TL;DR & Quick Summary
Client demand for AI work has outpaced most agencies' ability to deliver it. The gap is real: a single competent AI engineer is expensive, hard to assess if nobody on the team can interview them technically, and idle whenever the pipeline dips.
White-label delivery closes that gap — but the arrangement you choose determines your margin, your risk, and whether you still own the client in two years.
The three models, and who each suits:
| Model | Agency does | Partner does | Best when |
|---|---|---|---|
| Fixed-scope subcontract | Sells, scopes, manages | Builds to spec | Work is well-defined and occasional |
| Retained capacity | Sells, prioritises | Supplies ongoing engineering | Demand is steady but uneven |
| Embedded partner | Sells, co-delivers | Joins calls under agency brand | Projects are complex and evolving |
- Key Takeaway: The commercial terms decide the margin, but the operational terms decide whether you survive a partner dispute. Own the repository and the cloud accounts from day one — that single decision converts a potential outage into an inconvenience.
- Get Started: Looking for delivery capacity behind your brand? Book a Strategy Call with Cogniq AI or review our custom AI development capabilities.
Why Agencies Are Hitting This Wall
The pattern is consistent across marketing, design, and general software agencies.
Clients now arrive asking for things they did not ask for two years ago: a support assistant trained on their documentation, a system that reads inbound enquiries and routes them, an automation that keeps two systems in sync without a person copying fields. These requests do not fit neatly into a retainer built around campaigns or websites.
The agency has three options. Decline the work and watch a competitor take the relationship. Hire, which means committing to a salary before the revenue is proven and evaluating candidates in a discipline nobody in-house can assess. Or partner.
Partnering is frequently correct, and not merely as a stopgap. AI delivery spans skills that rarely coexist in one person — retrieval pipelines, prompt and evaluation design, integration against uncooperative legacy APIs, plus the infrastructure to run it reliably. A partner supplies the combination; a single hire supplies one corner of it and learns the rest on your clients' budget.
The Three Models Compared
Fixed-scope subcontracting
The agency wins the project, defines the specification, and contracts the partner to build it for an agreed price.
Strengths. Predictable cost, clean margin arithmetic, no ongoing commitment. Best when you can define the deliverable precisely.
Weaknesses. Every change is a change order, which is friction on projects that genuinely need to evolve — and AI projects evolve more than most, because the first working version reliably changes the client's understanding of what they wanted.
Suits: integrations with clear boundaries, defined automations, discrete builds.
Retained capacity
The agency buys a recurring block of engineering time and directs it across whatever is live that month.
Strengths. Flexibility without renegotiation, a partner who accumulates context across your client base, and priority access when something breaks. Multi-client agencies get the most from this, because unused hours on one account absorb overrun on another.
Weaknesses. You pay for the block whether or not you fill it. Underuse it for two consecutive months and the arrangement stops making sense.
Suits: agencies with several AI clients live simultaneously and lumpy demand.
Embedded partner
Partner engineers join client calls under the agency's brand, as apparent members of the agency team.
Strengths. No translation loss. On complex builds, the round trip from client to account manager to partner and back is where requirements degrade — embedding removes it, and technical questions get answered in the room rather than a day later.
Weaknesses. Demands genuine trust and disciplined brand handling. The partner is talking directly to your client, which is exactly why the contractual protections below stop being paperwork and start being the point.
Suits: large, evolving engagements where requirements are discovered during delivery.
The Margin Maths Agencies Get Wrong
The common error is quoting from the partner's build cost while forgetting everything the agency absorbs around it.
A worked example. A partner quotes £12,000 for a customer support assistant. The agency quotes the client £24,000 and books 50% margin.
What the naive calculation omits:
| Agency cost | Typical hours | At £75/hr |
|---|---|---|
| Pre-sale discovery and proposal | 12 | £900 |
| Client comms and account management | 30 | £2,250 |
| Revision cycles and clarifications | 16 | £1,200 |
| Testing and client-side UAT support | 10 | £750 |
| Total absorbed | 68 | £5,100 |
Real margin: £24,000 − £12,000 − £5,100 = £6,900, or 29% — a respectable number, and a very different one from 50%. Agencies that price against the fantasy figure discover the real one at year end.
Three practices protect the spread:
Charge for discovery separately. A paid discovery phase covers the pre-sale time, qualifies the client, and produces the specification the partner needs to quote accurately. It also converts your riskiest unpaid work into revenue.
Cap revision rounds in the client contract. Not in the partner contract — in the client contract. Unlimited revisions are absorbed by you, not by your partner.
