The situation
Digital agencies and consultancies are increasingly winning work that includes AI: a knowledge assistant inside a client portal, an automated document workflow, an agent that handles part of a service process. The proposal was written by people who understand the client’s business. The delivery needs someone who has taken an AI system into production and knows where it fails.
Hiring for that takes months, and the work cannot wait. The usual alternatives each carry a risk: a white-label subcontractor the client never meets, a staffing supplier who swaps engineers mid-engagement, or the existing team learning on the client’s time. The question is how to bring senior AI delivery into the engagement while keeping accountability clear, to you and to the client.
How the co-delivery works
Who owns the client. You do. You lead the relationship, the commercial conversation and the engagement governance. I do not discuss commercial terms with your client unless you ask me to.
Who contracts. You contract with your client; I contract with you as a disclosed subcontractor. Your client’s terms on confidentiality, IP, data protection and security flow down to me through our agreement.
Who delivers. Your team delivers the engagement. I deliver the AI technical work assigned to me in the responsibility matrix, such as retrieval design, agent controls, evaluation and integration, through your client’s review and change processes. Your engineers pair with me so the capability stays with your firm.
How specialists are approved. If another specialist would help, for example a security reviewer for a penetration test or a data engineer for a large migration, I name them, describe their role and basis, and they start only after you and the client have approved in writing. They appear in the responsibility matrix like everyone else. There is no bench of unnamed engineers behind this.
How payment works. I invoice you, on the contract day rate or as scoped delivery with milestones tied to deliverables, on terms agreed in writing before work starts. You invoice your client on your own terms. Expenses are pre-approved.
The signature deliverable, illustrated
The responsibility matrix is the document everyone signs up to. Illustrative example, not taken from a client engagement:
| Area | Agency | Dipankar | Client |
|---|---|---|---|
| Client relationship and commercial terms | Owns | Informed | — |
| Scope and acceptance criteria | Owns and agrees with client | Drafts AI technical scope | Approves |
| AI architecture decisions | Consulted | Owns, within agreed scope | Approves where it affects their systems |
| Implementation | Owns non-AI work | Owns assigned AI work | Reviews through their process |
| Data access approvals | Requests | Uses only as approved | Owns |
| Specialist approval | Approves | Proposes, by name | Approves |
| Handover | Coordinates | Delivers technical handover | Names the internal owner |
The client-facing scope that accompanies it states the AI deliverables, acceptance tests, data the work depends on, exclusions and the change process.
How acceptance is judged
The client accepts against the criteria in the client-facing scope, through your engagement governance. For AI work that usually means an evaluation run on an agreed set of cases, controls demonstrated live, and documentation handed to the named owner. You accept my work against the same criteria, so there is one definition of done.
Ownership and handover
The delivered system belongs to the client as your contract specifies. Technical handover goes to the client’s named owner with your team present, so you can offer continuing support if the client wants it. My involvement ends at handover unless you extend it.
When to choose something else
If you are filling a single contract role, the recruitment agencies route is faster. If you are an MSP wanting a repeatable adoption service, see AI enablement partner for MSPs. To draft the client-facing scope, the project scope and acceptance template helps. All partner routes are compared on the partners overview.