Contract engineering · Policy template · As of November 2026

Contractor AI Use Policy Template for Regulated Codebases

Approved tools, prohibited data classes, disclosure in pull requests, IP and offboarding, in one practical policy.

Free Word file for a work email. Or read the whole policy on this page.

Engineering manager in a navy sweater and a contract engineer in a grey shirt read one blank page together at a wooden table, a closed laptop and a mug with an orange rim at its edge

A contractor AI use policy template names approved tools, data AI must never see, how AI-generated code is disclosed in pull requests, where IP ownership lives and what gets cleared at offboarding. A practical policy, not a ban.

Last updated: November 19, 2026

Your contractors already use AI tools. The open question is whether you know which ones. A ban doesn’t stop it. It moves the use onto personal accounts you can’t see, which is the worst place for it.

So write the rule down instead. Keep it short. Name the tools, name the data that never goes in, ask for one line in every pull request, point at the IP agreement and clear the history on the last day. That is the whole policy. The rest of this page is the paperwork.

This isn’t a fringe problem. The October 2026 revision of NBER working paper 34836 surveys nearly 9,000 executives in seven countries and reports that 78% of firms use AI, so your engineers are inside that number and so are the contractors you brought in last month.

This template is for teams with a regulated codebase, the lending, fintech, payments and health shops where a customer or an auditor will ask to see it. It assumes contract engineers who arrive through an engineering staffing partner, the way regulated industry software engineers and contract engineering teams usually do. It goes with my post on setting the contractor AI tool policy before the code leaves the building.

Access is a separate document. Use the contractor access policy template for that, and the contractor onboarding checklist for the first day.

Download

Download the Policy Template

One work email and the Word version is yours, with the same text as the policy below and every field shaded so that nothing gets skipped when your policy owner fills it in.

  • Five sections

    Approved tools, prohibited data, disclosure, the IP pointer and offboarding, written as clauses you can adopt as they stand.

  • The data class table

    Six classes with a three-way answer for each, ready to rename to match your own classification.

  • Numbers you will change

    Retention days, report windows and review cycles sit in shaded brackets, like the 30-day setting and the one-hour report.

Where should we send it?

    Work email, please. We use your details to respond to this request. See our privacy policy.

    The template

    Contractor AI Use Policy Template, Section by Section

    Replace every shaded field. Each section starts with one symbol and one plain word, so a contractor can hold the whole policy in their head after one read. Start with the label.

    KORE1 contractor AI use policyTemplate, November 2026

    1. Useapproved tools and accounts
    2. Neverregulated data and secrets
    3. ShowAI help named in the pull request
    4. Ownthe IP assignment still decides
    5. Cleartool history gone at exit
    The policy as a care label. Five symbols in the order the sections come, each with one plain word. Sections 1 to 5 below print every clause.

    Policy

    Contractor AI Use Policy

    Applies to
    [Company name] and everyone who is not an employee and holds a credential to its systems, source code or data, including contract engineers, contract-to-hire engineers, statement-of-work staff and subcontractors
    Policy owner
    [VP Engineering, CTO or CISO]
    Version and effective date
    [1.0], [date]
    Next review
    [six months after the effective date]

    Approved tools and accounts

    On the label Use. Fill in the approved tools list, the tool owner, the review cycle.

    1.1

    The list. Contractors use only the AI tools on [the approved tools list]. Each entry names the tool, the plan, the data classes allowed in it, the owner and the date of the last review. A tool that is not on the list is not approved, however popular it is.

    1.2

    Company accounts only. Contractors sign in with a [Company name] single sign-on account. No personal accounts, no free tiers, and no browser extensions, editor add-ons or plugins that are not on the list.

    1.3

    Settings the owner holds. Each approved tool is set so that [Company name] content is not used to train models and prompts and outputs are kept for no more than [30] days. The tool owner records the settings and checks them every [quarter]. Contractors do not change them.

    1.4

    Agents count. A tool that acts on its own, such as an agent that runs commands, opens pull requests or calls other systems, gets its own entry on the list and runs with no more access than the contractor’s tier allows.

    Access tiers cap what an agent can reach. Our contractor access policy template sets them.

    1.5

    Asking for a new tool. The contractor or the hiring manager files a request that names the tool, the task and the data classes. The policy owner and [the security lead] approve it or refuse it in writing within [five] business days, and an approval adds the tool to the list.

