The right answer, its source, and nothing you are not cleared to read.
Circulars, guidance notes, SOPs and case notes sit in SharePoint, Confluence, shared drives and line-of-business systems. We build search that keeps each source’s access rights, answers with references to the passages used, or tells the user honestly that it has no answer.
For Irish public bodies and companies whose guidance is scattered across systems, often in two languages, and whose access rules must hold.
Sources that owners have approved are connected along with their access rights and indexed with metadata. A member of staff asks a question, and the search only covers what that person is entitled to read. If the evidence is thin or contradictory, the question is passed to the owner of the content; if not, the reply references its sources. Feedback is used to improve both the sources and the test set.
Before and after
How the day changes for your staff
Today
With the system
Today
Today: A search that finds too much
A keyword query brings back hundreds of hits, drafts and duplicates among them, plus circulars that were replaced long ago.
With the system
With the system: Brief answers, every claim referenced
The reply quotes the passages behind it, links to each one and shows the date each source was last reviewed.
Today
Today: Know-how held by a handful of people
The quickest way to get an answer is to ask whoever has been in the unit longest. That does not scale, and the knowledge goes when they retire.
With the system
With the system: Access rights stay attached
At the moment of asking, results are filtered on the user’s own rights, so one shared index never exposes material to the wrong reader.
Today
Today: No proof behind the answer
Staff pass on what they recall, and no one can say which version of a circular or procedure it came from.
With the system
With the system: No answer is better than a guess
If the sources do not back an answer, the assistant says so and sends the question to whoever owns that subject.
Today
Today: Two languages, one question
The guidance was published in Irish and English, but only one version was updated, and the person asking found the other.
With the system
With the system: Irish and English together
A question in English can find guidance written in Irish and the other way round, and the reply comes back in the language used to ask.
Workflow
The workflow, step by step
How good the answers are depends more on the sources than on the model. So the workflow opens with owners choosing what goes in, and closes with feedback flowing back to those same owners.
Illustrative workflow
7 steps · 2 human checkpoints
Legend
System
Person
Human checkpoint
Exception path
Enterprise knowledge: illustrative workflow and hold-back route
Read the diagram as text
The usual route starts by linking approved sources and indexing them with labels, then a member of staff asks, the search is limited to their rights and a referenced reply comes back, followed by feedback and testing.
The first point where a person steps in comes before indexing: an owner approves every source.
If the search finds too little support or sources that disagree, the assistant gives no answer. The question is taken off the usual route and passed to the content owner, who replies and decides whether a source should change. From there it returns to the route at feedback and testing.
01
Step 1: Link the approved sourcesSystem
Connectors read SharePoint sites, Confluence spaces, shared drives, helpdesk tools and document stores in the ERP, and keep each item’s access-control list in step with its content.
Human checkpoint
Every source and its scope is approved by the owner of the content before indexing starts, and a source that is withdrawn leaves the index.
02
Step 2: Index with labelsSystem
Each document is cut along its headings and sections and tagged with owner, language, status (draft, approved, replaced) and review date. One hybrid index serves keyword and vector search.
03
Step 3: Ask where you already workPerson
Staff put their question in the intranet, Microsoft Teams, the helpdesk console or a business application, and their identity goes with it.
04
Step 4: Search within each user’s rightsSystem
The question is widened to cover both languages, results are filtered on the user’s rights and ranked again. Replaced documents are left out, and sources that disagree are flagged.
05
Step 5: Pass to the content ownerPerson
A question with too little support, or with sources that contradict one another, goes to the owner of that subject together with the passages found.
Exception path: Too little or contradictory support· Returns to step 7
Human checkpoint
The assistant holds back instead of guessing. A person replies and decides if a source needs to be written, revised or withdrawn.
06
Step 6: Reply with referencesSystem
The system writes a brief reply using only the passages found. Every 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. Before any change to search or prompts, we measure again whether answers stay faithful to the sources and cite them correctly.
Components
What we would put in place
01
Connectors that sync access rights
Incremental connectors that bring access-control lists, deletions and version status from every source system into the index.
Chunking that follows document structure, embeddings that work in Irish and English, and keyword search in a single index, filterable by status, language and age.
Rewriting of the question, rights filtering when the question is asked, re-ranking, and replies grounded in the passages, with reference checks and the option to hold back.
Test sets for each subject area, scoring of faithfulness and references, and reports showing owners which questions went unanswered and which sources are out of date.
Deployment in your own tenancy in an EU region, with private network links to the sources and model endpoints, and your content never used for training.
Directory and identity provider, such as Microsoft Entra ID
Knowledge and console in Salesforce Service CloudPlatformSalesforce
Document stores and business objects in SAPPlatformSAP
The intranet and Microsoft Teams
IT service management platform
Controls
Safeguards built in
Access boundaries
We never merge access rights into a single shared index. Each item keeps its own access-control list, and each question is filtered on the user’s identity and group memberships at the moment it is asked.
Sensitive areas such as HR, legal and IT security can be left out altogether, or indexed for their own readers only.
Human review
Owners choose what is indexed and get alerts about unanswered questions and sources that disagree. In subject areas with a high impact, replies can be restricted to direct quotes.
