Dashboard

How to Subcontract Part of an AI Build

Subcontracting an AI build breaks the usual method, because you can no longer judge a subcontractor by reading their output. Selection moves from reading to testing.

Manuele Estivo
Manuele Estivo
Growth & SEO Lead
19 October 20261 min read

Subcontracting part of an AI build works the same way as any other subcontracting, with one difference that breaks the usual method: you can no longer judge a subcontractor by reading their output. Plausible-looking code is now free. Anyone can send you four files that read like competent work and do not function under load, and you will not catch it by skimming.

So the selection moves from reading to testing, and the contract has to cover something it never used to: what the subcontractor is allowed to put into a third-party model.

Decide what to hand over before you decide who to

Subcontract at a seam, not in the middle of a thing. The parts that hand over cleanly share three properties: a defined input, a defined output, and a way to tell whether it is right without understanding the rest of the system.

Good candidates in a typical AI build:

  • A data import or migration path with a sample file and expected result

  • An integration against a documented third-party API

  • A self-contained UI area, such as a settings screen or an admin table

  • Test coverage for an existing module

  • Infrastructure work: deployment, backups, monitoring

Bad candidates, in roughly descending order of regret: the prompt layer, anything touching your data model, and "the bits I have not thought about yet". The last one is not a scope, it is a hope.

If you cannot write down the acceptance condition in two sentences, the piece is not ready to hand over. Spending an hour getting it to that state is the highest-value hour in the whole arrangement, and it usually reveals that you understood the work less well than you thought.

Selecting: pay for a small real task

Portfolios and interviews have both lost most of their signal. Replace them with a paid trial task, small, real, and from your actual codebase if you can manage it legally.

What makes a good trial task:

  • Two to four hours of work, paid at full rate. Unpaid trials select for the desperate.

  • A real requirement you were going to build anyway, not a puzzle.

  • An acceptance condition stated up front.

  • One deliberate ambiguity in the brief.

That last one is the whole test. Everyone will produce working code. The question is what they do with the ambiguity: ask, state an assumption and proceed, or silently pick an interpretation and say nothing. The third behaviour is the expensive one, and it is the one that a conventional interview never surfaces.

Then read the work, with a specific question in mind: does the shape of this match the brief, or does it match what a model produces when given the brief? Generated work that nobody shaped has a texture to it, generic variable names, defensive handling of conditions that cannot occur, abstractions nobody needed. It is not that AI involvement is a problem. It is that unreviewed AI involvement is, and you are hiring for the review.

The two clauses that are new

Standard subcontractor agreements predate all of this. Two additions are worth insisting on.

Disclosure and permission on AI use. Not a ban, which is both unenforceable and silly. A statement of what tools they use, and an obligation to comply with whatever constraints your own client contract imposes on sending data to third parties. If your client's agreement says their data does not leave your systems, that obligation has to travel down, and it will not travel on its own.

Responsibility for what they ship. Make explicit that the subcontractor is accountable for the correctness of what they deliver regardless of how it was produced. This sounds obvious and it is worth writing because it forecloses a conversation that genuinely happens, in which a defect is attributed to the tool rather than the person who shipped it.

If you are subcontracting under a client NDA, check the NDA permits subcontractors at all. Many require prior written consent, and many people discover this after the fact. Our guide to writing an NDA for an AI project covers the clauses from the other direction, which is useful for knowing what you are agreeing to when a client hands you theirs.

Acceptance: define "done" as a test, not an opinion

This is where AI-era subcontracting most differs from the old version. "It looks right" is no longer evidence of anything.

Write the acceptance condition as something executable before work starts:

  1. Here is the input. Here is the expected output. It passes.

  2. Here are three edge cases that must not break it.

  3. It runs in the project's own test setup, not just on their machine.

  4. No new dependencies without asking.

Point four matters more than it used to, because agents add packages freely, and a subcontracted module that quietly pulls in four new libraries is a supply-chain decision somebody made on your behalf. It is also the most common cause of a module that works alone and fails in your build.

Pay against passing, not against delivery. The gap between "sent" and "works" is where the margin goes.

Pricing it so you do not lose money

The arithmetic that catches people: you quote the client, subcontract part at a rate that leaves a margin, and then spend unbilled hours specifying, reviewing and integrating. Those hours are real and they are yours.

A rough rule from the work this usually involves: budget review and integration at about a quarter of the subcontracted hours for someone you have worked with before, and nearer a half for someone new. If the margin does not survive that, subcontracting that piece is not worth doing, and you should either do it yourself or raise the client price. This is exactly the kind of invisible cost that distorts your effective hourly rate if you never measure it.

If subcontracting turns out not to pay on your current pricing, the problem may be the pricing rather than the arrangement, and the broader AI monetization strategies guide covers the alternatives.

Pay the subcontractor on a schedule that does not depend on your client paying you, if you can afford to. It is a straightforward way to be the client people work hardest for, and it costs nothing when the project goes normally. If cash flow genuinely will not allow it, say so before they start rather than after, and set payment milestones that line up honestly with your own.

The handover back

Getting the work in is not the end. Require, as part of acceptance:

  • A short note on anything they decided that the brief did not specify

  • The test cases, not just the code

  • Any prompts or agent configuration used, if the work involved them

  • A list of the files they touched

The last one takes them thirty seconds and saves you an afternoon when something breaks in four months and you are trying to remember what was theirs.

FAQ

Should I tell my client I am subcontracting?

Check the contract first, because many require consent. Beyond the legal position, telling them is usually better. Clients mind far less than people expect, and they mind a great deal about finding out accidentally.

Can I subcontract the prompt work?

You can, but it is the piece that most rewards knowing the client's domain, and it is the hardest to accept against an objective test. If you do, insist on test cases as part of the deliverable.

What rate should I pay?

Enough that the better candidates say yes. Subcontracting at a steep discount selects for people who could not get work elsewhere, which is the opposite of what you want on a piece you will not be reviewing line by line.

How do I handle a subcontractor who used AI badly?

As a quality problem, not a tooling one. The deliverable failed the acceptance test. That is the entire conversation, and the contract clause above is what keeps it there.

How did this land?

About the author

Manuele Estivo
Manuele Estivo

Growth & SEO Lead

Manuele covers distribution: SEO, content strategy, and how AI-built products find their first thousand users. He tests everything he recommends.

Share

Get the next post in your inbox

One email a month. Product updates, engineering posts, and the best of Built with Swarmz.

I agree to receive emails about AI building tips and Swarmz product news. Unsubscribe any time.