We built a multilingual AI assistant that answers employees' HR policy questions, checks global policies against local law and keeps personal data out of the logs. It resolves 85% of queries and has saved 12k hours.
Client: Nova MissionIndustry: HR TechService: AI & Automation
Overview
Client
Nova Mission
HR team supporting about 20,000 employees across 14 countries
Engagement snapshot
The challenge
The client's HR knowledge sat in 1,200+ static PDF policies, local law addendums and scattered Slack threads. For 20,000 employees, a simple benefits question often needed someone in HR to step in, which built up 48-hour ticket backlogs and pulled managers into answering questions on the side.
Employees searched long PDFs by keyword or asked colleagues, and those answers were often wrong for their country. When local laws changed, the policy documents were slow to catch up.
The volume of repeat questions drove high turnover in the HR support team and delayed wider people projects.
Unverified advice from colleagues on local rules created legal risk in tightly regulated labour markets.
Older chatbots did not reliably mask personal data, so sensitive employee details could leak into training or answers.
The solution
We built a retrieval-augmented generation (RAG) workflow that splits policies into sections by meaning, so context is kept when answers are given in another language.
AI agents check global master policies against local labour law addendums, so each employee gets an answer that is correct for their country.
Each conversation is scrubbed of personal information before it is stored. The logs are used for quality checks and to spot policies that have drifted out of date.
Key features
A retrieval-augmented assistant searches HR policies by meaning and answers employee questions in 8 languages.
Business impact
85% of questions resolved automatically
AI agents reconcile global policies with local labour law addendums for country-correct answers.
Business impact
65% lower HR support load
Each answer cites the exact handbook paragraph, and personal data is masked before reaching the model.
Business impact
2.4s p95 answers
The assistant compares each draft answer with the source policy before it replies.
Business impact
Zero hallucinated answers
Each answer links to the exact paragraph in the official handbook, so employees can check it and HR has an audit trail.
Business impact
100% policy accuracy
The system spots surges in similar questions, for example at tax time, and suggests policy updates to HR leaders.
Business impact
65% fewer tickets
Before and after
Before
Keyword search
Employees scrolled through long PDFs to find the clause they needed.
After
Plain questions
Employees ask in their own words and get a direct answer with a link to the source policy.
Before
Manual updates
Policies often fell behind local law changes in different regions.
After
Automatic sync
Policy changes flow into the vector store automatically, so the AI always works from the latest version.
Before
24-48 Hours
Tickets queued for manual sorting and waited on staff in other time zones.
After
Instant (<3s)
Questions are answered automatically, 24/7, in every region, and this held as the business grew.
Results
85% Resolution Rate
The share of employee questions the AI answers fully, without passing them to someone in HR.
Before: ManualAfter: 85% Auto
65% Less Support Load
Far fewer manual tickets means HR business partners can spend their time on higher-value work.
Before: CongestedAfter: 65% Lower
2.4s p95 Response Time
Answers stream back quickly even during busy benefits enrolment periods.
Before: < 5.0sAfter: 2.4s
Testimonial
“About 20,000 employees in 14 countries ask us policy questions in 8 languages. Coretus built an assistant that now resolves 85% of them automatically, citing the exact handbook paragraph. It has saved 12k HR hours.”
Olivia RogersHead of HR Operations, Nova MissionKey results
Trust and control
Personal data is removed from every prompt in real time, so no employee identity data reaches the LLM providers.
PrivacyThe system logs the exact policy version and passage used for each answer, so legal teams can verify it.
TraceableCloud-based inference scales to 20,000+ staff using the assistant at the same time.
Scales with demandThe client owns 100% of the IP in the knowledge platform, vector embeddings and custom agent logic.
100% ownedFaster delivery
About 10 weeks of setup was saved by starting from these components. That time went on local law rules and answer checking.
Loaded the 1,200+ policy PDFs and addendums into the vector store and kept them in sync.
Removed employee identity data from every prompt before it reached the LLM provider.
Tracked answer quality, response times and escalations to HR, so weak answers could be fixed quickly.
Let one set of policies answer questions in all 8 languages without keeping separate copies.
Its 20,000 employees in 14 countries had to dig through 1,200+ PDF policies, local law addendums and old Slack threads. Most benefits and tax questions ended up with HR, creating 48-hour ticket backlogs. Answers from colleagues were often wrong for local law, which created legal risk.
It uses retrieval-augmented generation (RAG), so the assistant finds the relevant policy passages and then writes an answer from them. It checks global master policies against local labour law addendums, compares its draft with the source before replying, and links to the exact handbook paragraph it used.
Personal data is removed from every prompt before it is sent, so the LLM providers never receive employee identity data. Conversation logs are also scrubbed of personal information. Each answer is logged with the policy version and passage it came from, which lets legal teams verify it.
If your HR team spends its days answering the same policy questions, we can help you build an assistant that answers from your own handbook and cites its sources.