
Buy when the task is common to every company, assemble on top of a model API when the value comes from your own data and workflow, and build your own models only when the capability itself is what customers pay you for and you have the data and people to sustain it. In my view the middle option fits most business cases. "Build vs buy" is really three options, and the costly mistakes come from treating it as two.
This is the framework I recommend to leadership teams before a budget is approved. It avoids prices and product names on purpose, because both change faster than a strategy should.
The three options
- Buy: a finished product with AI inside, such as an assistant in your office suite, a support tool or a coding assistant. You configure it; you do not engineer it.
- Assemble: your own application built on a model you access through an API or a cloud platform, combined with your data, prompts, retrieval, tools and evaluation. You own the workflow and the user experience. The model is rented.
- Build: training or substantially fine-tuning a model, and running it yourself. You own the capability and all of its costs.
Six questions to ask
1. Does this set us apart?
If every competitor needs the same function, buy it and move on. Spend engineering time where a customer would notice the difference.
2. Is our data the advantage?
Proprietary data that is clean, permitted for this use and hard for others to obtain is the best reason to assemble or build. If the data is thin or messy, the first project is the data, not the model.
3. What happens when it is wrong?
A wrong draft email costs little. A wrong credit decision or medical summary costs a great deal. The higher the cost of an error, the more control you need over evaluation, human review and audit trails, and the less a closed product will satisfy you.
4. Where may the data go?
Check contracts and data protection rules before choosing a vendor: where data is processed, whether it is retained, and whether it is used for training. I covered the UAE side in my guide to the AI rules that apply in the UAE.
5. Who will run it in a year?
A model in production needs monitoring, evaluation, updates and someone on call. If you cannot name the team that will own it after launch, you are not ready to build.
6. How fast must it work?
Buying is the fastest route. Assembling takes longer and needs engineers. Building takes longest and needs specialists. Match the option to the deadline that the business actually has.
The costs people miss
- Evaluation. You need a test set of real cases and a way to score answers before and after every change. Without it you cannot tell whether a new model or prompt is better.
- Integration. Connecting to your systems, permissions and data often takes more effort than the AI part.
- Usage. Costs that look small in a pilot grow with adoption. Model the cost per task at full use, not per month in the trial.
- Change. Models are replaced and retired by their providers. Plan to re-test when that happens.
- People. Staff need training and a reason to change how they work. A tool nobody uses has no return.
Protect yourself from lock-in
You do not need to avoid vendors. You need to be able to leave one. Four habits make that possible:
- Keep your prompts, evaluation sets and business logic in your own repository, not inside a vendor's console.
- Put a thin layer between your application and the model so that a second provider can be added without a rewrite.
- Own your data pipelines and embeddings, and know how to rebuild them.
- Read the exit terms: data export, deletion and notice periods.
A simple way to decide
- Write the use case in one sentence, with the person who benefits and the measure that should move.
- Check whether a product you already license does it. If it does well enough, stop there.
- If not, assemble a pilot on a model API with a small evaluation set and real users.
- Measure for a fixed period. Keep it, change it or stop it.
- Consider building a model only when the assembled version has proven the value and has reached a limit that only your own model can remove.
Frequently asked questions
Is fine-tuning the same as building?
It sits between assembling and building. Light fine-tuning of a hosted model is closer to assembling. Training and hosting your own model, with the team that requires, is building.
Should a startup build its own model?
Only if the model is the product and the founders have the expertise and data for it. Otherwise assemble, prove demand, and revisit later.
How do I compare vendors fairly?
Run the same evaluation set through each candidate, using your own cases. Compare quality, cost per task, speed and contract terms. Public rankings are a starting point, not a decision.
Can we change our choice later?
Yes, if you kept prompts, data and evaluation under your control. That is the practical reason for the lock-in habits above.
Where I can help
I am Ali Fouad Kwika, a Dubai-based CTO and founder who has worked in technology since 2011. My doctoral research at Clermont School of Business was on AI adoption in GCC firms. I help leadership teams choose AI use cases, run pilots and set up the engineering around them. See my AI consulting in Dubai, fractional CTO and digital transformation consulting services, or contact me.
Reviewed on 15 November 2026. This article states no prices, versions or statistics, because they change quickly; check current vendor terms before deciding.