White-Label Development in 2026: Selling AI Work Without an AI Team

White-label development means a specialist partner builds the AI feature while your agency sells it, brands it and manages the client, who signs with you, sees only your name and gets the code from you. Vendor guides put agency gross margins on resold work at 50–70%, but on senior AI capacity a 50% margin already prices you at the top of the US market (the arithmetic is below), so profit comes from scoping well more than from markup. It works when the contract chain (NDA, non-solicitation, IP assignment, and a GDPR Article 28 agreement if personal data is involved) is signed before the first commit, and when you sell a fixed-price discovery before you quote the build.

Key takeaways

You own the relationship, the scope and the liability. The partner answers to you. Your client has no contract with the partner and usually never meets them.
The margin benchmarks are vendor figures. They come mostly from cheap wholesale work. On senior capacity at about $123 an hour, check the resulting client rate against market rates before you pick a markup.
Resell bounded AI work: LLM features, assistants over the client’s own content, automations, and rescues of AI-built apps. Avoid open-ended agents, decisions about people, and accuracy promises.
IP doesn’t pass to your client by default. Commissioned software is generally not “work made for hire” under US law, so you need written assignments partner to agency and agency to client.
Personal data creates a processor chain. Client is controller, you’re processor, the partner is sub-processor. Article 28 requires the client’s authorization and a flow-down contract.
Sell discovery first. A short fixed-price discovery turns “add AI” into a scope you can quote.

How white-label works day to day

Who talks to the client. You do. The partner works in your Slack, your tracker and your client’s repository. If you want an engineer on a client call, they join as part of your team, often on your email domain. Price, scope and deadlines always come back to you, even when the client asks the engineer directly.

Whose brand. Yours, on reports, commits, documentation and demos. The partner doesn’t put its name in the work or your client in its portfolio. The one limit is the law: when personal data is involved, the client may be entitled to know who processes it, and enterprise security questionnaires often ask who has system access. White-label means invisible in marketing and daily contact, not hidden from a regulator.

Who owns the code. It should end up with your client, in two steps: partner to agency, agency to client. US copyright law treats a commissioned work as “made for hire” only if it falls into nine listed categories and the parties sign a written agreement. Software isn’t one of them. A copyright transfer is not valid unless it’s in writing and signed. Without assignments in both contracts, the code may legally belong to whoever typed it.

Who carries liability. You, toward your client. The partner is liable to you under your contract, usually up to a cap. And nobody gets to blame the software: in Moffatt v. Air Canada, a Canadian tribunal held the airline responsible for wrong fare advice from its website chatbot, calling the argument that the bot was a separate entity “a remarkable submission”. Your client answers for the chatbot to its customers; you answer to your client.

In a normal week you set priorities and acceptance criteria, the partner’s engineers work in your tools, and on Friday you get a written report you can forward as-is. You accept the work from the partner; your client accepts it from you.

The business model: markup and margin

What the published figures say

We found no independent survey of white-label development markups. Every figure comes from a firm that sells white-label services, so treat these as vendor estimates:

KrishaWeb recommends “a gross profit margin of 50 to 70%”, calculated on full delivery cost (partner fee plus your overhead).
BSS Commerce suggests reselling at “roughly 2–3× partner cost / 50–70% gross margin”.
Codeable writes that “most agencies aim for 50–70% gross margins on resold white-label work”.

Markup and margin aren’t the same number: markup is profit over cost, margin is profit over price, so a 100% markup is a 50% margin. And these benchmarks mostly describe brochure sites, Shopify builds and SEO, where wholesale cost is low. Then there’s the net figure: Promethean Research’s 2026 State of Digital Services report found agency net margins slipped to 13%. Gross margin has to pay for account management, sales and rework before any of it is profit.

A worked example (illustrative arithmetic, not a client story)

A client wants an assistant that answers questions from its help center. Discovery is done, and the build fits one two-week block of 65 senior hours from a partner at $8,000, about $123 an hour. You add 12 hours of your own project management and QA at an assumed internal cost of $60 an hour ($720), for a total delivery cost of $8,720.

