Onboarding docs and a policy Q&A bot
Turn your scattered onboarding and policy documents into a Q&A assistant that answers new-hire and employee questions with quoted sources, and escalates anything sensitive to a human.
Tools you'll use
A policy Q&A bot is an AI assistant pointed at your real onboarding and policy documents that answers new-hire and employee questions in plain English, quotes the source so the person can verify it, and escalates anything sensitive or unclear to a named human.
The problem it solves is the relay job. Every HR team answers the same questions over and over: how much leave do I have, when does the new policy start, who approves expenses, where is the laptop form. The answers already exist, but they are spread across a handbook PDF, a benefits portal, three Google Docs, and the memory of one person who is on holiday. That scatter is the usual failure point: a Gartner survey of 5,728 customers (December 2023) found self-service resolves only 14% of issues fully, and the most common reason is people simply cannot find the relevant content. New hires feel lost, and your team spends its week relaying facts instead of doing real people work.
This matters because onboarding is where most teams are weakest. Only 12% of employees strongly agree their organization does a great job onboarding new staff (Gallup). A good bot attacks the highest-volume, lowest-judgment part of your inbox, gives new hires answers in seconds instead of a day, and makes your written policies the single source of truth, because the bot is only as good as the documents behind it, which finally gives you a reason to clean them up.
Moriva's take
This clears Gate 1 easily: answering policy and onboarding questions is real work your team does every single week. Gate 2 holds if you build it as files plus a small, owned automation rather than a black-box vendor product, so your people can update a document and see the answer change the same day. The reason this is CAREFUL not GO is Gate 3-adjacent: HR data is sensitive and a wrong answer about pay, leave, or a complaint carries real consequences, so this needs quoted sources, a tight scope, and a hard human-handoff rule before it goes wide.
How do you onboarding docs and a policy Q&A bot?
- 1
Gather and clean the source documents first
Pull every onboarding and policy document into one folder: handbook, leave policy, benefits summary, expense rules, IT setup, code of conduct, the new-hire checklist. Use Claude Cowork to read across all of them and flag contradictions, out-of-date dates, and gaps before you build anything. A bot trained on stale or conflicting documents will confidently give wrong answers, so this cleanup is the work, not a preamble to it.
- 2
Decide the scope and the no-go list
Write down exactly what the bot may answer (general policy, process, where-to-find) and what it must never answer (anything about a specific person's pay, performance, a live complaint, harassment, discrimination, accommodation, or mental-health situations). This list becomes a literal instruction in the build. Keeping scope narrow is what keeps risk at medium rather than high.
- 3
Build the assistant over your files with Claude Code or Codex
Point Claude Code (or Codex) at the document folder and describe the job in plain English: answer employee questions using only these files, quote the exact passage you used, and if the answer is not clearly present, say so and hand off. Start with agentic search over the plain files rather than a vector database; it is simpler, more transparent, and easy for your team to debug. You end up owning the script and the prompt, so you can change behavior without calling anyone.
- 4
Force grounded answers with citations
Configure the assistant to always return the source document and quoted line behind every answer, and to refuse rather than improvise when the documents do not cover the question. This single rule is the difference between a useful tool and a liability. Spend your build time here: a bot that says 'I don't have that in the policy, here is who to ask' is far safer than one that guesses.
- 5
Wire in the human handoff for sensitive topics
Add explicit triggers so the bot escalates instead of answering: any item on your no-go list, any request for a person, and any case where it would have to guess. The handoff should name the right human and pass along the question so the employee feels progress was made. Treat this as a hard rule in the prompt, tested deliberately, not a nice-to-have.
- 6
Run a red-team pass before anyone real uses it
Use Claude Cowork to generate a few hundred realistic questions, including the awkward ones (How do I report my manager? Can I see someone's salary? What if I'm pregnant?), and check every answer for accuracy, correct citation, and correct escalation. Fix the documents and the prompt until the failures are gone. Owning the test set means you can re-run it every time you change a policy.
