Context

A large real estate conglomerate with more than 5,000 employees across 15+ subsidiaries — spanning residential development, commercial leasing, building management, brokerage, and construction — was struggling with a persistent operational issue: employees were spending too much time searching for internal policies, HR procedures, compliance guidelines, and operational SOPs scattered across intranet portals, shared drives, archived emails, and printed handbooks. HR and Administration were receiving an estimated 500+ repetitive questions per day on topics such as leave policy, expense reimbursement, onboarding, commission structures, and subsidiary-specific rules. The group wanted to launch an internal AI chatbot, accessible through the intranet and internal messaging platforms, that could answer employee questions instantly and accurately while reflecting the right policy version based on role, entity, and context.

The Challenge

• The client's knowledge was highly fragmented across multiple systems, formats, and business units, making it difficult to turn raw documents into a unified and searchable knowledge base.
• Many policies were context-dependent, meaning the same question could have different answers depending on tenure, seniority, subsidiary, department, or location.
• Historical employee questions came from multiple channels and varied widely in language, tone, abbreviations, and specificity, which increased the complexity of intent design and paraphrase coverage.
• There were frequent policy overlaps and contradictions between group-wide rules and subsidiary-level exceptions, creating a real risk of incorrect chatbot answers if conflicts were not explicitly labeled.
• The knowledge base required continuous post-launch maintenance because internal policies, legal requirements, and operating procedures changed regularly.
• Sensitive topics such as disciplinary action, salary disputes, whistleblowing, and legal compliance required careful escalation logic rather than full automation.

Our Approach

• We worked cross-functionally with HR, Legal, Compliance, Finance, and Operations teams to collect approximately 2,500 internal documents across Word, PDF, PowerPoint, scanned files, and internal wiki formats.
• Our annotation team transformed these raw materials into a structured knowledge base by extracting key policy content, assigning metadata such as department, applicable subsidiary, effective date, version, and approval level, and breaking long documents into standalone knowledge units suitable for question-answering.
• We analyzed 50,000 historical employee inquiries from email, Teams, and helpdesk systems, and labeled them into 60 main intents and 200+ sub-intents, reflecting the operational complexity of a large multi-entity organization.
• We annotated domain-specific entities such as business unit, subsidiary, job grade, role title, project name, policy name, clause reference, timeline, monetary threshold, and approval level so the chatbot could return contextually correct answers.
• We created 30,000+ context-aware Q&A pairs, mapping policy logic across multiple conditions such as tenure bands, management levels, and subsidiary-specific exceptions.
• We designed and labeled approximately 5,000 multi-turn conversation flows for more complex employee journeys, such as inter-company transfers, onboarding escalation, equipment requests, and compliance-related inquiries.
• To improve robustness, we generated 40–60 paraphrases per major intent, covering formal language, colloquial phrasing, departmental jargon, and common abbreviations used by different employee groups.
• We explicitly labeled policy conflicts and edge cases so the chatbot could respond safely when group policy and subsidiary policy differed, instead of forcing a single answer where ambiguity existed.
• After go-live, we supported a human-in-the-loop BPO layer that handled low-confidence or high-sensitivity escalations, updated policy content within 24–48 hours of any change, monitored unanswered questions daily, and generated monthly management reports on usage trends and knowledge gaps.

Result

• After 4 months of annotation and 2 months of UAT with 200 pilot users across 5 subsidiaries, the chatbot successfully went live on Microsoft Teams and the group intranet.
• In the first quarter after launch, the assistant was able to resolve 85% of internal policy-related questions without human intervention.
• Average response time dropped from 4–24 hours to under 10 seconds for standard employee inquiries.
• HR and Administration reported a 60% reduction in repetitive questions, freeing up approximately 120 staff hours per week for higher-value work.
• Employee satisfaction with access to internal information improved from 2.8 to 4.3 out of 5.
• The ongoing monitoring process identified enough knowledge gaps to drive 45 policy clarifications and 12 new policy documents within the first 6 months.
• The client expanded the assistant's scope in phase two to cover IT helpdesk and facilities-related requests, supported by the same annotation and maintenance workflow.