Get the answer, check its source, and see nothing you are not cleared for.
Policies, procedures, service tickets and technical files sit in SharePoint, Confluence, shared drives and business systems. We build search that keeps the access rights of each source intact and replies with citations, or states clearly that the sources do not answer the question.
For organisations in Malaysia whose knowledge is spread over many systems and across Malay and English sources, and whose access rights must hold.
Sources approved by their owners are connected with their access rights and indexed with metadata. When someone asks a question, retrieval returns only what that person is allowed to see. If the evidence is thin or contradictory, the question goes to the topic owner; if not, the reply cites its sources. Feedback flows back into the sources and the evaluation set.
Before and after
How the working day changes
Today
With the system
Today
Today: Search that finds too much
A keyword search returns hundreds of files, drafts and duplicates among them, along with versions withdrawn long ago.
With the system
With the system: Brief answers with sources
Every answer quotes the passages behind it, links to them and shows the date each source was last reviewed.
Today
Today: Answers depend on a few colleagues
The quickest way to an answer is to ask whoever has been there longest, which does not scale and is lost when that person leaves.
With the system
With the system: Access rights stay attached
Results are filtered by the access rights of the person asking at the moment of the query, so one index never becomes a way around permissions.
Today
Today: No evidence behind the answer
Colleagues pass on what they recall, and nobody can tell which version of a policy it came from.
With the system
With the system: Silence when evidence is thin
If the sources do not support an answer, the assistant says so and passes the question to whoever owns the topic.
Today
Today: Two languages, one question
The procedure is an English document, and the question arrived as a Malay message.
With the system
With the system: Search across Malay and English
A question typed in one language finds sources in the other, and the answer comes back in the language of the question.
Workflow
How the flow runs
Answer quality depends more on the sources than on the model. So the flow begins with owners choosing what goes into the index and ends with feedback that returns to them.
Illustrative workflow
7 steps · 2 human checkpoints
Legend
System
Person
Human checkpoint
Exception path
Enterprise knowledge: illustrative flow with a referral path
Read the diagram as text
On the main path, approved sources are linked and indexed with tags, a person asks a question, retrieval stays within that person’s access rights and a reply with citations follows; the path ends with feedback and evaluation.
The first human checkpoint comes before indexing: content owners approve every source.
If retrieval finds too little evidence or sources that contradict each other, the assistant gives no answer. The question leaves the main path and goes to the topic owner, who replies and decides whether a source needs to change. The referral then joins the main path again at feedback and evaluation.
01
Step 1: Link approved sourcesSystem
Connectors read SharePoint sites, Confluence spaces, shared drives, ticketing systems and ERP document stores, and keep the access-control list of each item in step with its content.
Human checkpoint
The content owner approves each source and its scope before indexing starts; a source that is withdrawn leaves the index.
02
Step 2: Index and tagSystem
Documents are divided along their headings and sections and tagged with owner, language, status (draft, approved or superseded) and review date. One hybrid index serves keyword and vector search together.
03
Step 3: Ask a questionPerson
Staff ask from the tools they already open each day: the intranet, Microsoft Teams, the service console or a business application. Their identity is carried with every question.
04
Step 4: Retrieve within access rightsSystem
The question is widened across Malay and English, results are filtered by what the person asking may open, and then ranked again. Superseded documents are left out, and sources that contradict each other are flagged.
05
Step 5: Pass to the topic ownerPerson
Questions that the evidence cannot support, or where sources disagree, are sent to the owner of the topic together with the passages that were found.
Exception path: Evidence thin or contradictory· Returns to step 7
Human checkpoint
The assistant declines to guess. A person replies and decides whether a source should be written, revised or withdrawn.
06
Step 6: Reply with citationsSystem
A brief reply is written only from the passages retrieved. Each statement links to its source, and the owner and review date of that source are shown.
07
Step 7: Learn from feedbackSystem
Ratings, corrections and passed-on questions are turned into test cases. Groundedness and citation accuracy are measured again before retrieval or prompts are changed.
Components
Components we would deliver
01
Source connectors with access sync
Connectors that pick up changes and bring permissions, removed items and version status from every source system into the index.
Query rewriting, filtering by access rights when the question is asked, re-ranking, and grounded generation with citation checks and the option to decline.
Test sets for each domain, scoring of groundedness and citations, and reports that show owners which questions went unanswered and which sources are out of date.
Deployment in your tenancy at a hosting location agreed with you and recorded in the contract, with private network links to sources and model endpoints, and no training on your content.
Directory and identity provider (Microsoft Entra ID or an equivalent)
Salesforce Service Cloud knowledge base and consolePlatformSalesforce
Document stores and business objects in SAPPlatformSAP
The intranet and Microsoft Teams
Service management platform
Controls
Controls built in
Access boundaries
We never merge permissions into one shared index. Every item keeps its own access-control list, and each query is filtered by the identity and group memberships of the person asking, at the time of asking.
Areas with sensitive material, such as HR, legal and security, can be left out altogether or indexed for their own users only.
