Buy a managed service when the scope is closed, the acceptance test is already written, and you are willing to pay 20% to 40% more per engineer-day for somebody else to carry the estimation risk. Buy staff augmentation when the backlog changes faster than you can renegotiate a statement of work, which describes most AI and data work for its first two years.
The two models are usually presented as points on a spectrum of vendor involvement. That framing hides the decision. They differ on one contractual variable: who is liable when the estimate turns out to be wrong.
What each contract actually sells
| Staff augmentation | Managed service | |
|---|---|---|
| Unit you buy | An engineer-day | A defined outcome or service level |
| Who owns the backlog | You | The provider |
| Who absorbs a bad estimate | You | The provider |
| Changing direction | A conversation, same day | A change request, one to three weeks |
| Typical exit | 2 to 4 weeks’ notice | End of the statement of work |
| Where the knowledge settles | Partly in your team | Mostly with the provider |
A managed service costs more per unit of engineering because the provider prices in a buffer against its own misjudgement, plus a delivery manager, a QA function and an account layer that you fund whether the work needs them or not. On equivalent seniority the loaded difference usually lands between 20% and 40%. You can sanity-check the underlying engineering component against our published daily rates for AI and data roles before a vendor tells you the premium is standard.
Where the managed service earns its premium
- The definition of done is testable by someone who is not an engineer. A warehouse migration with a row-count and reconciliation check qualifies. “Improve retrieval quality” does not.
- The work is repeatable operations with a measurable service level: pipeline monitoring out of hours, scheduled retraining, incident response against a response-time target.
- You have no internal engineering manager with capacity to set priorities weekly. Somebody has to own sequencing, and if it is not you, you should be paying for it.
- Procurement wants contractual remedy rather than best efforts, and will accept slower change control as the price.
Where staff augmentation wins
The mechanism is change frequency. In a managed service, every movement in the definition of done costs a change request, and change requests are where providers recover margin and where your calendar goes to die. On a retrieval or agent build, the definition of done moves after each evaluation round, because you learn what the system gets wrong only by running it against real documents and real users.
Count the number of times your target changed in the last quarter. Above roughly one material change a month, the administrative cost of a fixed-scope contract exceeds the risk it removes. That is the point at which buying engineers by the day into your own team becomes the cheaper structure, not just the faster one.
Augmentation also keeps the operating knowledge in-house. When a managed service ends, the people who know why the feature store is partitioned that way leave with it, and the second vendor charges you to rediscover it.
The split most buyers end up with
Run the build with augmented senior engineers under your own lead. Contract run-and-maintain as a managed service with an SLA once the system is stable and the failure modes are known. The boundary is not the go-live date. It is the point at which incidents stop being novel.
Where our model is the wrong choice
If nobody on your side will do sprint planning, staff augmentation fails, and it fails quietly. Rented engineers with no product owner will produce good code against the wrong priorities for two months before anyone notices.
It also fits badly for short closed deliverables. Onboarding a senior engineer into an unfamiliar data estate takes one to two weeks of reduced output. On a six-week piece of work that is a quarter of the engagement, and a fixed-price arrangement will beat it on total cost.
If your governance model requires one supplier accountable for an outcome with financial penalties, a time-and-materials contract cannot give you that. Which type of supplier to approach is a separate question, covered in our comparison of consultancies, agencies and augmentation providers.
Four questions that settle it
- Can a non-engineer verify that the work is finished? If no, augment.
- How often did the requirement change last quarter? Above monthly, augment.
- Who runs this in eighteen months? If your team, augment and keep the knowledge.
- Does the budget survive a 30% scope overrun? If not, buy the fixed scope and accept the premium.
If the answers point to augmentation, our staff augmentation model and contract terms set out notice periods and IP arrangements in full.