification claim, because it was written before anyone thought an agent needed to be treated like a worker.

Say you're a 40-person startup and you've replaced three contractor roles, customer support triage, lead qualification, and basic bookkeeping reconciliation, with agents from a vendor like Sierra or Decagon. You still tell the agent what to prioritize, you still review its outputs daily, you still set its hours of operation in practice even if nobody calls it that. That's the exact fact pattern that turns a vendor relationship into a joint-employer one when the thing being directed is a human contractor. Nobody has fully litigated whether it applies the same way to software, but the test that matters, control over the work, doesn't care whether the worker has a pulse.

The federal Department of Labor's economic-realities test asks six questions: who sets the schedule, who controls the manner and means of the work, how integral the work is to the business, whether there's an opportunity for profit or loss, the investment in equipment, and the permanency of the relationship. California's AB 5 ABC test is stricter still, presuming a worker is an employee unless the hiring company proves the worker is free from control, does work outside the company's usual business, and is customarily engaged in an independent trade.

Neither test mentions software. But both were written to catch exactly the arrangement a lot of founders are building right now: a vendor supplies the labor, the client company directs it daily, and everyone calls the relationship something other than employment to avoid the obligations that come with it. Uber and Lyft spent roughly nine years and more than $200 million fighting this argument in California courts and at the ballot box with Proposition 22, and they were arguing about drivers, not algorithms. The moment a plaintiff's attorney or a state labor board decides an AI agent's output is functionally indistinguishable from a contractor's, the control test gets pointed at the agent vendor, and by extension at you, the company directing the agent's work.

This isn't theoretical scaremongering dressed up as legal advice. It's a gap. Delaware General Corporation Law, which governs how most VC-backed startups are incorporated, says nothing about co-employment because co-employment isn't a corporate structure question, it's a contract and labor-law question that lives entirely in the agreement between you and whoever supplies the agent.

What the clause actually needs to say

A co-employment risk clause isn't boilerplate you paste in from a template site. It needs four specific components, and most staffing or SaaS agreements for AI agents currently have zero of them.

First, a scope-of-control provision that states, in writing, who configures the agent's instructions, who reviews its output, and who sets the cadence of its work. If you're the one doing all three, say so plainly rather than letting the contract imply the vendor retains control it doesn't actually exercise. Courts look at what actually happens, not what the contract calls it, so a clause that lies about control helps nobody and actually makes things worse if it's later used as evidence of bad faith.

Second, an indemnification trigger specific to misclassification claims, not just the generic IP and data-breach indemnification every SaaS contract already has. Name the trigger: a regulatory finding, a private lawsuit, or a state labor department inquiry that treats the agent's deployment as an employment relationship. Specify who pays legal costs and who controls the defense. Generic "mutual indemnification" language rarely addresses this because the drafters weren't thinking about it when they wrote the template.

Third, output and data ownership language that's airtight, because a court assessing the ABC test's "usual course of business" prong will look at whether the agent's work product is treated like an employee's work product (owned outright, integrated into core systems) or like a vendor's deliverable (licensed, bounded, separable). The more the agent's output looks indistinguishable from in-house labor, the weaker your independent-contractor argument becomes if a regulator ever tests it.

Fourth, a termination-for-reclassification right: a clause letting you exit the agreement without penalty if a court, regulator, or your own counsel determines the arrangement creates employer obligations you didn't price in, like payroll tax withholding, workers' comp coverage, or benefits eligibility under ERISA.

Why this is landing on founders now, not HR

Replit's CEO Amjad Masad has talked publicly about treating AI agents as part of the company's actual headcount planning, not just a tooling line item, and that framing is exactly where this risk hides. The moment a board deck or an internal planning doc describes agents as replacing N contractor roles, you've created a paper trail that argues your own case against you: you treated the agent as equivalent to a worker, which is the first thing a plaintiff's lawyer or state labor investigator will quote back to you.

None of this means don't use agents, and it doesn't mean every startup deploying one is sitting on a lawsuit. It means the exposure is real, it's underpriced in the contracts most founders are signing today, and it sits in a gap between corporate law, which doesn't touch it, and employment law, which was written assuming the worker had a Social Security number. Fix it in the contract before HR, legal, or a state labor board does it for you.

Also read: How to Forecast AI Agent Costs for Board Deck Without Guessing • How to Negotiate a Most Favored Nation Clause in Your First Enterprise Deal • How to Forecast AI Agent Usage for Pricing When Nobody Can Predict Volume

This article is posted in Startup News, check it out for more related stories.