Price recurring work as recurring. AI systems need monitoring, occasional retuning, and updating as the underlying models change. A support retainer priced from the start is straightforward; introduced six months in, it reads as a bill for something the client assumed was included.
The Contract Terms That Matter
Five clauses determine whether the arrangement is safe. Settle all five before the first project, when you have leverage and no live dispute.
1. Non-solicitation. Bars the partner from marketing directly to clients you introduce, for a defined period after the engagement ends. Include what happens if the client approaches the partner unprompted — that is the scenario people forget, and it is the one that actually occurs.
2. Brand and confidentiality. Specify how the partner may reference the work. "Client-approved case study only, no client name without written consent" is a reasonable default. Without this, your delivery partner's portfolio quietly becomes a list of your clients.
3. IP ownership. Decide explicitly among full assignment, perpetual licence with the partner retaining reusable components, or partner-retained ownership. The middle option is often the most honest: partners legitimately reuse internal frameworks across clients, and a contract pretending otherwise is either ignored or priced punitively. What matters is that the client's specific configuration, data, and customisations transfer cleanly.
4. Continuity. Covered in operational terms below, but it belongs in the contract too: repository ownership, credential custody, documentation standards, and a defined notice period.
5. Liability and support. If the system fails in production, who responds, how fast, and at whose cost? Agree a response window and whether it sits inside the retainer or bills separately. Also confirm the partner's insurance covers work delivered under someone else's brand — not all policies do.
Operational Protection: Own the Infrastructure
Contracts resolve disputes slowly. Operational arrangements prevent them from becoming outages. Four rules, all free to implement on day one and expensive to retrofit:
The repository belongs to the agency. Create it in your organisation and grant the partner access. Never the reverse. If the partner owns the repository, a disagreement over an invoice becomes a disagreement over whether your client's system still exists.
Cloud and API accounts are in the agency's or client's name. The partner gets credentials; the partner does not get the account. This also spares you the awkward migration when a client wants to bring the system in-house.
Documentation accrues during the build. "Full documentation at handover" is a promise that competes with a deadline and generally loses. Require a maintained README, environment setup notes, and an architecture summary as ongoing deliverables.
Insist on a walkthrough at milestones. A recorded session where the partner explains what was built and why is worth more than a document, and it surfaces the knowledge concentration risk early — while the person who holds it is still on the project.
Evaluating a Partner
Most agencies assess partners on portfolio and price. Both are weak signals: portfolios are curated, and the cheapest quote frequently reflects a narrower reading of the scope.
Better questions:
- "Walk me through a project that went wrong." The useful answer is specific and includes their own contribution to the problem. A partner who has never had a difficult project has either not done many or is not being straight with you.
- "What would you refuse to build?" A partner with no boundaries will accept an unsuitable brief and bill you for discovering it was unsuitable. You want someone who has said no before.
- "How do you test that the output is correct?" AI systems fail probabilistically rather than loudly — plausible wrong answers rather than crashes. A partner without a real answer about evaluation is planning to let your client find the failures.
- "Who specifically will do the work?" Sales engineers who vanish after signature are a familiar pattern. Ask to meet the actual engineers.
- "What happens if we stop working together?" Watch whether they are comfortable with the question. The answer matters less than the reaction.
A useful cross-check: ask how they would handle a build touching regulated data. A partner who cannot discuss retention, sub-processors, and data residency without preparation is not ready for your regulated clients — the questions in our guide to LLM data security work well as a screen.
Positioning the Work to Your Clients
Two decisions to make deliberately.
Disclosure. Most agencies do not name delivery partners, which is standard practice and generally uncontroversial. What matters is not being caught denying it. "We work with specialist engineering partners for certain builds, and we manage delivery end to end" is accurate, defensible, and closes the subject.
Expectation setting. AI projects benefit from being sold as iterative. The first version establishes what the client actually wants, which is rarely identical to what they specified. Selling that as the intended process rather than as a correction protects the relationship and matches how the work genuinely runs — the same reason we start client engagements with an automation audit rather than a build.
Conclusion
White-label AI delivery lets an agency say yes to work it cannot yet staff, and to do so without gambling a salary on a pipeline that has not proven itself. The model is well established; the failures are predictable and almost entirely avoidable.
They come down to three things: margin calculated without the agency's absorbed hours, contracts silent on solicitation and IP, and infrastructure owned by the partner instead of the agency. Fix those three at the start — when they cost nothing but an afternoon — and the arrangement is straightforward.
If you are an agency with AI demand you cannot currently deliver, book a strategy call. We work behind agency brands, and we are happy to start with the questions above pointed at us.
