A staff question assistant for your SOPs: what to load, what it may answer and who checks it

A staff question assistant answers employees' questions from the hotel's own written procedures, not from the open internet. Load one current, dated version of each procedure with a named owner. It may answer routine how-to questions, must refuse anything legal, medical, safety-critical or disciplinary, and every answer shows its source page. One person reviews it weekly.

In the first week of the season a new receptionist needs to know how to void a charge, whom to call in a power cut and which form the night audit uses. The answer exists somewhere in the building: in a folder nobody opens, in the memory of one colleague, and in a messaging group from two years ago. An AI assistant for hotel SOPs can make those answers findable, but only if the documents behind it are current, dated and owned. This guide explains what such an assistant does, what to load, what it must refuse, and how to test it before staff rely on it.

Which questions do new staff ask every season?#

The same twenty or so, in every department, and the answers usually live in three places at once: a folder, a colleague's memory and a messaging group. Collect the questions before you choose any tool. Keep one shared list for two weeks, one per department, and write each question down in the words the staff member used. That list is your requirement document, and later your test set.

The recurring shapes are predictable: how to void or move a charge on a guest bill, what to do in a power cut or a water leak, which form the night audit completes, who authorizes a late checkout, how to hand a complaint to a manager, what to do when a guest's document does not match the reservation. Every hotel has its own list, but the categories repeat.

The answer drifts because three versions of the same procedure are in use at once: the printed folder, the noticeboard, and the person who has done the job for nine years. The newest version is not always the one being followed. The cost is not dramatic, it is cumulative: the same request handled two ways on two shifts, and no way to show an auditor or an owner which rule was applied.

The written procedure is the precondition. If the step is not written down, no assistant can answer it. That gap is the first finding of the project, before any software is involved. If your procedures need work first, start with the practice of writing hotel SOPs that staff actually follow.

Where the answers live today#

What does an assistant over your own documents actually do?#

It searches your written procedures for the passages that match the question and writes an answer from those passages, quoting where each part came from. It does not remember your hotel, and it does not learn from the questions staff ask.

The mechanism in plain words: your documents are split into passages and stored with a search index; the question is matched against that index; the answer is composed only from the passages that were retrieved. The technique is usually called retrieval, and the answer is generated from what was retrieved rather than from general knowledge. For a fuller explanation, see answering from your own documents: retrieval explained for hotels.

Consequence one: the answer is capped by the document. A vague procedure produces a vague answer, and no amount of tooling repairs it. Consequence two: it cannot answer what is not in the loaded set, and it has to be built to say so instead of filling the gap with something plausible.

It is not a general chatbot. A general chatbot answers from what it absorbed in training, cannot show your page four, and cannot tell you whether your version or an older one is in force. It is also not an assistant that reads live system data, such as today's arrivals or a guest's open bill. That is a different tool with different risks, and blending the two in a single answer is where operational mistakes begin.

What it is genuinely good at is a stable lookup, in the language the staff member reads, at three in the morning, without waking the duty manager.

Three tools that look similar and are not#

What should you load, and what should you leave out?#

Load the current, dated version of each operational procedure with one named owner per document. Leave out anything personal, anything contractual, and anything that only a qualified person should decide.

One version per procedure. If two versions sit in the loaded set, the assistant will quote whichever it retrieves. When a new version is signed, the old one leaves the set the same day. Date every document and show that date in the answer: an undated procedure is indistinguishable from an outdated one, and staff cannot judge what they are reading.

Name one owner per document: the person who confirms it is current and who signs the next version. A document without an owner is a document nobody will retire.

Keep personal data out: no guest names, no copies of identity documents, no staff files, no disciplinary records. In Türkiye personal data falls under KVKK, the personal data protection law; where you are unsure whether a document may be loaded, take it to your legal adviser before it goes in. Leave out anything that changes daily, such as rate lists and availability, and point staff to the system that holds the live figure instead.

Include the forms, checklists and contact trees staff need at the moment of the question, not only prose. A procedure that says "use the incident form" is useless if the form is not in the set. Mark translations as translations, carrying the same version number as the original, so a translated answer can be traced back to the signed source.