Price to client	Markup on partner fee	Your gross profit	Gross margin	Implied rate per partner hour
$11,000	37.5%	$2,280	20.7%	$169
$13,500	68.8%	$4,780	35.4%	$208
$16,000	100% (2× partner fee)	$7,280	45.5%	$246

Look at the last column. Senior US teams charge roughly $100–$250 an hour, so the “2× partner cost” rule puts this project at the very top of the US range. That holds if your client pays US agency prices and you carry real account management. It’s hard to defend if the client could hire a specialist directly. The stable model is a moderate markup on the build plus work you deliver yourself: strategy, design, content and a support retainer. Quote the outcome, not hours, and keep the hourly arithmetic in your own spreadsheet.

What AI work is realistic to resell

Good white-label AI work is bounded: a clear input, a clear output, a way to check it, and a human in the loop.

LLM features in an existing product. Summaries, classification, drafting, document extraction, smarter search. Our AI feature cost guide has build ranges and run costs by tier.
Assistants over the client’s own content (RAG). Help-center bots, internal policy search. The engineering is in retrieval, permissions and evaluation, not the chat window.
Workflow automations. Lead routing, document intake, CRM enrichment. Success is checkable: did the right record land in the right place?
Rescues of AI-built apps. Clients arrive with a Lovable or Bolt prototype that demos well and leaks data in production. Productive’s March 2026 survey of 174 agencies describes one that won a client whose AI-built web app needed outside help to work and pass security compliance. It also found about a third of agencies have launched something AI-related, mostly with small impact so far. For costs, see fixing a vibe-coded app.

What to avoid, or resell only with an experienced partner and a lawyer:

Open-ended autonomous agents. Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, citing costs and unclear value. Scope an agent to one workflow or don’t sell it.
Decisions about people. The EU AI Act lists systems that filter job applications, evaluate candidates or score creditworthiness as high-risk, the category with its strictest obligations.
Accuracy promises. “95% accurate” before anyone has measured it on the client’s data is a refund waiting to happen.

Engagement models: who carries the risk

The models differ in who absorbs an overrun and who is exposed if the client relationship sours. (For deposits, caps and refund clauses, see fixed price vs. time and materials vs. retainer.)

Model	How you pay the partner	Overrun risk	Client-relationship risk	Typical use	Watch out for
Per-project fixed price	One price per defined deliverable	Partner	You	Discovery, a scoped feature, an audit	Scope disputes; padded quotes on vague briefs
Dedicated capacity (prepaid blocks or monthly)	A fixed fee per block of hours	You (you direct the work)	You	Ongoing client work, iteration after launch	Idle hours while the client is slow to decide
Hourly (time and materials)	Hours worked × rate	You	You	Unclear scope, maintenance	No ceiling unless you add one
Revenue share	A share of client fees or product revenue	Shared	Shared, and hard to unwind	Productized offers you sell repeatedly	Disputes over what counts as revenue; the partner gains a stake in your client

Fixed price suits the first piece of work with a new client, because you can quote it back-to-back. Capacity suits what follows, once you know how fast the client decides. Revenue share costs nothing up front but ties the partner’s income to your client’s account, which is the incentive a white-label deal should avoid.

How to choose a white-label partner

Questions to ask

Who exactly does the work? Names, seniority, location, and whether they use subcontractors.
What have you shipped like this? Ask for a walkthrough: how retrieval was tested, how permissions worked, what broke after launch.
How do you evaluate an AI feature before it ships? Look for a test set built from real examples, run before and after every change. “We test it manually” means nobody will notice regressions.
Where do API keys and model costs live? In the client’s own provider account, so the client pays usage directly and keeps access when you part ways.
What’s your rule on client contact? No contact unless you ask, and every commercial question referred to you, in writing.
Can I see a weekly report? If you can’t forward it to a client, you’ll be rewriting it every Friday.
How fast do you start and respond, and what happens to undelivered hours? In writing.

Red flags

They want an introduction to the client “to understand the requirements”.
Their case studies name other agencies’ clients.
A firm quote on a one-paragraph brief, with no discovery and no assumptions.
No way to measure output quality, which AI features need because code can pass every test while answers get worse.
They want the API keys, or the model bill, in their own name.
They won’t sign a non-solicitation clause.