    Prohibited data classes

    On the label Never. Fill in your own class names, the security mailbox, the report window.

    2.1

    One class, one answer. Every item belongs to a class in the data classification policy at [Company name]. The table says what a contractor may put into an approved AI tool for each class. If an item could belong to two classes, treat it as the stricter one.

    2.2

    The table.

    Six data classes and whether each may go into an approved AI tool
    Data classExamplesIn an approved AI toolCondition
    PublicPublished documentation, open source code, public API specifications.AllowedAny tool on the approved list.
    InternalYour own non-regulated code, design notes, tickets and runbooks that contain no secrets.AllowedCompany account only, with the retention and training settings from section 1.
    ConfidentialSecurity architecture, unreleased product plans, vulnerability reports, code under a customer NDA, code under a licence you have not cleared.Approved for this class onlyOnly a tool the policy owner has approved for this class, and only for the task named on the ticket.
    RegulatedCustomer information, cardholder data, protected health information, nonpublic personal information, and the source code of regulated systems.ProhibitedNo, unless the policy owner approves in writing a tool that runs inside your own boundary.
    Production dataLive records, production logs and database extracts, even a single row or one screenshot.ProhibitedUse masked or synthetic data. An employee produces any sample.
    SecretsPasswords, API keys, tokens, private keys, certificates and connection strings.ProhibitedNever, in any tool. If one goes in by mistake, section 2 says what happens next.
    • Allowed in any approved tool
    • Approved for that class only
    • Never goes in

    2.3

    Prompts count. A prompt, a pasted log, a screenshot and an uploaded file are all data. So is an output that repeats what went in.

    2.4

    Report a slip the same day. If regulated data or a secret goes into a tool, the contractor tells [the security mailbox] within [one hour] of noticing. The policy owner decides what to delete, rotate or notify. Reporting a slip the same day is never a disciplinary matter. Hiding one is.

    Disclosure of AI-generated code

    On the label Show. Fill in the pull request template line, the protected branches.

    3.1

    Say so in the pull request. Every pull request states whether an AI tool helped write the change, which approved tool it was, and which files or functions it touched. Put it in the template as one line, for example AI assistance [none, some or most], tool [from the list], parts [files or functions].

    3.2

    A person owns every line. The contractor who opens the pull request reads, understands and tests every line before asking for review, and answers for the change as if they had typed it. “The tool wrote it” is never an answer in review.

    3.3

    What reviewers check. An employee reviews every contractor pull request that discloses AI help. The reviewer checks that every imported package exists and is the one intended, that nothing copied from elsewhere carries a licence you have not cleared, and that no secret or customer data is in the change.

    A review checklist for the same pull requests is in our contractor code review checklist.

    3.4

    Generated tests and documents. Tests, scripts, configuration and documentation made with an AI tool follow the same rule as code.

    3.5

    The record stays. The pull request is the record of where AI contributed. It is kept with the repository for [the life of the code] and is not edited afterward to remove the disclosure.

    IP assignment pointer

    On the label Own. Fill in the name of your assignment agreement, the counsel who reviews it.

    4.1

    Ownership is settled elsewhere. Who owns work product, including anything made with AI tools, is set by [the IP assignment agreement] that the contractor and the staffing partner have signed. This policy points to that agreement and does not change it. Where the two differ, the agreement controls.

    How the agreement itself is written is covered in our guide to contractor IP ownership agreements. This is not legal advice.

    4.2

    Check that it covers AI. Before this policy takes effect, [the policy owner] and [counsel] confirm that the assignment reaches output made with AI tools and that the contractor promises to have the right to hand over what they deliver.

    4.3

    No rights for a tool vendor. A contractor does not accept terms that give a tool vendor rights to [Company name] content or output. [Counsel] reads the terms of every tool before it joins the approved list.

    4.4

    Material that is not yours. A contractor does not paste third-party code into a prompt or a pull request unless [Company name] has cleared its licence, and reports output that looks copied from somewhere else.

    Offboarding of tool context

    On the label Clear. Fill in the purge steps for each tool, the return-or-destroy form.

    5.1

    Before access ends. On or before the last day, the contractor deletes or hands over to [Company name] everything they hold in each approved tool, including chat history, saved prompts, project files, custom instructions, memory entries and uploaded files.