A load list and a leave-out list#

Which questions must it refuse or hand to a person?#

Anything that turns on a legal rule, a person's conduct, a medical situation or money above a set limit goes to a named person, and the assistant must say so in one sentence instead of attempting an answer.

Legal and regulatory questions come first: whether a guest's document may be copied, what a notification must contain, what an invoice must show. The assistant may quote your own written rule; it must not interpret the law or guess at a date or a penalty. For how to verify legal texts yourself, see asking AI about a law in Türkiye: how to check article numbers and dates.

HR and disciplinary questions go to the manager, always, even when the general rule is written down: lateness, leave in an individual case, a warning, a dismissal. Guest safety and medical situations follow the written emergency procedure, and the safety officer or duty manager leads. An assistant is not part of the response chain and must not appear to be. For money above a threshold you set, refunds, compensation, rate overrides, the assistant may state the rule and name who approves; it never approves.

Anything not in the loaded documents gets one answer: "I could not find this in the current procedures", followed by the name of the person who owns the topic. Refusal has to be designed, not hoped for. A tool that always produces an answer will produce one for the question it should have handed over, and it will sound just as confident as when it is right.

How does an answer show its source?#

Every answer names the document, its version date and the section it came from, and the staff member can open that document from the answer in one step.

Three parts make an answer usable: the step itself, the source, and the person to ask when the case does not match the written rule. A citation is not decoration. If the answer cannot cite a passage from the loaded set, it should not answer at all.

The date matters more than the title. A correct answer taken from a superseded procedure is still wrong, and the staff member has no way to notice unless the date is on the screen. "Not found" must look different from "found": different wording, a different visual treatment, and a route to a named person.

Hedging words such as "usually" or "probably" are a warning sign in an operational answer. The procedure either states the rule or it does not, and the assistant should not smooth over the gap. Keep the underlying documents reachable by staff, so an answer can be checked rather than only trusted. A source nobody can open is a claim, not a citation.

Weak answer, usable answer#

How do you test it with fifty real questions before staff rely on it?#

Write down fifty questions staff actually asked last season, have each department head write the correct answer from the current documents, and check every assistant answer against that answer key before anyone uses it on shift.

The questions come from the two-week log, last season's messaging group, the front desk notebook, and the questions asked during shift handover. The answer key is written by the people who own the procedures, not by the person who bought the tool. If the department head cannot write the correct answer, the procedure is not ready.

Score four outcomes rather than one: correct with the right source; correct but citing an outdated document; wrong; correctly refused and handed to a person. The second and fourth reveal design faults that a simple pass rate hides. Fix the document before fixing the tool: a wrong answer that traces back to a vague procedure is a documentation failure, and changing the tool will only hide it.

Test in the language staff will actually use, and test the awkward phrasings, not the tidy ones. Real questions arrive as half sentences with a room number attached. Re-run the set whenever a procedure changes and once before each season, and keep the results, because the comparison between two runs is the only evidence that the thing is holding. Treat this as a working approach, not a standard or a certification. It is a method a hotel can run itself, and it should be described that way to owners and auditors. For the broader testing picture, see how to test a hotel AI assistant before staff or guests use it.

The test, step by step#

How do you keep it current through the season?#

Give one person the job, review the loaded documents weekly, and retire a superseded procedure on the same day the new one is signed.

There are two levels of ownership: one person owns the whole set and the weekly review, and one named person owns each document. The second is what makes the first possible. The weekly review is short and has one input: the questions the assistant could not answer or refused. That list is the cheapest missing-procedure list a hotel will get, and it comes from real shifts.

Keep a change log of what was loaded, when and by whom. It is the only way to explain an answer given three weeks ago to an owner, a guest or an auditor. Re-load before high season, when temporary staff arrive and procedures change fastest, and again when a department rewrites a routine.

When the tool and the document disagree, the document wins and the tool is corrected. The reverse direction is how a hotel ends up following software instead of its own rules.

What to do next#

This week, before you look at any tool, write down the twenty questions your team asked most last season and check whether a current, dated, owned document answers each one. That list is your business case, your load list and your test set at once. The assistant is only worth buying once the documents behind it can survive the questions.