New-hire conversational assistant

A new-hire conversational assistant earns a reply only when that reply cites a passage from current policy, benefits, or IT documentation. If retrieval cannot attach that passage, the assistant returns empty. Empty is the designed outcome when the corpus does not support a claim, not a degraded mode.

HR still owns exceptions. Relocation, visa timing, equity treatment, manager-approved time-off, and anything promised in an offer letter sit outside a public handbook lookup. The assistant is a retrieval surface for routine questions in the first 90 days. It does not decide, and it does not close work.

The question mix is familiar: coverage start dates, dependent eligibility, expense tools, laptop access, parental leave versus PTO, and which handbook applies after a cross-border start. Those questions should hit documents you already maintain, not a model that invents a benefits rule because a nearby chunk looked similar.

Treat quality as a binary check. Cited passage: show the answer, keep the ticket visible, and let the hire or HR confirm they are unblocked. No cite: show nothing, escalate with the identifiers you have, and leave closure to a person.

Load the hire's question in the channel they already use, in their words, plus the identifiers retrieval needs: start date, country or state, employment type, job family, and benefits plan code. Without those fields a US medical summary and a global handbook both look relevant.

Do not add a second intake path that HR then has to watch. Capture the raw question and the identifiers as fields you can replay when someone disputes the answer. Authenticated identity from the HRIS beats unverified free text.

Retrieve against a bounded corpus: the current employee handbook, benefits summaries and SPDs, IT onboarding runbooks, and the FAQ HR already publishes. Keep drafts, expired SPDs, manager slide decks, and chat archives out of the index. Rank by document type and effective date, not by fluency.

Then cite or stay silent. A usable answer names the document, the section or page, the effective date, and the quoted or tightly paraphrased passage. If the top chunks disagree, or the plan code does not match the retrieved SPD, return empty and route to HR. Do not average two plans. Do not pick the more generous waiting period.

If the response schema cannot fill document, section, and passage, render empty even when the model produced a confident paragraph. An answer with no policy cite must not reach the hire.

HR handles exceptions after silence, and after a cited answer that still does not fit the hire. The assistant can quote a waiting-period section and still need a benefits partner to interpret a mid-month start. Interpretation is not retrieval.

Keep the ticket open on every path. A cited answer is evidence in the case. Auto-closing because chat replied hides offer-letter conflicts, missing state addenda, and contractors provisioned like employees.

A hire in week one asks whether a partner gets dental coverage on day one or after 30 days, because a cleaning is booked for the following Friday.

The assistant loads the question with employment type, country, start date, and the dental plan code from the HRIS. It retrieves the current dental SPD and the enrollment guide. The SPD states that coverage for eligible dependents begins on the first of the month following 30 days of employment. The reply quotes that section and shows the document name and effective date.

That is a complete quality outcome. The hire sees the waiting-period language. The assistant does not treat next Friday as approved or denied. It does not invent a new-hire courtesy exception, and it does not tell the hire to keep or cancel the appointment.

If retrieval returns a medical SPD and a stale dental FAQ, and neither passage names dependent waiting periods for this plan code, the assistant stays empty. It forwards the question to HR with the identifiers attached. Filling in "30 days, like most of our plans" is inventing a benefits rule.

If the hire then says the offer letter promised day-one dental, that is an exception. HR compares the letter, the SPD, and whatever recruiting promised. The assistant does not reconcile those sources.

The same pattern holds for IT and policy. A question about installing a password manager on a corporate laptop should retrieve the acceptable-use or endpoint standard and cite the allowed-software section, or stay silent if the standard only lists MDM-pushed tools.

Workday, ServiceNow, and Lattice belong here as a class of people, case, and program systems of record. They supply identifiers, hold the ticket, and store onboarding tasks. They are not the conversational model and they are not the policy.

Read start date, location, employment type, and plan code from the HRIS after the hire authenticates. Write the question, the retrieved document IDs, and the cited passage (or the empty result) onto the case. HR ops then has a replayable trail. Chat never becomes the system of record.

Case tools in that class keep the work visible. Deflection means the hire did not need a live agent because a cited passage answered them. The case still waits for a human status change. HRIS fields prevent retrieving the wrong country's handbook. People-program platforms may show onboarding checklists; they do not replace an SPD.

If the hire is asking what they should learn this month, or what their manager needs before day one, do not stretch this assistant. Send them toward role-based learning auto-assignment or a manager briefing generator. Those pages own a different outcome. This page owns a cited answer or an empty one.

An answer with no policy cite is the usual false win. The prose sounds like HR: waiting periods, eligible dependents, typically 30 days. There is no document ID, no section, no effective date. Hires will book the cleaning. Payroll and the carrier will not. Require the cite in the response schema. If the field is missing, render nothing and escalate.

Treating chat as a closed ticket is the second false win. Volume looks better when every bot reply auto-resolves. You then lose the exception queue. Keep status changes with HR. The assistant may hint that the hire appears unblocked. It may not flip the ticket to closed.

Inventing a benefits rule creates trust and compliance debt. It appears when retrieval is close: the wrong year's SPD, the wrong plan's waiting period, parental leave applied to a sick-leave question. Ban generation of waiting periods, eligibility lists, and coverage amounts unless those tokens appear in the cited passage. If the passage is incomplete, empty stays empty.

A quieter miss is answering pay or rewards from a handbook blurb instead of the statement the hire should receive. Route that to a personalized total-rewards statement rather than summarizing competitive benefits from a leftover careers snippet in the index.

Once cited answers and empty escalations behave, connect this assistant to the rest of the first 90 days without merging the jobs. A 30-60-90 plan generator can use role and start date. It should not answer when 401(k) matching begins. Learning assignment can fire after IT access is confirmed. It should not explain VPN exceptions.

Sample weekly tickets where the assistant replied. For each, check three things: Is there a cited passage? Does that passage answer the question that was asked? Did a human still own closure? Flag every invented number, every auto-close, and every empty state you skipped because the prose sounded sure.

Publish corpus rules next to the assistant: which documents are in, which are out, who updates effective dates, and how a hire files an exception. When a policy changes, the index updates before any chatbot phrasing. The assistant has no standing copy of the rule.

The practitioner test is the question you took last Tuesday. If the reply cites the current passage, you can show it to the hire. If it cannot cite, the screen stays empty and HR gets the case. Anything in between is not this use case.