AI Implementation Partner for Agencies: Three Models

An agency can work with an AI implementation partner in exactly three ways. Referral: the partner contracts your client directly and you take a fee. White label: you contract the client under your own brand and the partner builds without ever appearing. Co-delivery: both firms are named to the client and the scope is split down the middle in writing. Which one fits depends on how often AI requests actually reach you, whether you employ anyone technical, and how much of the client relationship you are prepared to put in someone else’s hands.

The three models, defined

Referral moves the client. You make an introduction, the partner signs them, you are paid once and you are out of the delivery chain. White label moves the work but not the client: your contract, your invoice, your brand, and a subcontract behind it that the client never sees. Co-delivery moves neither cleanly: the client knows both firms, and each firm owns named scope lines it is separately accountable for.

The distinction that matters is not branding. It is who the client sues when the system stops working at 2am.

The three models side by side

	Referral	White label	Co-delivery
Holds the client contract	Partner	Agency	Agency, partner named
Brand the client sees	Partner’s	Agency’s	Both
Who speaks to the client	Partner	Agency only	Both, on defined topics
Agency revenue shape	One fee, once	Full spread on the build	Spread on your scope lines only
Carries delivery risk	Partner	Agency, backed by a fixed partner price	Split by scope line
Needed in-house	Nothing	Someone who owns scope	Scope owner plus delivery staff
Usual failure mode	Partner wins the adjacent work	Partner quotes time and materials	Nobody owns the seam between the two halves

Referral fits when AI requests are rare in your client base

Look at what your clients do before you build a partnership around them. The US Census Bureau’s Business Trends and Outlook Survey, reported 26 May 2026 and covering 14 December 2025 to 3 May 2026, put AI use at 39.7% of firms in Information and 33.9% in Finance and Insurance, against roughly 14% in Retail Trade.

That spread decides the model. If your book is software and financial services, AI requests arrive often enough to justify standing up a repeatable delivery motion. If your book is retail and hospitality, you may see two such requests a year, and building white label machinery for two projects costs more in process than it returns in margin. Take the fee, keep the goodwill, spend your attention elsewhere.

Referral has one honest downside and agencies consistently underprice it: you are handing a competent technical firm a direct line to your client. Cap that risk with a written non-solicit on adjacent services, or accept it and price the fee accordingly.

White label fits when the work is regular and the client must see one company

This is the standard answer for agencies with steady demand and no engineers, and we have written the mechanics of it separately in white label AI automation, including what to require from a partner before your logo goes on their build.

The short version: the partner quotes you a fixed number before you quote your client, so your margin is known on signature day rather than discovered at handover. If a prospective partner will only work time and materials, white label is not available to you with that partner, whatever they call it in the proposal.

Co-delivery fits when you already employ technical people

Co-delivery exists to keep your own delivery staff busy while a partner covers the part you cannot staff. Design work, integration into the client’s existing site, change management and training stay with you. Agent design, backend build, evaluation and the operations layer go to the partner.

The AICPA and CIMA 2025 National MAP Survey, published December 2025 from 1,073 accounting firms with FY2024 data, gives a useful sense of scale here. Firms in the top quartile by net remaining per partner ran 5.78 billable professionals for every equity partner, against 3.00 across all respondents. Third-party capacity is how firms reach ratios like that without hiring ahead of demand.

The same survey shows partnering scales with size: 74% of firms above $10M in net client fees used offshoring, against 29% of all respondents. Small firms refer. Large firms build a delivery bench and fill the gaps. Co-delivery is the middle of that curve.

Technology work is the least subcontracted line, and that is the opening

The MAP survey asked which service lines firms actually send outside. Individual tax led at 51%, business tax 42%, client accounting services 38%, audit 28%. Technology consulting came in at 8%.

Read that as unclaimed ground rather than as proof it cannot be done. The demand underneath it is documented. Eurostat’s e-business release of 20 May 2026, covering the 2025 reference year, found 53% of EU enterprises used specialised e-business software, but the split by size is severe: business intelligence software ran 11% at small enterprises against 69% at large ones, ERP 41% against 89%, and CRM 25% against 65%.

Those gaps are the work. Mid-market and small companies are missing exactly the reporting and data layer that AI agents need in order to be useful, which is why so many AI requests to agencies turn out to be data plumbing requests wearing a different hat.

No credible dataset exists on referral fees, so use published prices

We looked. There is no tier one or tier two source on white-label margin norms, channel referral percentages, or agency reseller economics for AI work in 2025 or 2026. Everything circulating on those questions comes from vendors selling partner programmes, with no methodology attached. We are not going to cite it and neither should you.

What you can price against is published numbers. Ours are on one page: an operations audit at $1,900 over ten working days, an AI Workforce Sprint from $9,500 over four to six weeks, and an ongoing fractional service at $2,900 a month. A partner who will not publish or fix a price cannot be resold at a predictable margin, because you would be quoting a number they have not committed to.

Worth saying plainly: most of what we build is direct for the operating company that will run it, and partner-routed work is a second route in, not the business. That is deliberate. A partner channel built on top of a studio with no direct clients has nothing behind it to learn from.

What all three models need in writing

Whichever model you pick, three terms carry it. A fixed price against a fixed scope, agreed before you quote your client. Full handover of code, repositories, credentials and documentation to the end client, so the build cannot be held hostage later. And stated failure behaviour: what gets logged, what stops a run, who owns each agent, and what happens when a malformed input arrives overnight.

That last one separates the partners worth reselling from the rest. IBM’s Institute for Business Value 2025 CEO Study, published 6 May 2025 from 2,000 CEOs across 33 countries, found only 25% of AI initiatives had delivered the expected return, while 85% of those CEOs expected positive returns on scaled AI efficiency investments by 2027. The gap between those two numbers is operational discipline, not model quality.

We publish ours because it is the thing being resold. Our agent fleet completed 5,450 runs with zero failures over 13 days, measured to July 2026, across 41 registered agents with 9 live in production. That holds because each one is a tomte, our word for one production agent with one defined job, an owner and a kill switch. If you would rather automate your own delivery than resell someone else’s, that is a different exercise, covered in AI automation for agencies.

Send us a scope you cannot staff and say which of the three models you want to run it under. You get a fixed number, a delivery window and the model spelled out, in writing, within one business day. No meeting required. Start async.