Three ways to outsource AI development
“Outsourcing” covers three different purchases, and most disappointment comes from buying one while expecting another.
| Managed delivery | Team extension | Hiring in-house | |
|---|---|---|---|
| You buy | An outcome, at a fixed scope | Capacity, on your backlog | A person, permanently |
| Our price | $15,000–$30,000 for a feature; $40,000–$120,000 for a product | $4,000–$7,000 per engineer per month | A salary; recruiting into your team is a separate service of ours |
| Time to start | Written scope within 48 hours of the discovery call | 1–2 weeks from signing | Months, realistically |
| Who directs the work | We do, against the agreed scope | You do | You do |
| Right when | The result can be written down | You have a product owner and the gap is AI capacity | AI is a permanent part of the product |
| Wrong when | The scope is still moving | Nobody on your side can prioritise | You need something shipping this quarter |
Outstaffing and staff augmentation are the same thing as team extension, under the names different markets use for it — every role we staff that way. The engineers work in your repo, your standups and your review process, and you hold 100% of the IP from the first commit.
What an outsourced AI system costs to run
A build quote is one number. The monthly bill once it is live is another, and it is the one that decides whether the feature survives its first budget review. Ask for it before you sign, and ask how it was measured.
One production system, measured. The multilingual logistics platform behind our AI case study runs at about $0.011 per answer across roughly 40,000 queries a month — about $0.002 of retrieval and $0.009 of generation. Answers take 1.9 seconds on average and 4.2 at p95, measured to the last token. The split is the transferable part: what retrieval really costs arrives as extra input tokens in the generation line, so context length is the cost lever, not the vector database.
The five risks, and how to price them in
- No evaluation set. Without scored test cases nobody can say whether the system works, only whether it demos. On that platform 99.2% extraction accuracy is a measurement because a harness was built to take it — and the harness is the line item most often cut to hit a price.
- Data nobody audited. Clean, permissioned data means a retrieval pipeline in weeks; scanned documents in four languages with no metadata means the data work is the project. A quote given before anyone has seen a random sample of your documents is a guess.
- Lock-in. Put the model provider behind one interface so changing it is days of work, and get the code, the prompts and the evaluation set assigned to you in the contract — not licensed.
- Data that cannot leave your boundary. That is a design constraint, not a clause: it means a self-hosted pipeline, and it changes the architecture and the price. Say it on the first call.
- No baseline. Measure the process you are replacing before you replace it. We did not on that platform, so we publish 99.2% as what the system scores, not as an improvement on anything. It costs an afternoon beforehand and cannot be bought back afterwards.
Six questions to ask any AI outsourcing vendor
- What does one answer cost to run on a system you shipped, and how was it measured?
- Is your latency figure to the first token or the last?
- Who writes the evaluation set, and when in the project does it exist?
- Who owns the code, the prompts and the evaluation data on the day they are written?
- What changes if our data cannot leave our infrastructure?
- What did you measure before the system you are showing us replaced anything?
A vendor that answers the first two with a number and its conditions has built and run something. One that answers with a range of model prices has not yet.
When outsourcing AI is the wrong call
- The data does not contain the answer. No vendor fixes missing information. Sample the real corpus before buying anything.
- Nobody can say what a correct answer is. Then there is nothing to evaluate against, and the project becomes an argument about opinions. Settle the definition first.
- AI is the core of your product for years. Then you want the knowledge inside your company. Extending your team while you hire is a sensible bridge; outsourcing the core indefinitely is not.
If none of those apply, the choice comes down to the table above. We sell all three: managed AI development, including AI integration into an existing product, and AI engineers on your team.