For regulated banks
The system of record for AI answers.
Intrados answers questions from your own policy, procedure and regulatory documents, in Arabic and English, and cites every claim to the exact clause it came from.
An assistant that is confidently wrong is worse than no assistant.
General-purpose AI cannot tell you which clause it relied on, whether that clause is still in force, or whether it was allowed to read it. In a bank those are the first three questions, not refinements.
No provenance
A plausible paragraph, and no way to check it against the document it claims to summarise.
No sense of time
Last year's policy and this year's look identical to a search index. Withdrawn clauses keep answering.
No access control at retrieval
Apply permissions after the search and restricted content has already shaped the answer.
Capabilities
Four things it does that a general assistant does not.
Every claim cites its clause
Not "see the AML policy". Section 3.2.1 of a named policy, at a stated version and effective date. Open the citation and read the passage the answer was built from.
Sources are immutable, and history stays answerable
A new version never overwrites the old one. Both stay queryable, so a question can be asked as of a date, and a superseded clause is never quietly cited as current.
It says when it does not know
Ask something the corpus does not cover and it says so, and names the document it would need. Where two sources disagree it surfaces the disagreement instead of choosing one.
Arabic is a first language, not a translation layer
Arabic documents are indexed as Arabic, with orthographic normalisation so spelling variants match. Answers come back in the language asked, with Arabic citations and a right-to-left interface.
What that looks like
It found a policy conflict nobody had asked it to look for.
A bank raised its enhanced due diligence observation period for politically exposed persons from 12 months to 18. The English policy was updated. The Arabic one was not.
Asked for the current period, Intrados answered 18 months from the English policy, then reported that the Arabic version of the same policy still said 12, gave the effective date of each, and said the discrepancy needed resolving.
Nobody configured that behaviour. It is what happens when a system is built to surface disagreement rather than resolve it silently.
Governance
Built to survive a security review, not to pass a demo.
| Access control inside the search | Permissions are enforced by the database during the index scan. Content a user may not see does not shape their answer, and its existence is not revealed by what comes back. |
|---|---|
| A complete, tamper-evident record | Append-only audit log with a per-event hash chain. Every query records what was retrieved and how, which prompt and model version ran, where the time went and what it cost. Exportable to your SIEM, signed, with optional redaction. |
| Nothing a model infers is trusted automatically | Relationships the system extracts from your documents wait in a review queue for a person. Until approved they influence no answer. |
| A fail-safe you can actually reach | AI operations can be disabled without a redeploy, for one system, one risk level, or the whole tenant. It takes effect immediately and never reaches a model. |
| Human review as a right, not a support ticket | Any answer can be escalated by whoever received it, routed by risk to the correct queue with a deadline. |
| Your data does not train a model | No fine-tuning and no adapters. Everything it knows about your policies is retrieved at query time and cited. |
Built against the frameworks you are examined on: the CBJ AI Governance Framework in Jordan, the CBUAE guidance on responsible AI in the UAE, and the SAMA, PDPL, NDMO and SDAIA layers in Saudi Arabia. Most of what they require overlaps, so Intrados implements the shared requirements as machinery and treats the differences as configuration.
The product enables your compliance team to satisfy your regulator. It does not claim to discharge your obligations, and it never makes a legal determination in code. It records facts; your people make findings.
Deployment
Your data stays where your policy says it stays.
One database
Documents, search index, knowledge graph and audit log all live in one PostgreSQL. No second datastore to put through procurement.
In your own tenancy
Deploy into your cloud account, in your region, under your identity provider. No design choice here depends on a specific cloud provider.
Or nothing leaves the building
A fully on-premise configuration with self-hosted models, for an air-gapped environment, with the quality difference documented rather than hidden.
What it deliberately does not do.
- It answers questions. It does not take actions in your systems.
- It does not fine-tune on your data.
- It does not reuse one person's answer for another's question, because a near-miss returns the wrong answer carrying a real citation.
- It does not hide uncertainty to look more capable.
See it answer a question from your own policies.
A working session takes an hour, on a corpus we load beforehand.