- 01
Market
Saudi Arabia
- 02
Written for
CIO, Head of Digital
- 03
Decision
Give staff a single assistant over internal knowledge and systems
- 04
Languages
EN · AR
Access
What has to be true before the first question
Four properties. An assistant missing any of them will eventually answer something it should not, and the incident will be discovered by the person who receives the answer.
Permissions at query time
Retrieval filters on the asker's identity against the source system's own access control, not against a copy made at indexing time.
How it is verified: Two accounts, one question, different answers — tested as a release gate.
Citations
Every answer names the document and version behind it, so a reader can check rather than trust.
How it is verified: An answer with no source is a defect, not a stylistic choice.
Freshness
Superseded policies are removed from the index, not ranked below the current one. Old documents are the most common wrong answer.
How it is verified: A retirement process owned by whoever owns the policy.
A refusal that is honest
Where the corpus does not answer the question, the assistant says so and points at who does, instead of assembling something plausible.
How it is verified: Measured on a held-out set of questions the corpus genuinely cannot answer.
The corpus
Most of the work is deciding what not to index
A shared drive holds the current policy, three drafts of it, a version somebody exported to comment on, and a folder from a project that ended in 2019. Indexing all of it produces an assistant that is confidently wrong in a way that is hard to detect, because every answer has a real document behind it.
So the first phase is unglamorous and decides the outcome: naming an owner for each document set, marking what is current, and excluding what is superseded. Organisations that skip it get a demo that impresses and a rollout that is quietly abandoned within two months.
Adoption
Three reasons staff stop using it
- 01
It was wrong once, visibly
One confidently wrong answer about leave entitlement costs more trust than fifty correct ones build. Citations are the mitigation: an answer that shows its source is checkable, and a checked answer survives being wrong.
- 02
It cannot answer the questions people have
Staff ask about process and entitlement — what am I owed, who approves this, what is the current limit. A corpus of strategy documents answers none of them, and the assistant is judged on the first week of real questions.
- 03
It is slower than asking a colleague
The competition is not a search engine; it is a message to someone who knows. An assistant has to be faster than that and has to be reachable where people already are, which usually means the chat tool rather than a new portal.
Our role
What we build, in what order
- 01
Permission-aware retrieval
Access enforced at query time against the source systems, with the two-account test in the release pipeline rather than in a checklist.
- 02
Corpus curation with owners
Each document set gets a named owner and a currency marker, and superseded material leaves the index rather than being outranked.
- 03
Evaluation on real questions
Built from what staff actually ask the service desk, including the ones the corpus cannot answer, so refusal is measured as carefully as accuracy.
Policy interpretation stays with the function that owns the policy; the assistant's job is to put the current document in front of the person asking, with the source named.
Common questions
What buyers ask
Can staff see things they should not?+
Only if the assistant was built to retrieve first and filter afterwards, which is the pattern to avoid. Retrieval filtered by the asker's own access, evaluated against the source system at query time, means the assistant cannot surface what the person could not already open. It is tested by asking the same question from two accounts and requiring different answers.
Do we need our own model?+
Rarely, and the question is usually about data handling rather than about the model. What matters is where inference runs, what is retained, and what contractual terms apply — all of which can be satisfied with managed services under the right agreement or with self-hosted models where the requirement genuinely demands it. Deciding that first prevents an expensive answer to a question nobody asked.
How long until it is useful?+
Six to ten weeks for one department, most of which is corpus curation and the permission work rather than the assistant itself. Starting with one department is not caution for its own sake: it is how the curation problem stays small enough to be finished, and an assistant that answers HR questions well is worth more than one that answers everything approximately.
In this cluster
Next step
Run it against your own corpus
An assistant running over one department's corpus with permissions enforced and answers cited, before it is opened to the organisation.