- 7
Pilot with one cohort, then connect to where people already are
Launch to the next onboarding group or one department first, with a feedback button on every answer. Once it holds up, connect it to Slack or your help channel via MCP so employees ask in the tool they already use. Keep the source documents as the only knowledge source so updates stay a one-place job.
- 8
Measure, then keep the documents fresh
Track question volume, deflection rate (answered without a human), escalations, and any flagged-wrong answers. Review the wrong-answer log monthly and fix the underlying document, not just the bot. That review loop is what makes this a durable Gate-3 win you can point to: fewer repeat questions in the inbox and faster new-hire ramp.
What could go wrong (and how to handle it)
The bot states a confident but wrong answer about pay, leave, or benefits, and an employee acts on it.
Require a quoted source on every answer and a hard refuse-and-handoff when the documents do not clearly cover the question. Run a red-team test set before launch and after every policy change.
Sensitive topics (harassment, discrimination, accommodation, mental health, a specific complaint) get an automated answer instead of a human.
Maintain an explicit no-go list and route those topics straight to a named HR person. Test these triggers deliberately; treat any miss as a launch blocker.
The assistant reaches confidential files it should not, exposing one person's data to another.
Scope the bot to a folder of general, non-personal policy documents only. Never connect it to individual employee records, payroll, or case files. Keep personal data out of the knowledge source entirely.
Documents go stale and the bot keeps repeating an outdated rule.
Make the source folder the single source of truth and assign an owner. Re-run the test set whenever a policy changes, and review the wrong-answer log monthly.
Over-automation: people stop talking to HR for things that genuinely need a human.
Keep scope to high-volume, low-judgment questions and make the handoff prominent and friendly. The goal is to clear the relay work, not to replace the relationship.
Compliance exposure under data-protection rules (for example GDPR) if personal data is ingested or answers are logged carelessly.
Exclude personal data from the knowledge source, review what gets logged, and have your legal or compliance owner sign off on scope and retention before launch.
Prompts to get started
FAQ
Won't it just make things up like every other AI chatbot?
Only if you let it. The whole build centers on grounding: it answers from your documents, quotes the line it used, and refuses when the answer isn't there. A bot that says 'that's not in our policy, here's who to ask' is the goal, and that behavior is something you set and test, not hope for.
Is it safe to point AI at HR data?
Point it at general policy and onboarding documents, not at personal records. The bot should never touch payroll, performance files, or case notes. Scoped that way, the sensitivity is medium, not high, because no individual's data is in the knowledge source.
What happens when someone asks about a harassment complaint or their own situation?
It hands off. Those topics sit on an explicit no-go list and route straight to a named HR person, with the question passed along so nothing is lost. This handoff is built and tested as a hard rule, not left to the model's discretion.
Do we need engineers to keep this running?
No, that's the point. Because it's built from your documents plus a small automation your team owns, updating an answer usually means editing a document. When you do want to change the bot's behavior, you have the prompt and the script and can adjust them, or re-run the same agentic tool to do it for you.
How do we know it's actually saving time?
Track repeat-question volume in the HR inbox, how many questions get answered without a human, escalations, and any flagged-wrong answers. The recurring win is fewer repeated policy questions and faster new-hire ramp, both of which you can measure against your starting baseline.
Sources
- Only 12% of employees strongly agree that their organization does a great job onboarding new employees. — Gallup
- Only 14% of customer service and support issues are fully resolved in self-service (survey of 5,728 customers, December 2023); the most common point of failure is the inability to find content relevant to the issue. — Gartner, 2024
More from HR & People
Job descriptions and structured interview kits
Turn a role's real requirements into a clean job description and a matching structured interview kit (questions plus a scoring rubric) that your team owns, reuses, and runs the same way for every candidate.
Resume-screening assist and structured scorecards
An AI assistant that reads each resume against a job-specific rubric and returns a structured scorecard with evidence, so your recruiters spend their time on real candidates instead of the first pass. A human still makes every decision.
Prep renewals and spot upsell openings
Build a repeatable agent that pulls usage, support, and account data into a per-customer renewal brief flagging churn risk and concrete upsell openings, 90 days ahead of each renewal date.
Want help shipping this?
We'll build it with your team on your real work — and leave you owning it, not renting it.