An AI policy for contractors is not a document problem. It is three decisions: whose account they work in, what the contract says about the work they hand back, and who spends forty minutes showing them your rules on your material. Nearly every template you can download covers contractors in a single scope sentence — “this policy applies to all employees, contractors, interns, and third parties acting on our behalf” — and then spends four thousand words on controls that only exist for employees. The sentence is true and it binds almost nothing.
One clarification, because the search term is ambiguous. If you sell AI systems to federal agencies, “AI policy for contractors” means OMB’s acquisition memoranda and the GSA contract clause — a different article. This one is for the company on the other side. You hire freelancers, fractional staff, and agencies, you have an AI use policy for your own team, and you have realized that some of the people touching your material never signed it.
Why the scope sentence binds almost nothing
Every control an employee policy relies on assumes employment. Approved-tool lists are enforced by device management you do not own on a freelancer’s laptop. Prohibited-data rules are backed by a data-loss tool that never sees their browser. The consequences section points at a disciplinary process that does not apply to someone you pay by invoice.
Even the acknowledgment checkbox usually lives in an onboarding system contractors are kept out of, for sound employment-classification reasons.
What remains is the contract and the account. Those are the only two instruments that reach a non-employee, and both have to be chosen rather than inherited. This is the same gap that opens between a policy and actual behavior inside your own team, which we worked through in the safe AI use playbook — with contractors it is simply wider, because you cannot fall back on management.
The scale is why this stopped being an edge case. MBO Partners put the US independent workforce at a record 74.9 million in its 2026 State of Independence study, and Upwork’s Future Workforce Index 2026 found skilled freelancers now make up 38% of US knowledge workers, up from 28% a year earlier. If your policy covers only badged employees, it may cover a minority of the people producing your work.
Four ways a contractor is not an employee
| Difference | What breaks | What you can still use |
|---|---|---|
| Not on your identity provider | No single sign-on, no device management, no central logging of their use | A seat in your account, or a contractual tool restriction |
| Works in their own tools | Your approved-tool list is advisory | Named approved and prohibited tools in the statement of work |
| Paid for output, not hours | Nobody pays them to read your policy | A paid onboarding line item |
| No disciplinary path | “Consequences up to termination” means cancel the contract | Cure periods, withholding, indemnity |
Take the fourth row. With an employee, a first AI mistake is a coaching conversation. With a contractor it is a contract event, which means you either overreact or you do nothing.
Most companies do nothing, which is how a policy becomes decorative.
The third row is the one people resist. A freelancer reads the clause they are asked to sign and nothing else. If you want them to know which fields in your customer records can never be pasted into an assistant, that knowledge has to arrive in paid time.
Pick the posture before you pick the clause
There are only three defensible arrangements, and the clause you need follows from which you choose.
| Posture | What it means | Best when | Real cost |
|---|---|---|---|
| Your tenant | Contractor gets a seat in your business AI account and works there | Long engagements, your data, sensitive material | Seat cost, a domain decision, offboarding discipline |
| Their tools, constrained | They use their own AI subscription within named limits | Short projects, their own working material, deliverable-based work | Weak visibility; relies entirely on drafting |
| No AI on this engagement | Contractual prohibition on AI in the deliverable or the process | Regulated data, disclosure-sensitive output, unresolved rights questions | Slower, more expensive, often unverifiable |
Most companies default to the middle option without naming it, which is the worst of the three in practice: no seat, no visibility, and a clause nobody read. Choose the first for anyone touching customer or client material for more than a month. Choose the third only where you can say what you would actually inspect, because an unverifiable prohibition is a liability dressed as a control.
The posture question is the account question. The argument for putting contractors on a paid business account rather than tolerating personal logins is the one we make for staff in the case for paying for enterprise AI accounts, and a consumer login held by someone outside your company is both problems at once.
The contract language that survives the second hop
Five clause elements carry the weight. Four of them now appear routinely in contracts and do different jobs:
- Tool restriction. Names the approved tools and the prohibited ones. Enforceable only if the named list is short and current — a list referencing a tool nobody uses signals a template.
- Data restriction. States what may never be entered into any AI system, by data category, not by example. This is the clause that does the most work and the one most often left vague.
