Detect
Identify direct identifiers and selected combination risks in Arabic and English text.
A privacy gateway for AI.
Use any AI model. Keep your customers' identities in the Kingdom.
DataSitr checks each request for personal data, applies tenant policy, and routes the resulting payload to a configured provider—or blocks the request.
Synthetic example
Policy checked
What your team typed
Summarize the order for Aisha Al‑Otaibi, phone +966 50 000 0000.
What the model receives
Summarize the order for [[PERSON:01]], phone [[PHONE:01]].
PERSON · PHONE
Green route eligible
Reads
Arabic and English
Arabic names, Saudi number formats, and the usual identifiers in both languages.
Decides
Three plain outcomes
Send with placeholders, keep in the Kingdom, or block. You can read the rule for each.
Keeps
A record of every decision
Dated, hash-chained, and readable by your DPO or a reviewer.
Shows its work
181 controls, published
Each tied to a test or a dated piece of evidence. Benchmark and trust report, no sign-in.
How it works
One request. One policy decision.
The request is checked before a provider route is selected.
Minimize
Replace direct identifiers with typed placeholders when policy allows.
Route
Route the permitted payload to a configured external provider, use a configured in-Kingdom path, or block the request.
In
A request from your team or system
Exactly as written, in Arabic or English.
DataSitr
Detect, minimize, decide
Original values go to an encrypted vault in the Kingdom and come back into the answer on the way out.
Green
Tokenized before eligible external processing.
Amber
Pseudonymized, then routed to a configured in-Kingdom path.
Red / block
Raw sensitive content goes only to a configured in-Kingdom path; otherwise the request is blocked.
Tokenization reduces direct-identifier exposure. It does not remove every contextual or legal transfer risk.
What it finds
The identifiers that actually show up in Saudi work.
A model fine-tuned for Arabic, plus pattern recognizers for Saudi formats. Not a generic English filter with Arabic bolted on.
National ID
Iqama number
Passport number
IBAN
Saudi phone numbers
Names, Arabic and English
Addresses
Commercial registration
Email addresses
Card numbers
Dates and birthdays
Medical record numbers
Combinations that identify someone
Production detection threshold: 0.7 confidence. Amber marks combination risks, which are flagged rather than replaced. Measured results are on the benchmark page.
Who this is for
Teams whose customer data is already reaching an AI.
Usually through a chat window, a summary tool, or a vendor's feature. The gateway puts a policy in front of it. The assessment tells you where you stand first.
Pilgrim files pass through many hands.
Passports, health forms, and visa details for whole groups. Ask an AI to summarize a group file and the identifiers ride along.
Which third parties receive them, and what leaves the Kingdom?
Patient notes are written in Arabic.
Medical record numbers, dates of birth, and diagnoses in free text. A transcription or summary request should not carry them out of the Kingdom.
Can you show what the model actually received?
IBANs and IDs hide in support tickets.
Claims histories, underwriting notes, and customer messages get pasted into AI tools because it is faster. The gateway keeps the fast part and removes the identifiers.
Who gets called at hour zero of a breach, and is the 72-hour path written down?
Customer lists and consent, side by side.
SMS campaigns, loyalty data, and support chat. The records the gateway keeps are the ones a reviewer asks to see.
Can you show consent for each message you sent?
Start here
Start with a 10-business-day readiness assessment.
If a PDPL notice arrived tomorrow, could your team produce the evidence within five days? The assessment answers that on paper, with fixed scope, documents and synthetic material only, and four deliverables your team can act on.
01
Findings register
Risk-ranked, with evidence references and named control owners.
02
Control map
Mapped to PDPL requirements with implemented, partial, missing, and evidence-gap states.
03
Remediation plan
Ordered by risk, dependency, and practical sequence.
04
Executive readout
Material exposure, immediate decisions, and the next evidence gates.
Evidence
Inspect the work.
Published benchmark snapshots, system boundaries, and deployment details are available without a sign-in.
Straight answers
The questions people ask first.
Do you see our data?
For the assessment, no. It uses documents plus synthetic or redacted material, with no access to your systems and no production personal data. For the gateway, original values stay in an encrypted vault hosted in the Kingdom; an external model receives the placeholder version, and only when your policy allows it.
Which AI models can we use?
Any provider configured for your tenant. The pilot has routes for OpenAI, Anthropic, Google, Groq, and STC SambaNova, and in-Kingdom paths are configured per tenant. The full provider list is on the sub-processors page.
Is DataSitr approved by a regulator?
No. DataSitr is registered on the National Data Governance Platform as a data services provider, which means registered, not licensed. The status of our accreditation application is written up with dates on the Trust page, and we will not describe it as anything more than it is.
How big is DataSitr?
One founder-operator in Jeddah, who is the registered DPO. The gateway runs as a live pilot, not a proven high-availability platform. We say this plainly because you would find out anyway, and because the evidence pages are built to survive that question.
What does the assessment cost, and how long does it take?
A fixed fee in two milestones, ten business days from accepted intake, and at most two starts a month. The exact figures and payment terms are on the assessment page, so that nothing here is a surprise on the call.
What happens after I write to you?
You hear back from the operator, not a sales team. We hold a 20-minute call to check fit and pick a start date, then you receive a written scope. No documents are needed for the call.
Where we are
Small, honest, and in the Kingdom.
The gateway runs as a live pilot. Continuity has been drilled on dated occasions; it is not a standing availability guarantee.
Every control we claim is tied to a test or a dated piece of evidence, and the trust report lists what it does not prove.
The assessment is a technical and operational review. Regulatory standing, in writing, is on the Trust page.
Evaluation
Evaluate it on one defined workflow.
Start with a scoped pilot and an agreed evidence plan.