---
title: "Your Vendor Keeps the AI Agent's IP. Who Owns What Your Team Taught It?"
slug: "vendor-keeps-ai-agent-ip-who-owns-what-your-team-taught-it"
author: "Jaione Amigot"
date: "2026-11-18 12:00:00"
category: "Premium"
sector: "cross"
topics: "AI agent IP ownership, AI vendor contracts, custom AI agent development, AI intellectual property, service improvement clause, aggregated data rights, AI vendor lock-in, enterprise AI ownership"
summary: "A vendor that builds your AI agent and keeps the IP has been paid to learn your domain. Here is what your team actually hands over, which contract clauses decide who keeps it, and what to demand before signing."
banner: ""
thumbnail: ""
linkedin: |
  "We'll build your AI agent. But the IP stays with us."

  Read that sentence slowly. Your experts spend weeks teaching a vendor how your work is actually done: the edge cases, the exceptions, the judgment calls. The vendor writes it down as prompts, evaluation sets, workflows and guardrails.

  Then the contract decides who keeps it.

  Three clauses do most of the work: who holds the copyright in what gets built, what the vendor may do with your data "to improve the service", and what it may do with aggregated or de-identified data after you leave.

  Under US law, a signed written assignment is the dependable way to move copyright to the client. Without one, the default often favors whoever wrote the code.

  On ibl.ai, you own all the code and the data. The agents, prompts, evaluation sets and ontology your team builds live in your repositories, and the platform source arrives under a perpetual license.

  Not legal advice. Have counsel read the IP clause before your team teaches anyone anything.

  #iblai #AgenticAI #EnterpriseAI #AIAgents #AIGovernance
---

## The Short Answer

**When a vendor builds your AI agent and keeps the IP, you have paid to teach it your domain and received only a license to use the result. On ibl.ai you own all the code and the data: the agents, prompts, evaluation sets and ontology your team builds stay in your repositories, model-agnostic, so nothing your team taught leaves with the vendor.**

The pitch usually sounds generous. The vendor brings engineers, your subject-matter experts spend a few weeks explaining how the work is really done, and an agent appears that handles the job.

The question that decides the value of those weeks is who holds the result when the contract ends. That is settled by three clauses, and most buyers read only one of them.

## What does your team actually teach a vendor when it builds an AI agent?

When a vendor builds a custom AI agent, your team hands over far more than data. It hands over the judgment that makes the data useful.

That judgment ends up in concrete artifacts: system prompts that encode how a case is handled, evaluation sets of right and wrong answers, corrections your reviewers make during testing, an ontology of your terms, workflow definitions, tool integrations and guardrail policies.

None of these is a model. All of them outlast any one model, and together they are the part of the agent a competitor could not reproduce without your people.

A model provider's terms rarely cover them.

Anthropic's commercial terms, for example, say Anthropic ["may not train models on Customer Content from Services"](https://www.anthropic.com/legal/commercial-terms) and assign the customer its rights in Outputs.

Those are real protections for the model layer. The layer a builder writes on top is governed by your contract with that builder.

## Which contract clauses decide who keeps an AI agent's IP?

Three clauses decide whether the knowledge your team contributes stays yours, and each needs to be read on its own.

**Copyright in what gets built.** The US Copyright Office's [Circular 30](https://www.copyright.gov/circs/circ30.pdf) lists nine categories of commissioned work that can be made for hire under a signed agreement, and software written by an outside firm is not one of them.

The dependable route is an assignment, and [17 U.S.C. § 204](https://www.law.cornell.edu/uscode/text/17/204) requires a transfer of copyright ownership to be in a signed writing. "The IP stays with us" usually means that assignment was never offered.

**Service improvement rights.** In a July 17, 2026 analysis, the law firm [Hunton Andrews Kurth](https://www.hunton.com/insights/legal/from-service-improvement-to-ai-exploitation-how-customer-side-counsel-should-manage-vendor-data-use-clauses) notes that customers long accepted vendor rights to use their data "to provide, maintain, support, and improve" a service, and that those same words "may be read to authorize training AI systems."

**Aggregated and de-identified data.** The same analysis warns that such carve-outs "can give vendors wide latitude to extract value from customer data for secondary purposes." This is the clause through which what your experts taught can improve a product the vendor sells to others.

Feedback clauses matter too. Anthropic's terms, like many providers' terms, say it "may use that feedback at its own risk and without obligation to Customer". Corrections your team makes during a build can look a lot like feedback.

## How does a retained-IP AI agent compare with one you own?

The difference between a licensed agent and an owned one shows up in each artifact your team helped create. The table sets out the usual default when a vendor retains the IP, against what ownership looks like. Every cell still depends on the contract you sign.