    5.2

    The tool owner purges. Because accounts belong to [Company name], the tool owner then removes the contractor’s seat, purges the workspace content and shared conversations, and revokes the API keys, editor tokens and integrations created under that account.

    5.3

    The device. Local agents, extensions, caches and logs leave with the managed device, which is wiped when it is returned. Nothing stays on a personal device, because nothing should have been there.

    5.4

    Evidence kept. The tool owner’s purge record and a signed return-or-destroy certification that names AI tool content as well as code and data.

    5.5

    The leftover check. Thirty days after exit, list every seat, token, workspace and integration still named for the contractor. Anything still live goes to the policy owner.

    The last-day order of operations, with the places access hides, is in our contractor access offboarding checklist.

    Approval

    Approved by
    [Policy owner name], [title], [date]
    Reviewed by
    [Counsel or security lead name], [date]
    Next review
    [date six months out]

    A policy template, not legal advice. Have counsel read it once for your industry, your customer contracts and your IP assignment before you adopt it.

    Senior engineer in a charcoal sweater sits on a window ledge and talks with a younger contract engineer in a plum overshirt who holds a closed laptop, an orange notebook and a ceramic cup beside him
    Why a rule

    Why Write a Rule Instead of Banning AI Tools?

    Because a ban doesn’t remove the tools. It removes your view of them. Bans fail quietly, and a ban rarely survives a deadline.

    A rule you can audit beats a rule you hope people follow. If the approved tool is easy to reach and the list of things that never go in is short and specific, people follow it, and if the list is long and vague, they stop reading it and go back to whatever they used before.

    So I split the question in two. Which tools may a contractor use? Which data may never go into any of them? The first answer changes every quarter. The second almost never does, so it belongs in the policy itself, while the tool list belongs in an appendix that the owner can edit without calling a meeting or asking anyone to sign again.

    Some teams do ban AI outright for one codebase, a payments core or a signing service. The table still works there. Set every class to prohibited, and say so in section 1.

    None of this is new to anyone who has run engineering through an audit. I have written about how review changes when AI writes part of the code and about getting AI pilots into production. The contractor version adds two facts, people arrive with their own habits and they leave with their own history.

    78%

    of firms use AI, from a survey of nearly 9,000 executives in seven countries, NBER working paper 34836, October 2026 revision

    3 days

    the median wait between an accepted offer and the first shift, counted over 1,997 KORE1 starts

    37%

    of contractors started the same day they accepted, in the same 1,997 starts

    116 days

    typical length of a KORE1 contract, about 17 weeks, across 1,887 assignments

    KORE1 figures come from W2 contract placements dated 2019 to 2026, across all roles, the same records behind the Contract Rate Index 2026-Q4. The 78% comes from the NBER abstract.

    The risk

    What Breaks If AI Tools Leak Regulated Code or Data?

    Four things. None of them needs a headline breach. Each one is a reason a clause above exists.

    • The argument that it was secret

      Trade secret law protects information only when its owner has taken reasonable measures to keep it secret. Unreleased code pasted into a tool with no agreement and no settings you own makes that harder to show, and sections 1 and 2 are the measures you can point to when someone asks what you did.

    • Regulated data with no contract behind it

      HIPAA lets a covered entity disclose protected health information to a business associate only once it has satisfactory assurance that the vendor will safeguard it. A free tool a contractor opened on a personal account gave you none. Cardholder data and customer financial records carry duties of their own.

    • Code you may not be able to protect

      The U.S. Copyright Office’s January 2025 report on copyrightability finds that copyright does not extend to purely AI-generated material and that prompts alone do not give a user enough control to be the author. If nobody records where AI wrote code, you can’t say what you own. Section 3 is that record.

    • A history that walks out the door

      Chat history, saved prompts and uploaded files hold pasted code, schemas and ticket text. On a personal account they leave with the person, and nothing on your side can delete them. That is why sections 1 and 5 insist on company accounts.

    Contract engineer in a white shirt hands a closed laptop with a coiled charger and a small kraft sleeve across a desk to an IT administrator in a slate-blue cardigan, a potted plant and an orange mug at the edge of the desk
    Adopt it

    How Do You Adopt a Contractor AI Use Policy?

    Six moves, in order. Most teams finish the first four in an afternoon.