References and a trial

Ask two agency references one question: “Did they ever go around you?” Then run a small paid trial, such as a codebase audit or a one-week proof of concept, before the partner touches a client’s critical path. It shows code quality, communication and honesty about problems better than any sales call.

Contract essentials

You need two contracts that line up: partner agreement and client agreement. Wherever they don’t match, you carry the gap.

NDA and confidentiality. Cover the client’s code, data and credentials, plus the client’s identity and the fact that you use a partner.
Non-solicitation of your clients. The partner may not pitch your client or accept work from them directly, during the agreement and for a set period after it. Look for sensible exceptions, such as companies the partner already worked with, and a real remedy, such as a fee on revenue taken.
IP assignment that flows to the client. The partner assigns the work product to you as it’s created or on payment, and you may pass it on without consent. The partner keeps its pre-existing tools but licenses them to you with the right to sublicense to the client. Every person on the partner’s side must have assigned their rights to the partner, or the chain breaks one level down.
SLAs. Start date, working hours and time zone, incident response times, reporting cadence, and what happens to undelivered hours.
Liability caps. Expect the partner to cap its liability to you, and read where the cap sits. If your client contract has uncapped liability for data breaches, the difference is yours.
Data processing. With EU or UK personal data, Article 28 says a processor may not engage another processor without the controller’s prior written authorization, must impose the same obligations on it by contract, and stays fully liable for its failures. So: get your client’s written approval to subcontract, sign a data processing agreement with the partner that mirrors your client’s, and list the model provider as a sub-processor. For transfers out of the EU, the standard contractual clauses include a processor-to-sub-processor module.

How to scope and quote AI work without overpromising

“Add an AI assistant to our site” can mean a weekend or a quarter. Quoting the build from that sentence is the fastest way to lose money.

Sell discovery first, at a fixed price. One to two weeks: look at the client’s real data, prototype on a sample, define “good” with examples, then estimate. Your partner can fix a price for discovery, so you can quote it back-to-back.

Write success criteria as examples. “Answers 50 sample questions with the right help-center article, and says so when the answer isn’t there” is testable. “Accurate and helpful” isn’t.

Price run costs separately. Model usage is a monthly bill that grows with traffic; put it in the client’s provider account. OpenAI doesn’t train on API data by default but keeps abuse-monitoring logs for up to 30 days, which belongs in your data processing paperwork.

Disclose the AI. Since 2 August 2026, the EU AI Act requires that people using a chatbot be told they’re dealing with AI unless it’s obvious. Put it in scope.

Fix the price, flex the scope. Agree budget and date, rank features, and let the low-priority ones drop if time runs short.

The risks

Quality you can’t see. You’re selling code you didn’t write. Veracode found 45% of AI-generated code introduced a security flaw in its tests. Require code in the client’s repository from day one, visible pull requests, tests on critical paths, and an independent review before launch for anything touching personal data or money.

Dependency. If only the partner understands the system, you can’t switch or bring it in-house. Make documentation and a handoff session deliverables, and keep repositories and cloud accounts in the client’s name.

Partner poaching. The non-solicitation clause is the legal answer. The practical one is a partner whose business doesn’t depend on end clients.

AI-specific data handling. Prompt injection tops OWASP’s Top 10 for LLM applications, followed by sensitive information disclosure. An assistant that can read every document in the client’s drive will eventually show one to the wrong person. Scope its data sources and permissions explicitly, and ask the partner how they test for both.

Checklist before you sell the first AI project

Working with coderband as your AI team

If your clients want AI features, a real backend or GPU work and your team does design, Webflow, Shopify or marketing, Agency Partner Capacity is built for that. You book a prepaid two-week block of 65 senior hours for $8,000. Once the partner agreement and NDA are signed and the block is paid, we start within two business days, in your repository, tools and standups.

It’s white-label by default, with NDA and non-solicitation in the contract. We never contact your client unless you ask, and on calls we can join as your team under your email domain. Each week you get our hours and a written report you can forward to the client as-is. Hours we fall short on carry over or are refunded. If hours go unused on your side, up to 13 (20% of the block) carry into your next block if it starts within 30 days.

We’re senior and priced that way, so we’re the wrong fit for cheapest-possible offshore capacity. If a client is asking you for AI work this month, book a call and bring the brief; we’ll tell you whether it’s a good first project to resell.