Auditability
For each reply we log the question, the passages found, the sources referenced, the model and prompt version and the user’s rating, kept for a period you decide.
Data protection
Personal data inside the sources remains under the access rules of those sources. Logs are pseudonymised where we can, and none of your content trains outside models.
AI transparency
The screen tells users that replies are produced by an AI system, shows the sources behind each one and offers a single click to flag a wrong answer, in line with the transparency duties of the AI Act.
Measures
How we would judge it
We settle these measures with you in discovery, built on real questions gathered from the teams who will use the assistant.
How we would judge it
Metric
Why it matters
How we would measure it
01Faithfulness to sources
Why it mattersA reply the sources do not support does more harm than no reply.
How we would measure itProportion of statements in a reply that the referenced passages back up, scored on a test set and on a sample checked by staff.
02Correct references
Why it mattersStaff need to check any reply with a single click.
How we would measure itProportion of references that lead to the passage which really supports the statement.
03Questions resolved
Why it mattersTells you if staff get what they came for, not merely whether text appeared.
How we would measure itRatings, follow-up questions and hand-offs to owners by subject area, from usage logs and feedback.
04Time to answer a set task
Why it mattersTies the assistant to actual work, for example a helpdesk answering a question on a circular.
How we would measure itTimed runs of the same tasks before and after, on a fixed list of typical questions.
No targets are set before a baseline exists.
Rollout
How we would introduce it
Phase 01
Discovery and review of sources
A single subject area with named owners, for example HR circulars or support for one product. We gather real questions and assess the condition of the sources.
Exit criteria
Sources and owners confirmed
Real questions gathered
Access model checked with IT security
Phase 02
Pilot with one team
One team uses the assistant in one place, on approved sources and nothing else.
Exit criteria
Faithfulness and reference targets reached
Access tests show no leaks
Feedback to owners working
Phase 03
Add subject areas
Further sources and user groups, each brought in after the same check of owners, rights and test questions.
Exit criteria
A test set for every subject area
Content lifecycle settled with owners
Phase 04
Run
Your team takes over monitoring, test runs and content reports, and one named person owns the service.
Circulars, guidance notes and earlier decisions for public servants in Government departments, State agencies and local authorities, in Irish and English.
Underwriting guidelines at insurers, product terms and compliance procedures for advisers and operations teams at banks and fund administrators.
Illustrative example
Circulars and guidance for staff in a Government department
Situation
A Government department issues circulars, HR guidance and procedures in Irish and English across several divisions. Staff send the same questions to HR and the corporate services unit week after week, and older circulars are often quoted after they were replaced.
System
An assistant on the intranet answers from current circulars and the guidance each user is cleared to read, in the language of the question, quoting the paragraph it relies on.
Human control
Each division owns its guidance and approves it before indexing. Anything the sources do not cover goes to HR or the corporate services unit, and the assistant does not answer it.
What we would measure
Faithfulness and correct references on a set of questions in both languages, and the number of questions passed to HR.
Ireland
Guidance in two languages, rights intact
Many public bodies in Ireland publish in Irish and English, and the Official Languages Act places Irish-language duties on them. Irish operations of international groups add policies written abroad. We design search that keeps each system’s access rights and finds the right passage whichever language holds it.
Hosting region, logs and retention are agreed early with your data protection officer under the GDPR and the Data Protection Act 2018. Where an assistant supports staff in their work, we design for the transparency duties of the AI Act, which the Regulation of Artificial Intelligence Act 2026 gives effect to in Ireland, with the AI Office of Ireland as the coordinating authority.
01How do you stop people seeing content they should not?
We copy the access-control lists from every source and apply them each time a question is asked, based on who is asking. Then we prove it: ahead of any pilot, we log in with accounts from different groups and confirm that no content crosses over.
02And if the assistant gets it wrong?
It will now and then, which is why each reply shows its sources and anyone can flag a problem with one click. Flags go to the content owner and turn into test cases. When the support is weak, the assistant is built to hold back rather than guess.
03Our office suite already has an assistant. Why build one?
Often there is no need. If your content sits in one suite and the assistant that comes with it meets your quality bar, we help you set it up and govern it. A tailored service is worth it when sources are spread over several systems, when you want control over search and testing, or when data residency rules out the standard option.
04Our content is untidy. Must we clean it all up first?
No, although content governance is usually what holds things back. We start with one subject area, flag replaced and draft documents, and send owners reports on duplicates and out-of-date sources, so quality improves first where it matters most.
05Can it work across Irish and English guidance?
Yes. Embeddings that handle both languages, and questions widened across them, find sources in either one. The reply comes in the language of the question and references the original passages. Where a public body publishes in Irish under the Official Languages Act, both versions can be indexed and checked against each other.
06How long would a pilot run?
Mostly it depends on how quickly we get access to sources and owners. Pilot scope and exit criteria are fixed in discovery. See how we work and our AI engineering work.
Show us where your knowledge sits
Tell us which systems hold your guidance and who relies on it. We will review sources, access rights and languages with you and suggest a first subject area.