- Disclosure. Requires the contractor to tell you where AI was used in the delivered work. Ask for the useful version — which parts and which tool — not a prompt log you will never read.
- Warranty and indemnity. The contractor warrants the deliverable does not infringe third-party rights, including rights arising from AI use, and carries the consequence if it does.
The fifth is the one templates miss. When you hire an agency, your clause binds the agency; the person typing is their subcontractor. Without an explicit flow-down obligation requiring equivalent terms downstream, your data restriction stops at the agency’s front door.
Ask for it in writing, and ask who specifically will do the work.
One more trap: your master agreement, the statement of work, your data processing agreement, and whatever policy document the contract incorporates by reference frequently say different things about AI. Contractors notice this before you do.
Read your existing confidentiality terms first. Most of them already prohibit what you are about to prohibit again, as we covered in what your NDAs already say about pasting client material into AI. Drafting a new AI clause that contradicts an NDA you signed in 2023 creates an argument, not a control.
If your contractors handle your clients’ material rather than yours, the obligation runs further still, and the professional-services boundaries in the rules of engagement for client confidentiality apply to your subcontractors as fully as to your staff.
The two admin settings that decide the outcome
“We will just add them to our account” is where the plan usually dies, for two checkable reasons.
Allowed email domains. On Claude’s Team and Enterprise plans, a new member’s email must use one of your organization’s allowed domains, and member-to-member invites follow the same restriction. There is no documented guest role.
A freelancer on a personal address therefore cannot be added until someone allows that domain — which, for a consumer mail provider, means allowing every address at it. The clean answer is to issue contractors an address on your own domain for the engagement, which also gives you a revocation point.
External domain invites. On ChatGPT’s business tiers, inviting someone outside your domain requires a workspace owner to switch on external domain invites first, and the setting applies only to new invitations. Turn it on deliberately, for named people, rather than discovering it mid-deadline.
Both platforms reclaim the seat the moment you remove the member, which makes offboarding the easiest control here and the one most often skipped. Put contract end dates in the same calendar as seat review. If an engagement ended in March and the seat is still live in September, your access log is fiction. That is exactly the finding that surfaces when a client or insurer sends the questionnaire described in what your insurer and clients now ask about your AI controls.
The forty minutes nobody schedules
Here is the step that separates a policy from a practice: a short, paid, applied session for the contractors who will work on your material.
Not your annual security module. Not a slide deck.
Forty minutes on your own documents, covering four things: the data categories that never go into an assistant, which account they work in, what verification you expect before a deliverable comes back, and who to tell when something goes wrong. Run it quarterly as an open session and point new contractors at the next one.
This is not optional in Europe. EU AI Act Article 4, applicable since 2 February 2025, requires providers and deployers to take measures supporting AI literacy among “their staff and other persons dealing with the operation and use of AI systems on their behalf.” The 2026 Digital Omnibus softened the verb — the original “ensure… a sufficient level” became “support the development of” — but the population covered did not change, and that phrase reaches contractors by name.
The obligation is capability, not a signature. NIST’s AI Risk Management Framework points the same way in its governance function, which asks for policies and procedures addressing AI risks arising from third-party entities, software, and data.
AI enablement is the whole point here, for contractors as much as for payroll staff, and it is why we keep arguing that buying licenses is not the same as building capability. A contractor who understands why a rule exists will apply it to the case your clause did not anticipate. A contractor who signed a clause will not.
What this costs for twenty contractors
| Line item | Market range | Notes |
|---|---|---|
| Business AI seats | Roughly $20–30 per seat per month | Claude Team lists $20 per seat billed annually, $25 monthly; ChatGPT Business lists the same, as of October 2026. Verify before budgeting |
| Enterprise tier | Seat minimums apply | Claude Enterprise lists a 20-seat minimum, which can exceed the contractor count you are solving for |
| Contract amendment | A few hours of counsel time | One reusable clause set, not twenty one-off drafts |
| Paid onboarding time | Forty minutes at their rate | The cheapest line here and the one that gets cut first |
| An applied AI enablement session | Workshops run roughly $2,000–15,000 in the open market | Depends on group size, custom material, and whether it covers staff as well |
A twenty-person pool typically runs four figures a year in seats, plus a one-time legal and session cost. Weigh that against one confidentiality incident involving a client’s data.
