Bilingual reference
AI and data glossary, the terms as they are used in Saudi Arabia
One term gets translated five different ways in a single meeting, and the decision gets lost with it. This is a short list of exactly what we mean by each term on Saudi data and AI work.
Each definition also says what the term means in practice in front of Saudi regulators.
Data foundation
- Data governance
- The set of owners, definitions, classifications and access rules that decide who may use which data, for what, and on what evidence.
- In the Saudi context: In Saudi Arabia this is measured against the NDMO national data management controls, so governance has to produce artefacts, not intentions.
- Data classification
- Labelling data by sensitivity at the moment it is ingested, so downstream storage, sharing and model access follow the label automatically.
- In the Saudi context: Saudi classification tiers drive residency: what is labelled restricted or confidential decides what can leave Saudi Arabia, and usually nothing does.
- Data residency
- The rule that specified data is stored and processed inside a defined jurisdiction, enforced by the platform rather than by policy documents.
- In the Saudi context: For most Saudi enterprises this means in-country cloud regions or an on-premises deployment, with an audit trail showing where each record has lived.
- Lakehouse
- A single governed store that keeps raw and modelled data together, giving warehouse-grade tables and machine learning access over the same files.
- In the Saudi context: Usually built on Databricks, Snowflake or Microsoft Fabric in a Saudi region, with classification and lineage wired in from the first pipeline.
- Data lineage
- The recorded path of a value from source system through every transformation to the number an executive reads.
- In the Saudi context: Lineage is what turns a regulator question into a five-minute answer instead of a two-week reconciliation exercise.
- Master data management
- Resolving the same customer, supplier or asset across systems into one governed record with a single identifier.
- In the Saudi context: Arabic name variants, national ID formats and legacy Latin transliterations make matching a Saudi-specific engineering problem.
- Semantic layer
- A certified definition of every business metric and dimension, so revenue or active customer resolves the same way in every tool.
- In the Saudi context: It is also the safety rail for Arabic question answering: the model queries certified metrics instead of guessing at raw tables.
- Data quality
- Measured completeness, accuracy, timeliness and consistency, expressed as tests that run on every load rather than as an annual survey.
- In the Saudi context: Most stalled Saudi AI pilots are quality problems wearing a model costume.
AI and models
- Generative AI
- Models that produce text, code, images or speech rather than only scoring an input, usually built on large pretrained language models.
- In the Saudi context: Enterprise value in Saudi Arabia comes from grounding these models in your own governed data, not from open-ended chat demos.
- Large language model (LLM)
- A model trained on very large text corpora that predicts language, and through that can summarise, extract, classify and reason over documents.
- In the Saudi context: Arabic performance varies sharply by model, so the choice is decided by evaluation on Saudi enterprise Arabic, not by leaderboard position.
- Retrieval augmented generation (RAG)
- Retrieving the relevant passages from your own documents and giving them to the model as context, so the answer cites your material.
- In the Saudi context: For bilingual archives the retrieval step has to handle Arabic morphology and mixed Arabic-English documents, or recall collapses.
- Fine tuning
- Continuing training of a base model on your own labelled examples so it adopts your terminology, format and task behaviour.
- In the Saudi context: Worth doing when the task is narrow and repetitive, for example Arabic claim coding or contract clause extraction.
- Model router
- A layer that sends each request to the right sized model based on task, sensitivity and cost, instead of pointing everything at one frontier model.
- In the Saudi context: It is also a residency control: sensitive prompts stay on in-country models while low-risk work can use a broader portfolio.
- OCR and document AI
- Turning scanned paper and PDFs into structured fields, combining character recognition with layout understanding and validation.
- In the Saudi context: Arabic script, stamps, handwritten annotations and mixed-direction invoices mean generic OCR needs tuning before it is usable.
- MLOps
- The engineering practice that versions data, models and prompts, monitors drift, and lets you roll a model back the way you roll back code.
- In the Saudi context: Without it, an AI system cannot be handed over, and handover is what separates a pilot from production.
- Hallucination
- A confident but unsupported model output. It is controlled by grounding, retrieval, constrained tool use and evaluation, not by prompt wording alone.
- In the Saudi context: In regulated Saudi workflows the answer must carry its source, otherwise it cannot be used as evidence.
Agents and interfaces
- Agentic AI
- Systems that plan a sequence of steps and call tools or business systems to complete a task, rather than only returning text.
- In the Saudi context: Useful only when the agent has scoped permissions inside SAP, Oracle, Salesforce or Dynamics and every action is logged.
- Human in the loop
- A designed approval point where a person reviews or authorises an action before the system commits it.
- In the Saudi context: The practical pattern for financial postings, clinical steps and citizen-facing decisions in Saudi entities.
- Voice agent
- A speech interface that understands a caller, resolves intent against live systems and completes or escalates the request.
- In the Saudi context: Saudi deployment means handling Saudi dialect input while answering in clear MSA, with in-country recording and retention rules.
- NL2SQL
- Converting a plain-language question into a governed query that runs through the semantic layer with the asker's own permissions.
- In the Saudi context: Bilingual by requirement: the same question in Arabic and English must return the same governed number.
- Guardrails
- Enforced limits on what a model or agent may read, say and do, including masking, injection filtering, read-only execution and rate limits.
- In the Saudi context: Guardrails are the part auditors ask to see, so they are built as system controls rather than written policy.
Saudi regulation
- PDPL
- The Saudi personal data protection law, covering lawful basis, purpose limitation, disclosure, transfer and the rights of data subjects.
- In the Saudi context: In delivery it becomes consent and purpose metadata carried through the pipeline, plus deletion and export paths that actually run.
- NDMO controls
- The national data management and governance control domains that Saudi entities are assessed against, from cataloguing to quality to sharing.
- In the Saudi context: Best treated as platform requirements, so each control emits its evidence automatically as a by-product of normal operation.
- SDAIA
- The Saudi authority for data and AI, source of the national data governance framework and the AI ethics principles.
- In the Saudi context: Its ethics principles translate into documented use-case risk assessment, human oversight and explainability expectations.
- NCA Essential Cybersecurity Controls
- The national cybersecurity control baseline covering identity, segmentation, logging, hardening and third-party risk.
- In the Saudi context: AI platforms inherit these controls, which is why model endpoints, prompts and outputs all have to be logged.
- SAMA
- The Saudi central bank, whose cybersecurity and IT governance frameworks bind banks, insurers and payment firms.
- In the Saudi context: For financial AI it sets the model risk, change control and outsourcing expectations any deployment has to satisfy.
- Sovereign AI
- An AI stack whose data, models, keys and operations stay under national control, deployed in-country and operable without foreign dependency.
- In the Saudi context: It is an architecture decision made per data class, not a slogan applied to a whole organisation at once.
Let's put your AI ambition on the table
Sixty minutes with a senior Bilytica partner. We stress-test your plans against what actually works in your sector, agree which moves are worth funding first, and you leave with a written 90-day plan and honest numbers, whether you work with us or not.
Start the conversation