Human review
Content owners choose what is indexed and receive notice of unanswered questions and conflicting sources. In domains where a wrong answer carries high impact, replies can be restricted to exact quotations.
Auditability
Every reply is logged with the question, the passages retrieved, the sources cited, the model and prompt version and the user’s rating, for a retention period you decide. If a personal data breach has to be notified to the Personal Data Protection Commissioner, these logs help you establish quickly what was exposed.
Data protection
Personal data in source documents stays under the access rules of its source, in line with the Security Principle of the Personal Data Protection Act 2010 (PDPA). Logs are pseudonymised where possible, documents serve only the purpose they were indexed for, and nothing is used to train external models.
AI transparency
The interface makes clear that answers come from an AI system, shows the sources behind each one and offers a simple way to report a wrong answer.
Measures
Measures we would agree
We agree the measures with you during discovery, using real questions gathered from the teams who will rely on the assistant.
Measures we would agree
Metric
Why it matters
How we would measure it
01Groundedness
Why it mattersAn answer its sources do not support does more harm than no answer at all.
How we would measure itThe share of statements in an answer that the cited passages support, scored on a test set and through sampled review by people.
02Citation accuracy
Why it mattersStaff need to be able to check an answer with a single click.
How we would measure itThe share of citations that lead to the passage which actually supports the statement.
03Search success
Why it mattersShows whether people find what they came for, not only whether they get some text back.
How we would measure itRatings, follow-up questions and referrals per domain, taken from usage logs and feedback.
04Time to answer on set tasks
Why it mattersTies the assistant to real work, such as a service desk officer answering a question about a policy.
How we would measure itTimed comparisons before and after introduction, on a fixed set of typical questions.
No targets are set before a baseline exists.
Rollout
Rollout in stages
Phase 01
Discovery and source check
One domain with clear owners, such as HR policy or product support. We gather real questions and assess the condition of the sources.
Exit criteria
Sources and owners confirmed
Question set gathered
Access model checked with IT security
Phase 02
Pilot with one team
The assistant serves a single team, in a single interface, on approved sources only.
Exit criteria
Groundedness and citation levels reached
Access tests show no leaks
Owners using the feedback loop
Phase 03
More domains
Further sources and user groups, each added after the same check of owners, access rights and test questions.
Exit criteria
Test set for each domain
Owners agree how content is maintained
Phase 04
Day-to-day operation
Monitoring, evaluation runs and content reports pass to your team, with one named owner for the service.
Policy documents from Bank Negara Malaysia (BNM), guidelines from the Securities Commission Malaysia (SC), product terms and internal procedures for compliance and operations staff.
Illustrative example
Product compliance files at an electronics exporter
Situation
An electronics exporter with a sales company in the EU could keep test reports, declarations of conformity and customer requirements in several systems. Engineers and sales staff would ask the compliance team the same questions about each product again and again.
System
An assistant in the intranet would answer from the approved compliance files for the product in question, citing the report or clause it relies on, and only from files the person asking may open.
Human control
The compliance team owns the files and approves them for indexing. Questions the files do not answer go to a compliance engineer, who replies and decides whether a file must be updated.
What we would measure
Groundedness and citation accuracy on a question set per product family, and the number of questions passed to the compliance team.
Malaysia
Regulator texts and the PDPA
Compliance teams at banks, insurers and takaful operators work from the revised Risk Management in Technology (RMiT) policy document of Bank Negara Malaysia, in effect since 28 November 2025. Capital market intermediaries follow the Guidelines on Technology Risk Management of the Securities Commission Malaysia, in effect since 19 August 2024. We build search that cites the paragraph an answer rests on, so staff can check it against the regulator’s own text.
Internal documents also hold personal data. Under the Personal Data Protection Act 2010 (PDPA), the Security Principle and the purpose for which data was collected limit who may retrieve it and why, so access rights travel with every passage. For staff of group companies in the EU, the EU General Data Protection Regulation (GDPR) and the transparency duties of the EU AI Act apply to the same assistant, and the hosting location is agreed with you and recorded in the contract.
01How do you stop people seeing what they may not open?
Access-control lists are copied from every source and applied when the question is asked, based on the identity of the person asking. We also test it: before a pilot begins, we run access tests with accounts from different groups and confirm that nothing crosses between them.
02What if the assistant gets it wrong?
At times it will, which is why each answer shows its sources and users can flag a problem with one click. Flags go to the content owner and become test cases. Where evidence is thin, the assistant is built to decline rather than guess.
03We already use a workplace assistant. Why build another?
Often there is no need. If your content sits in one suite and its built-in assistant meets your quality criteria, we help you set it up and govern it. A custom service makes sense when sources span several systems, when you need control over retrieval and evaluation, or when the default hosting location does not meet your requirements.
04Our documents are untidy. Must we clean them all first?
No, though content governance is usually the real constraint. We begin with one domain, mark drafts and superseded documents, and give owners reports on duplicates and outdated sources, so quality improves first where it matters most.
05Can it answer from both Malay and English sources?
Yes. Multilingual embeddings and query expansion find sources written in the other language, and the reply is given in the language of the question, with citations to the original passages.