Skip to content
All insights

Guide6 min readAahlan team

Private AI or a public API? A decision guide for Saudi organisations.

A practical framework for deciding which AI workloads can use a public model API, which belong on capacity you control, and how to run both without losing track of your data.

  • Private AI
  • Strategy
  • Security

Most AI programmes reach the same fork early on. The quickest route to a working prototype is a public model API: sign up, send a prompt, get an answer. The route that keeps data under your control is a model hosted on capacity you run or reserve — on-premises, or in a private cloud region in the Kingdom. Both are legitimate. The mistake is choosing one for everything, or choosing by habit. This guide sets out the questions we work through with organisations, in the order that tends to settle the decision fastest.

Start with the data, not the model

The first question is not which model is best. It is which of your data may leave your boundary at all. Most organisations already have a classification scheme — government entities work to the NDMO framework, and banks, hospitals and many private companies keep their own. Map each candidate use case to the most sensitive class of data it will touch, including what users are likely to paste in, not just what the system is designed to receive.

  • Public: published content, marketing copy, public regulations. Usually fine to send to a reputable API.
  • Internal: procedures, internal e-mail, project documents. Often acceptable with contractual controls on retention and training use.
  • Confidential: contracts, pricing, board papers, source code. A case-by-case decision, and frequently a no.
  • Personal and regulated: customer records, health and credit data, anything covered by the PDPL or sector rules. The default is to keep it inside the boundary unless your legal team has approved a specific transfer.

Cost: renting per token versus owning capacity

A public API charges for what you use, usually per token in and out. There is no upfront cost and no charge when nothing runs, which makes it the natural choice for pilots, low volumes and demand that comes in bursts. The catch is that the bill grows in a straight line with usage — and usage tends to grow faster than planned. Retrieval-based assistants send long prompts full of document context, agents make several model calls per task, and Arabic text often costs more tokens per word than English on tokenisers built mainly for English.

Owned or reserved capacity works the other way round. You pay for servers or reserved GPUs, power, networking and the people who run them, whether the capacity is busy or idle. Each additional request costs almost nothing until you reach the ceiling, so the cost per request falls as utilisation rises.

Break-even is where the two lines cross. It comes earlier when demand is steady and high, when prompts are long, when a smaller open-weight model is good enough for the task, and when one deployment can serve several use cases. It comes later when volume is uncertain, when you need the largest frontier models, or when you do not yet have the skills to operate GPUs. Model it with your own traffic estimates — and include the operating effort, the line item most often left out.

Latency, availability and portability

Latency depends on distance and on queueing. A model hosted in the Kingdom avoids the round trip to a distant region, which matters for voice, live chat and agents that make many calls in sequence. On a shared API you also share the queue: rate limits, regional outages and slow periods are outside your control. On your own capacity you control the queue, but you also own the capacity planning — if demand doubles on a Sunday morning, nobody else absorbs it.

Model choice is the other half. The strongest proprietary models are often available only through their vendor’s API. Open-weight models such as Llama, Qwen and Mistral, and Arabic-first models such as ALLaM and Jais, can run wherever you decide. API vendors also retire models on their own schedule, which can change the behaviour of a system that was working. Whichever route you take, keep prompts, evaluation sets and integration code independent of any single model, so that switching is a configuration change rather than a rebuild.

Arabic quality and operational burden

Do not assume the model that tops an English benchmark is the best for your Arabic work. Quality varies widely across Modern Standard Arabic, Saudi dialects, mixed Arabic–English text and scanned documents. Test candidates on a small set of your own real tasks before you compare prices, and measure token counts on your own Arabic text while you are at it. We cover how to build that test set in a separate guide.

Then be honest about operations. Running models privately means someone patches the servers, updates drivers and serving software, monitors quality and latency, manages keys and access, and plans upgrades. A public API takes the infrastructure off your hands, but not the rest: you still own the integration, the evaluation, the logging and the controls on what gets sent. The question is not whether there is operational work, but how much of it you want to carry and who will do it.

The two options side by side

FactorPublic model APIPrivate model (on-premises or private cloud)
Data exposurePrompts and outputs processed by an outside provider under its termsStays inside infrastructure you control
Cost shapePer token; no upfront cost; grows with usageLargely fixed; upfront or reserved commitment; cost per request falls as usage rises
Best-fit demandLow, uncertain or burstySteady and high
Model choiceThe vendor’s catalogue, including frontier proprietary modelsOpen-weight and Arabic-first models you can deploy
Latency and availabilityShared queue, rate limits, provider outagesYour queue and your capacity planning
Model changesOn the vendor’s scheduleOn yours
Operational burdenIntegration and governanceIntegration and governance, plus infrastructure, serving and upgrades

Most organisations end up hybrid

In practice the answer is rarely one or the other. A common pattern is a single gateway in front of every model. Each request is tagged with the data class of its source — the system, the document library, the user’s role — and routed accordingly: personal and confidential work to a private model, public and approved internal work to an external API where that makes sense. Where external calls are allowed, names and identifiers can be masked before the request leaves and restored when it returns. One set of access controls and one audit log cover both routes, so you can always show what went where.

A decision checklist

  1. 1List the use cases and the most sensitive data class each one touches.
  2. 2Agree which classes may leave your boundary, and on what contractual terms — with your legal and security teams, not around them.
  3. 3Estimate volume, prompt length and growth for each use case, then compare per-token cost with owned capacity, including the operating effort.
  4. 4Test two or three candidate models on your own Arabic and English tasks.
  5. 5Check latency and availability needs: live chat, voice and agents are less forgiving than overnight batch jobs.
  6. 6Design the gateway, routing rules and audit log before the second use case, not after the fifth.
  7. 7Keep prompts and evaluation sets portable, and revisit the decision when volumes or models change.

If it would help, this is the kind of assessment we carry out at the start of an engagement: classifying the data, modelling both cost curves and testing candidate models on real work, so that the infrastructure decision follows from evidence rather than preference.

Want to talk this through?

We are happy to discuss how this applies to your organisation — no sales script.

Newsletter

Practical AI, in your inbox.

Real use cases, Saudi regulations and what works in Arabic — once or twice a month. No spam; unsubscribe any time.

We’ll e-mail you to confirm. Unsubscribe any time — see our privacy policy.