AI Development Subcontractor: What Makes the Handover Safe

An AI development subcontractor is a specialist studio that builds an AI system under your agency’s name and then transfers it out. The arrangement is safe when the contract treats the deliverable as an asset that changes hands, not a service that gets performed: a signed copyright assignment, accounts created in the client’s name from day one, and written acceptance criteria. Get those three wrong and you have not subcontracted a build. You have introduced your client to a vendor you cannot control.

Most of our own work is direct with the company that owns the problem. Partner delivery runs alongside that, and the terms below are the ones we sign either way.

The three-party structure puts the risk in the middle

In a subcontracted AI build there are three parties and only one of them is exposed twice. Your agency holds the client relationship and issues the invoice. The subcontractor writes the code. The client owns the result. If the build fails, the client’s claim lands on your agency, not on the party that actually wrote the failing code.

That asymmetry is the whole reason the contract matters more here than in most outsourcing. You are underwriting work you did not perform, in a discipline you cannot fully inspect, for a client who will never hear the subcontractor’s name. The terms are the only inspection you get.

A work-for-hire clause does not transfer ownership of code

In the United States, calling a commissioned build a work made for hire does not make it one. The US Copyright Office states in Circular 30 that a commissioned work qualifies as a work made for hire only if it falls within one of nine listed categories and both parties sign a written agreement saying so. Software is not among the nine.

The nine are narrow and specific: a contribution to a collective work, part of a motion picture or audiovisual work, a translation, a supplementary work, a compilation, instructional text, a test, answer material for a test, and an atlas. Custom software written by an independent contractor fits none of them.

What happens instead is set by statute. Under 17 U.S.C. 201(a), copyright vests initially in the author, which is the developer who wrote it. Under 17 U.S.C. 204(a), a transfer of copyright ownership is not valid unless there is an instrument of conveyance, or a note or memorandum of the transfer, in writing and signed by the owner of the rights conveyed.

So here is a test you can run in thirty seconds. Search your subcontract for the word “assign”. If the only ownership language in it is “work made for hire”, the copyright in the system your client just paid for probably still sits with the person who wrote it, and your client’s ownership is a sentence in a document rather than a fact.

Two practical notes. The assignment has to reach through the chain, so it must also cover anyone the subcontractor brings in. And this is US federal copyright law: other jurisdictions set their own default rules, which is why the agreement should name its governing law. Have a lawyer in your jurisdiction read the clause. The point is not that you can draft it yourself, it is that you now know which clause to look at.

Agencies subcontract this work because the capability sits elsewhere

The Federal Reserve Board’s FEDS Note of 3 April 2026, which triangulates the Census Bureau’s Business Trends and Outlook Survey with the Real-Time Population Survey and the Atlanta Fed’s Survey of Business Uncertainty, put AI adoption at roughly 18% of US firms as of year-end 2025, while about 78% of workers were employed at an AI-adopting company. Professional services sat near 33% adoption and finance near 30%.

Read those two numbers together. Capability is concentrated in large employers. The mid-market client sending an AI request to your agency is usually in the other group.

Deloitte’s State of AI in the Enterprise 2026 edition, fielded across 3,235 board, C-suite, VP and director-level leaders in 24 countries during August and September 2025, found insufficient worker skills was the biggest barrier to putting AI into existing workflows. The responses were educating the workforce at 53%, upskilling and reskilling at 48%, and hiring specialised talent at 36%. Hiring came third because it is the slowest of the three, not because it works better. We covered the staffing arithmetic in more detail in white label AI automation.

The quality gate is the half that nobody writes down

Ownership protects you after delivery. Acceptance criteria protect you at delivery, and this is where AI builds differ from ordinary software subcontracting.

Stack Overflow’s 2025 Developer Survey, published 29 December 2025 from about 49,000 responses across 177 countries, found 84% of developers use or plan to use AI tools while only 33% trust the accuracy of what those tools produce, against 46% who actively distrust it. The most cited frustration, at 66%, was “AI solutions that are almost right, but not quite”, and 45.2% said debugging AI-generated code takes longer than writing it.

“Almost right” is the failure mode you are reselling. It does not show up in a demo, because a demo uses clean inputs. It shows up six weeks later on a malformed record at 2am. Deloitte’s same 2026 study found only one in five companies had a mature governance model for autonomous AI agents, so assume your subcontractor’s governance is something you have to specify rather than inherit.

Write acceptance criteria as behaviour, not features. What the system does with a malformed input. What condition stops a run instead of guessing. What gets logged, and who reads the log. A partner who answers those in writing has built production systems. A partner who answers with model names has not.

What a safe subcontract specifies

Term	The version that fails	The version that holds	How to test it
IP ownership	”work made for hire”	signed assignment of copyright to the client, covering the subcontractor’s own contractors	search the document for “assign”, check the signature
Accounts	vendor’s repository and hosting	created in the client’s name from day one	log in yourself before final payment
Deliverable	”a working agent”	repository, documentation, recorded walkthrough, named acceptance criteria	can a new developer run it from the docs alone
Data	a mutual NDA	named data scope, no client data used for training, sub-processors disclosed	ask who else sees the data, in writing
Price	time and materials	fixed price against fixed scope, changes are a new order	can you quote your client before you sign
Continuity	goodwill	everything runs on the client’s accounts	if the partner disappeared on Friday, what runs Monday

The last row is the one that makes the other five enforceable. If the system runs on the client’s infrastructure, none of the other terms need to be litigated, because nothing the subcontractor can withhold is load-bearing. That is the argument in own every repo your AI agency builds, applied one layer up the chain.

Full handover, itemised

At our own handover the client receives the repositories, the database, the hosting and service accounts created in their name at the start rather than migrated at the end, documentation written for their team, and a recorded walkthrough. Sprints carry 30 days of fixes after that.

The operating discipline transfers with it, and that part is visible: every agent registered before it runs, an owner and a defined fence for each one, a run log for every action, and humans approving anything that leaves a draft.

Our own group’s fleet stood at 41 registered agents with 9 live in production as of 30 June 2026, and over the 13 days measured to 4 July 2026 those agents completed 5,450 runs with zero failures. Each of those agents is a tomte, our word for one production agent with one defined job. The full measurement, including what “zero failures” does and does not mean, is on the agent fleet proof page.

Handover is not an exit clause in this model. It is the ending of every engagement, which is exactly why the work resells cleanly.

Scope it before you quote it

The reason most subcontracted AI builds go wrong is that the scope was written by whoever was selling, not by whoever was building. An AI operations audit is $1,900 fixed over 10 working days and produces an automation map, a build list ranked by effort and return, and a 90-day plan, with a full refund if we find nothing worth building. An agency can put a client through that first and then quote from a ranked list instead of from a conversation.

Send us the scope you have, however rough. You get a fixed number and a delivery window in writing within one business day, and no call is required at any point. Start async.