Most of an AI integration budget goes to the systems around the model, not the model. On the work we staff, 70% to 80% of engineer days land on source-system access, data preparation, evaluation and production plumbing. The prompt and the model call are the cheap part, and that one fact should decide how you scope, price and staff the job.
What you are paying for
| Workstream | Share of effort | What eats the days |
|---|---|---|
| Access to source systems | 20-30% | Undocumented internal APIs, auth, rate limits, batch export windows |
| Retrieval and data preparation | 20-25% | Chunking, deduplication, keeping row-level permissions inside the index |
| Model call layer | 10-15% | Prompts, tool definitions, structured output, retries |
| Evaluation and regression tests | 15-20% | Building the test set, agreeing what a correct answer is |
| Production plumbing | 15-20% | Tracing, cost caps, timeouts, behaviour when the provider is down |
| Changes to the calling application | 10-15% | Review queues, audit trail, the screens a user actually sees |
If a proposal prices the model call layer in detail and folds everything else into “integration support”, the estimate is wrong by a factor of three. The same arithmetic seen from the project side is in our note on what drives AI development cost.
What moves the price
- Number of source systems. Each additional system with its own auth model adds roughly 5 to 15 engineer days before any AI logic exists. Systems whose original owners have left the company cost more than that.
- Permissions. If two users must get different answers from the same corpus, the access control list has to live inside retrieval, and that workstream roughly doubles. The mechanics are in our guide to how retrieval-augmented generation is assembled.
- Latency. A target under two seconds end to end forces caching, smaller models or parallel retrieval, which is 10 to 20 engineer days you do not spend on an asynchronous design.
- Evidence standards. In finance, health and the public sector the evaluation set gets signed off outside engineering. Expect four to eight weeks of calendar time that no amount of staffing compresses.
- Inference volume. Below roughly 100,000 requests a month, model spend is noise against engineer cost. Above a few million, someone owns routing, caching and quotas as a standing job.
Who sells this work, and how each category fails
- Global systems integrators price the programme, staff it thin at the senior end, and bill the coordination. Sensible when you need 40 people across five countries. Expensive when you need six engineers.
- Big-4 consulting practices sell the assessment and the target operating model well. Delivery is frequently subcontracted, so ask who will sit on the standup before you sign.
- Platform implementation partners optimise for consumption of the platform they resell. You get fast progress inside one vendor’s stack and a bill that is hard to move afterwards.
- Boutique AI agencies build the demo well. The regression test suite is where they thin out, so ask how they will prove the system still works after the next model version.
- Talent marketplaces supply individuals at a low hourly rate with no bench, no replacement cover and nobody accountable for the architecture those individuals produce.
Fixed price or time and materials
Fixed price works when the interfaces are documented, the data already sits in one place, and the acceptance test exists in writing. That describes maybe one integration in five. For the rest, a fixed-price bid gets padded 30% to 50% against unknowns, and every discovery turns into a change request negotiated by people who are not engineers. Buying engineers on a time-and-materials basis with a checkpoint every two weeks keeps the discovery cost visible while direction is still cheap to change.
Where renting engineers is the wrong choice
- You have no internal product owner. Without someone who can decide what a correct answer looks like, rented engineers build to their own guess and you pay to rework it.
- The scope is one documented connector. Buy that fixed price.
- You need the result operated around the clock. Staff augmentation covers build and business-hours support; a run contract is a different purchase with different economics.
- Procurement wants one supplier accountable for a business outcome rather than for engineer days. That is a consulting engagement, and you should pay consulting prices for it.
Two numbers settle most of these decisions early: the day rate for the seniority you need, which sits on our current daily rates, and an honest count of the source systems in scope.