From 15 hours a day on HR questions to 2: a case for AI HR Senior

We did not begin by uploading every HR file we could find. We first sat with recent support cases and watched how operators actually answered them. A question might start in the work rules, move to a benefits guide, then jump to a groupware notice, an expense policy, or a spreadsheet for foreign-worker guidance. That path mattered. The knowledge library had to follow how the team really worked, not how the folders were named.

The cleanup work was less glamorous than the agent, but it decided whether the agent would be useful. PDFs, Word files, Notion pages, web documents, and spreadsheets could not go in as-is. We removed repeated headers, split long documents into question-sized chunks, and tagged them with the words employees actually used: travel expenses, shift work, foreign-worker onboarding, performance bonus, groupware expense handling. That made retrieval work even when the question did not use the exact wording in the policy.

Once the library was usable, we connected AI HR Senior. The rollout still needed people work. HR operators, site managers, and affiliate contacts had different habits, so we agreed on operating rules together. Which questions can the agent draft directly? Which ones need admin review? Which ones reveal that the source document itself is unclear? OTOworks stayed close to real support cases while those rules settled. That is why AI HR Senior feels less like a bot and more like a senior teammate who pulls the right evidence before you answer.

Take a question like, 'How does annual leave carryover work for field employees?' AI HR Senior checks whether the office and field rules differ, whether statutory rules or internal policy apply, and whether an affiliate exception matters. Then it pulls the relevant policy chunks and FAQs into a draft with source links. The operator is no longer staring at a blank page. They are reviewing evidence and deciding exceptions. The same flow runs whether the question came in by email or chat.

The existing material was mostly files, and every operator explained it a little differently. We did not try to solve the whole area at once. We first mapped the basics: employee type, site, procedure, required documents, training items, and owner. That lightweight ontology gave the library enough structure to serve different guidance for new hires, factory workers, foreign workers, and affiliates. Pre-start instructions, factory safety training, residence-document guidance, and groupware onboarding could finally be found by context instead of folder location.

The client already had performance criteria and tools. The hard part came when ratings affected pay or bonuses. Office roles leaned on ratings and goal achievement. Factory roles added shifts, overtime, night work, and production-line rules. Foreign-worker cases added visa, contract, and allowance context. A black-box calculator would have been the wrong answer. We organized the rules, evidence, and review questions so AI HR Senior could assemble what an operator needed to check.

We kept this part deliberately conservative. Compensation questions should not be finalized by an AI system alone. AI HR Senior summarizes the performance rating, worker type, work pattern, allowance rule, and source policy. Sensitive or unusual cases stay in the admin-review queue. The team gets a faster review packet, but the final call remains with a person. That made adoption easier, especially for payroll and performance work.

The last piece was logging. Every question, source document, answer draft, and admin edit is kept. Those logs are not just audit history. They show what to fix next. Repeated questions become FAQ candidates. Questions with weak evidence become document-backlog items. Affiliate-specific exceptions get their own tags. The more the system is used, the cleaner the library gets, and the more stable the answers become.