A thirty-day sequence
Days 1–5. List every non-employee who touched your material in the last quarter. Include agencies, and name the individuals inside them — that list is usually longer than expected.
Days 6–12. Assign a posture to each engagement. Long and sensitive goes in your tenant; short and self-contained stays in their tools with constraints; regulated stays off AI entirely.
Days 13–20. Write one clause set covering tool restriction, data restriction, disclosure, warranty, and flow-down. Add it to your standard statement of work rather than renegotiating live contracts, except where the data justifies the conversation.
Days 21–25. Turn on what you need: allowed domains or external invites, issued addresses for tenant contractors, seat review tied to contract end dates.
Days 26–30. Run the first session. Record it so the next contractor watches it in week one instead of month four.
Frequently asked questions
Can I just add a freelancer to our ChatGPT or Claude account?
Not by default. ChatGPT’s business tiers require a workspace owner to enable external domain invites first, and Claude’s plans require the member’s email to sit on an allowed domain with no documented guest role. The cleanest path is issuing the contractor an address on your own domain for the engagement.
Does our AI policy legally cover contractors if the scope section names them?
Naming them in scope is necessary but it is not the binding instrument. A non-employee is bound by what they sign. Put the obligations in the statement of work or master agreement, and treat the policy as the document the contract incorporates by reference.
What if the contractor refuses to use our account?
That is a legitimate position for an independent professional with their own stack, and it tells you to pick a different posture. Keep them in their own tools with a tight data restriction and a disclosure obligation, and do not route sensitive material through that engagement.
Should contractors have to disclose when they used AI?
Yes, at the level of which parts of the deliverable and which tool. Prompt-level logging sounds thorough and gets ignored by both sides. A disclosure you will actually read is worth more than one you will not.
Do I have to pay contractors to attend AI training?
If you want them to attend, yes. They bill for time and will not absorb unpaid hours reading your material. Treating those forty minutes as free is why most of these sessions never happen.
Does the EU AI Act apply to our freelancers if we are a US company?
It can. Article 4’s literacy obligation reaches “other persons dealing with the operation and use of AI systems” on a deployer’s behalf, and the Act’s territorial scope can pull in US companies whose systems or outputs are used in the EU. Confirm your exposure with counsel rather than assuming distance protects you.
How do we handle an agency instead of an individual?
Bind the agency and require equivalent terms to flow down to anyone they assign, including their own subcontractors. Ask who specifically will do the work. Your data restriction is only as good as the furthest hop it reaches.
What belongs on the prohibited-data list?
Categories, not examples: anything identifying a customer or patient, anything covered by a confidentiality obligation you owe someone else, credentials and keys, unreleased financials, and personnel matters. Five categories people remember beat twenty they skim.
How often should this be revisited?
Twice a year, and whenever you change AI vendors, add a regulated client, or renew a major agreement. Tool names age fastest, so the clause should reference a maintained list rather than hard-code products.
Is a signed AI clause enough for a SOC 2 audit?
It is the evidence an auditor asks for first, and it is rarely sufficient on its own. Expect follow-up on how you verify compliance, how access is revoked at contract end, and whether the people in scope received any instruction. The clause proves intent; the seat log and the session prove operation.
Key takeaways
- A scope sentence naming contractors is not a control. The contract and the account are the only instruments that reach a non-employee.
- Choose a posture first — your tenant, their tools under constraint, or no AI on the engagement. The clause follows the posture.
- Five clause elements carry the weight: tool restriction, data restriction, disclosure, warranty, and flow-down to subcontractors.
- Two admin settings decide whether “add them to our account” works: allowed email domains and external domain invites. Both default to blocking, correctly.
- Seat reclamation on removal makes offboarding your easiest control. Tie seat review to contract end dates.
- Pay for forty applied minutes. EU AI Act Article 4 asks for literacy among people acting on your behalf, and a signature is not literacy.
If you are working out which posture fits your contractor pool, or what a short applied session for a mixed group of staff and freelancers would look like, book a free AI-readiness call. We will go through your engagement list and tell you where the real exposure sits.
Dirk Jan van Veen, PhD