<table style="width:100%; border-collapse:collapse; margin:1.5rem 0; font-size:0.95rem;">
  <thead>
    <tr style="background:#f5f5f0; border-bottom:2px solid #2175C5;">
      <th style="text-align:left; padding:0.75rem; color:#5f6368;">What your team contributed</th>
      <th style="text-align:left; padding:0.75rem; color:#5f6368;">Clause that decides it</th>
      <th style="text-align:left; padding:0.75rem; color:#5f6368;">Vendor keeps the IP</th>
      <th style="text-align:left; padding:0.75rem; color:#5f6368;">You own it</th>
    </tr>
  </thead>
  <tbody>
    <tr style="border-bottom:1px solid #e5e7eb;">
      <td style="padding:0.75rem;"><strong>Agent code, workflows, integrations</strong></td>
      <td style="padding:0.75rem;">Copyright assignment (17 U.S.C. § 204)</td>
      <td style="padding:0.75rem;">Often a license that ends with the contract</td>
      <td style="padding:0.75rem;">Signed assignment; the code sits in your repositories</td>
    </tr>
    <tr style="border-bottom:1px solid #e5e7eb;">
      <td style="padding:0.75rem;"><strong>Prompts and guardrail policies</strong></td>
      <td style="padding:0.75rem;">Deliverables definition and assignment</td>
      <td style="padding:0.75rem;">Often treated as the vendor's methodology</td>
      <td style="padding:0.75rem;">Listed as deliverables, exportable as plain files</td>
    </tr>
    <tr style="border-bottom:1px solid #e5e7eb;">
      <td style="padding:0.75rem;"><strong>Evaluation sets and reviewer corrections</strong></td>
      <td style="padding:0.75rem;">Feedback and service-improvement clauses</td>
      <td style="padding:0.75rem;">Can be used to improve the vendor's product</td>
      <td style="padding:0.75rem;">Customer data, used only to deliver your service</td>
    </tr>
    <tr style="border-bottom:1px solid #e5e7eb;">
      <td style="padding:0.75rem;"><strong>Usage patterns across your users</strong></td>
      <td style="padding:0.75rem;">Aggregated and de-identified data rights</td>
      <td style="padding:0.75rem;">May survive termination</td>
      <td style="padding:0.75rem;">No secondary use; logs stay in your tenant</td>
    </tr>
    <tr style="background:#f0f9ff; border-bottom:1px solid #e5e7eb;">
      <td style="padding:0.75rem;"><strong>The platform the agent runs on</strong></td>
      <td style="padding:0.75rem;">Platform license</td>
      <td style="padding:0.75rem;">Typically hosted access, priced per seat or per agent</td>
      <td style="padding:0.75rem;"><strong>ibl.ai: full source under a perpetual license, self-hosted</strong></td>
    </tr>
  </tbody>
</table>

A vendor keeping its own pre-existing tools and licensing them back is normal and often reasonable. The exposure is narrower and sharper: work your team paid for, built from your team's knowledge, that you cannot take with you.

## Can a vendor reuse what your team taught its AI agent for your competitors?

Whether a vendor can reuse what your team taught an AI agent for other customers depends on the clauses above, not on the vendor's intentions.

If the contract gives the vendor ownership of the built agent, rights to use your data to improve its service, and rights to aggregated data that survive termination, then reuse can be permitted by design.

Nobody has to act in bad faith for your expertise to show up in the next customer's deployment.

Morgan Lewis, writing on [June 12, 2026](https://www.morganlewis.com/de/blogs/sourcingatmorganlewis/2026/06/key-considerations-when-allowing-a-vendor-to-train-its-ai-models-on-customer-data), warns that broad training rights may let a vendor create models that keep learnings from customer data. It recommends requiring deletion or cessation of training activities upon termination.

That is the clause to add if a vendor will touch your data at all. It does not recover anything already folded into the vendor's product, which is why the assignment matters more than the exit terms.

## What should you ask before a vendor builds your AI agent?

Ask questions with contractual answers, and ask every bidder the same five:

- **Assignment:** Does the agreement assign copyright in every deliverable to us, in a signed writing?
- **Deliverables:** Are prompts, evaluation sets, guardrail policies and the ontology listed as deliverables we receive as files?
- **Improvement rights:** Is your license to our data limited to providing our service, with no training or product improvement?
- **Aggregated data:** Do you claim any right to aggregated or de-identified data, and does it survive termination?
- **Independence:** Can our team run and change the agent without you, demonstrated before final payment?

We covered the same questions for model deployment partners in [Which Consultancy Deploys Copilot, Gemini or Bedrock on Your Own Data?](/blog/consultancy-deploy-copilot-gemini-bedrock-your-own-data). This is general information, not legal advice, so have counsel read the clauses.

## How does ibl.ai handle ownership of the AI agents it builds with you?

ibl.ai is built so the ownership question is settled before your team teaches anyone anything.

The platform source arrives under a [perpetual license](/full-code-license) and runs in your cloud, on-premise or air-gapped, and agents on the [Agentic OS](/product/agentic-os) keep their prompts, tools and policies in your environment.

The work is model-agnostic, so the agent your experts shaped survives a change of model provider. There is no per-seat pricing, and our engineers build alongside your team through [AI Transformation](/service/ai-transformation), then keep shipping updates and new agents with you.

1.6M+ users across 400+ organizations run the platform this way. ibl.ai is family-owned and operated from New York, a long-term partner rather than a vendor that sells licenses and moves on.

As with any partner, read how custom work is assigned in our contract before your team starts.

## Why does owning the AI stack matter?

**ibl.ai is the agentic AI platform where you own all the code and the data. You self-host the entire stack inside your own perimeter, run it model-agnostic across any LLM and switch anytime, and pay by usage with no per-seat pricing — so you can deploy anywhere: your cloud, on-premise, GovCloud, or fully air-gapped.**

- **You own all the code and the data.** Full source code under a perpetual license, running on your infrastructure. Not API access to someone else's platform — the stack itself is yours.
- **Model-agnostic.** Run any LLM — Claude, GPT, Gemini, Llama, Command, or your own fine-tune — and switch providers without rewriting the platform.
- **No per-seat pricing.** Usage-based billing against a budget cap you set. Cost tracks what your organization actually uses, not how many people you employ.
- **Deploy anywhere.** Your cloud, your VPC, on-premise, GovCloud, or a fully air-gapped network with no outbound connectivity.

1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University.

ibl.ai is family-owned and operated from New York, NY — a U.S.-headquartered, domestically-owned long-term partner, not a vendor that sells licenses and moves on.