    1. Name the policy owner and the tool owner.

      The policy owner approves exceptions. The tool owner keeps the list and the settings. One person can hold both at a small company, but write both names down.

    2. Build the approved tools list.

      Start with what contractors already use. Ask them. Then approve, move or retire each tool, and write the data classes allowed in each one.

    3. Rename the six classes to match yours.

      Use the words from your own data classification policy. Can’t say which class a database table belongs to? Fix that first.

    4. Add one line to the pull request template.

      AI assistance, tool, parts. Three fields and thirty seconds.

    5. Check the IP assignment with counsel.

      Does it reach output made with AI tools? If it doesn’t, change the language before the next contractor signs.

    6. Run the offboarding purge on a test account.

      Use an account an employee owns. Time it. Fix whatever takes longer than the last day allows.

    Then put the review on the calendar. Every six months is enough, and again after an incident, an audit finding or a change in your tool list. Change the text? Change the date.

    Start before the offer. The median KORE1 contractor starts three days after accepting, so a policy that waits for day one is already late. Hand over the one-page version with the offer. If the work is bigger than a policy, our AI governance staffing page covers the people who run it, and the knowledge transfer and exit planning guide covers what leaves with them.

    KORE1

    How KORE1 Sources, Vets and Places Contract Engineers for This Model

    This part is us. Not Kris. A policy is only as good as the people who walk into it, so here is what we do on our side of the table.

    • Sourcing

      Engineers who have shipped under audit

      Our regulated bench is screened by engineers who have shipped under SOC 2, HIPAA, PCI DSS and TRID themselves.

      How we screen for regulated work
    • Vetting

      A recruiter who knows the candidate

      Every KORE1 placement starts with a recruiter who has spoken with the candidate before you meet them.

      Contract engineering teams
    • Placing

      W2 contractors with dates on paper

      KORE1 is the W2 employer of record, so each assignment has a start date, an end date and a named approver.

      How contract staffing works

    The room on our client calls has changed. Devin Hornick, a KORE1 partner, puts it this way. “With AI, I’m seeing that so many more managers are involved… HR kind of ran things. HR’s there and… they just sit there on the calls now.” For a policy, that means the hiring manager is the person who gets asked how a contractor uses AI. Give that manager one page on day one.

    Send the one-page version with the req so the recruiter can raise it with the candidate before the offer, because our engineering searches average 17 days to hire and 92% of KORE1 placements are still in the seat twelve months later. For the model behind that, see contract staffing.

    Questions

    Common Questions

    What should a contractor AI use policy include?

    It should cover approved tools and accounts, prohibited data classes, disclosure of AI-generated code in pull requests, a pointer to your IP assignment and removal of tool context at offboarding. Those five sections answer most of what a security review or a customer questionnaire asks. Add an owner, a review date and a place to record exceptions.

    Should you ban contractors from using AI tools?

    Usually no. A ban pushes use onto personal accounts you can’t see, which is worse than a rule you can audit. Name approved tools, keep regulated data out of them and ask for disclosure. Some teams do ban AI for one codebase, and the data table still works there. Set every class to prohibited.

    Who owns code a contractor writes with an AI tool?

    The signed IP assignment agreement decides who owns contractor work product, so check that it reaches AI-assisted output. Separately, the U.S. Copyright Office’s January 2025 report says purely AI-generated material is not protected by copyright. Keeping a record of where AI helped matters. Ask counsel to read both.

    How should contractors disclose AI-generated code?

    In the pull request, with one line that says whether AI helped, which approved tool it was and which parts it touched. The contractor still owns the review of every line. An employee reviews the pull request before it merges to a protected branch.

    What happens to AI tool history when a contractor leaves?

    Delete or hand over the chat history, saved prompts, memory and uploaded files on or before the last day, then remove the seat and revoke tokens and integrations. Keep the purge record. Company-owned accounts make this possible, because a personal account is not yours to purge.

    About the author

    Kris Drouet has spent 25 years building, leading and advising engineering organizations in fintech, mortgage technology and regulated SaaS. His author page lists everything he writes for KORE1.

    Next step

    Need Contract Engineers Who Already Work This Way?

    Tell us the roles, the approved tools and the start date. A KORE1 recruiter takes the req from there.

    Find Contract Engineers →

    Not yet? Download the template or read the contractor AI tool policy post first.