There are four honest responses to a client asking for AI when nobody in-house has shipped it. Hire, if AI is becoming your positioning and you have booked work for two quarters - it costs around $260,000 in year one and takes three to six months. Use contractors for scoped work under about $5,000. Use a white-label partner when the work is lumpy or you need to answer this month. Or refuse the work, which is the right answer more often than anyone writes about - particularly when the client wants AI because it is on a slide rather than because it solves something.
How can an agency deliver AI work without hiring?
There are four honest options: hire an AI engineer, use contractors, bring in a white-label partner, or refuse the work. Hiring suits an agency making AI its positioning with work already booked; contractors suit small, scoped jobs; a white-label partner suits lumpy demand; and refusing is right when the client wants AI for its own sake.
Three clients have asked about AI in the last quarter. You said something non-committal to two of them and lost the third to an agency that said yes. Nobody on your team has shipped a retrieval system or an agent, and the roadmap for the next two quarters does not have room to learn.
There are exactly four responses to this, and three of them are legitimate. What follows is the honest version of each, including the one most articles about this skip.
The four options
| Option | Time to first delivery | Cost shape | Main risk |
|---|---|---|---|
| Hire an AI engineer | 3-6 months | $220k+/year, fixed | Utilisation - an idle specialist is the most expensive thing on this table |
| Use contractors | 1-3 weeks | $60-150/hr, variable | No continuity, and you become the integrator |
| White-label partner | 1-2 weeks | Fixed per project | Choosing badly, and a layer between you and the code |
| Refuse the work | Immediate | Zero | Losing the client's other work too |
Option 1: when should an agency hire an AI engineer?
The right answer if AI is becoming your positioning rather than a service line. An in-house engineer can stand in front of a client and be the expert, scope accurately in the first meeting, and accumulate the knowledge that makes the fifth project cheaper than the first.
The trap is utilisation. A salaried specialist is the most expensive thing on that table when the pipeline is lumpy, and AI work at a 20-person agency is almost always lumpy at first. By our estimate the fully-loaded cost is around $260,000 in year one once recruiting and the empty seat are counted - the arithmetic is worked through in white-label AI vs hiring in-house.
Choose this when: you have booked AI work for two quarters, you can technically assess a candidate, and there is someone to review their code. Fail the third test and you will not find out for six months.
Option 2: when do AI contractors make sense?
Fast, flexible, and genuinely right for a scoped piece of work under about $5,000. You talk directly to the person writing the code, which is worth more than it sounds.
It goes wrong at the seams. One contractor for retrieval, another for the front end, and the integration work lands on you - an agency owner who is now a technical project manager on a stack nobody in the building understands. It also has no cover: an illness in week four is your deadline, not theirs.
Choose this when: the scope is small and clear, your team can maintain the result, and you only need one specialism. The full comparison is on AI agency vs freelancer.
Option 3: when does a white-label AI partner make sense?
You keep the client, the contract and the margin; a specialist builds invisibly under your brand. Cost scales with projects rather than months, so a quiet quarter costs nothing, and you can say yes next week rather than next quarter.
What you give up is immediacy and institutional memory. A partner is not in the room when you are scoping a pitch, context is rebuilt at the start of each project, and you carry the risk of having chosen badly - which is a real risk, and the next two sections are about reducing it.
Choose this when: the work is lumpy, you need to answer a client this month, or the project needs a skill your hire would not have had anyway.
Option 4: when should an agency refuse AI work?
When the client wants AI because it is on a slide rather than because it solves a problem, or when the budget is a fraction of the ambition. It is the option nobody writes about, and sometimes the correct one: refusing well, with a better use of the same budget, often keeps the account and leads to work you can deliver.
If the client wants AI because it is on a slide rather than because it solves something, the project will be judged against an expectation nobody wrote down, and no delivery model saves you from that. If the budget is $3,000 and the ambition is $40,000, saying so costs you one project and keeps the relationship.
Refusing well is a skill. "We do not do that" loses the account. "That specific thing is not worth what it would cost you - here is what I would do with the same budget" usually keeps it, and quite often turns into different work you can actually deliver.
What does a good white-label partner engagement look like?
An NDA before scoping, a fixed price with an explicit out-of-scope list, work done in your repositories under your identity, a weekly update you can forward unedited, client-owned infrastructure and API accounts, and a documented handover. If you take the third option, these are the things that distinguish a working arrangement from a bad quarter.
- NDA and non-solicit before scoping, not after the proposal. You should be able to speak freely about your client on the first call.
- Fixed scope, fixed price. You are selling your client a fixed price and cannot absorb an hourly overrun. A partner who will not quote fixed is telling you they cannot estimate.
- An explicit out-of-scope list. The more useful half of a quote. It is what prevents a scope argument in week five.
- Work in your repositories, under an identity you control. The commit history should carry your agency, and you should be able to see progress without asking.
- A weekly written update you can forward unedited. If you have to rewrite it every week to remove the partner, that is a tell.
- Client infrastructure, client API accounts. Model spend billed at cost to your client rather than resold with a margin, so a busy month is not your problem.
- Handover as a deliverable. Documentation, runbooks, the evaluation harness and a recorded walkthrough, in your voice.
Ours are on how a white-label engagement works, with the margin table on the white-label page.
What are the red flags when choosing a white-label AI partner?
Five signals are worth walking away over: they will not put you on a call with the engineer, no case study contains a number, the proposal has no evaluation harness, they say yes to everything, and they are vague about where the engineers sit. Each one is easy to check before you sign anything.
They will not put you on a call with the engineer
The single most reliable signal. If the person who pitched is the only person you meet, you are buying a sales process. Ask to talk to whoever would build it, before you sign.
No case study contains a number
"Transformed their workflow" is not a result. Ask what was measured, what it was before, and what it was after. Anyone who has shipped production AI can answer immediately; anyone who has shipped demos changes the subject.
No evaluation harness in the proposal
Ask how you will know the system still works in six months. If the answer is testing rather than a labelled evaluation set with agreed thresholds, they have not maintained an LLM system through a model deprecation. The detail of what to expect is in the evaluation harness we ship with every build.
They say yes to everything
A partner who never says "that is not worth building" is a partner who will build the thing that is not worth building, on your client's budget and your reputation.
Vague about where the engineers are
Offshore delivery is fine and common - what is not fine is finding out at the first standup. Ask where the team sits, how many hours overlap your working day, and who holds the contract. A partner who states it plainly has thought about how you will explain it if your client asks.
Where do most agencies end up?
Most agencies that get good at this end up with both: one in-house engineer who owns architecture, standards and the client conversation, and a partner absorbing overflow and the specialist work that engineer has not done before. The hire stops being a bottleneck and the partner stops being the only source of expertise.
Getting there usually starts with the partner rather than the hire, because delivering three projects teaches you what you are actually hiring for - which is a much better brief than the one you would write today.