Solution

Answers for staff who weren't here last season

An AI knowledge base for hotels and restaurants answers the question a new server, a covering front-desk agent, or a maintenance tech would otherwise have to interrupt a supervisor mid-service to ask. Calfy builds these systems around the brand standards, allergen matrices, menu and wine notes, property-specific detail, and supplier manuals a hospitality business already has — and every answer comes back with the source document attached, so a person can check it before repeating it to a guest.

· Reviewed by Artur Horimoto, Founder & CEO

The problem turnover creates

Hospitality runs on some of the highest staff turnover of any industry, and that single fact changes what "institutional knowledge" can mean for a hotel or restaurant group. A manufacturer or a law firm can lean on a senior employee who has been there twenty years. A restaurant that rebuilds a third of its floor staff every season cannot. The person who knew that table 14 is the one with the wobbly leg, that the Thursday special always runs out by 8pm, or that the walk-in cooler drain backs up in July — that person may not be working here by the time any of it matters again.

The standard fix is more training: a thicker onboarding binder, a longer shadow shift, a laminated card behind the pass. All of it helps a little and none of it survives contact with a Friday dinner rush. A new server three shifts in does not stop and reread a binder when a guest asks whether the mushroom risotto has dairy in it. They either guess, or they pull a supervisor off the floor to ask — and during service, both of those cost something. A knowledge system does not replace training. It gives the new hire, and everyone covering a shift that is not usually theirs, somewhere to ask the question in the fifteen seconds they actually have, and get an answer that names where it came from.

What the system does on a normal shift

The material varies by property and by role, but the retrieval pattern stays the same across all of it.

  • Brand standards and SOPs. A new hire asks how a specific procedure is done at this property — the folding standard for a turndown, the sequence for opening the bar, the script for a comped meal — and gets the current version, not whatever a trainer happened to say on day one.
  • Allergen and ingredient information. A server checks whether a dish contains a guest's stated allergen and gets the controlled allergen document for that item, current as of the last recipe change, with the source clearly shown.
  • Menu and wine knowledge. A server unfamiliar with tonight's specials or a guest's wine question gets the tasting notes, pairing suggestions, and preparation details a sommelier or chef would give, without pulling either of them away from service.
  • Property-specific detail across a group. Someone covering a shift at a sister property, or a regional manager visiting for the week, asks a question that has a different answer at every site — pool hours, accessible room count, parking, the closest pharmacy — and gets the answer for the property they are actually standing in.
  • Supplier and equipment documentation. A maintenance tech troubleshooting an ice machine or a laundry press gets the manufacturer's manual and the unit's service history for that specific piece of equipment, instead of guessing from memory or waiting on a call to the supplier.

Three questions that used to depend on whoever had been here longest

A new server's first week, without pulling a supervisor off the floor

It is a Tuesday dinner shift, and a server who started nine days ago has a table asking whether the seared scallop appetizer can be made without butter for a dairy allergy. Interrupting the chef mid-ticket to ask is the kind of thing a new hire hesitates to do, and guessing is worse. The server checks the system on the floor tablet. It returns the dish's current prep card, showing butter used in the pan sauce and confirming a dairy-free version exists on request, prepared with oil instead. The server relays exactly that back to the table, seats the order correctly the first time, and never breaks stride to find someone senior. What used to require finding the one person who'd know now takes the length of the walk back to the table.

The allergen question, sourced rather than answered from memory

A guest tells a server he has a severe tree nut allergy and asks whether the pesto on the appetizer menu contains pine nuts. This is the highest-stakes retrieval in the whole system, and it is built to behave differently from every other one. The system does not compose a reassuring sentence about the dish. It pulls up the current, controlled allergen and ingredient document for that exact menu item — the one the kitchen maintains and updates whenever a recipe changes — and shows it to the server directly, with the effective date visible. The server reads the actual document, sees pine nuts listed under tree nuts, and tells the guest the dish is not safe for them, then flags a server-safe substitute from the same document. Nobody guessed, and nobody trusted an AI-generated summary of an allergen with a guest's safety riding on it. A wrong allergen answer is not an inconvenience — it is a safety incident — so this is the one place the system is deliberately built to surface the source and stop, not to sound helpful.

The property-specific answer, in a group where no two sites are the same

A guest-services agent normally based at a group's downtown property is covering a short-staffed weekend at the airport property across town. A guest asks about the pool's closing time. At her home property it closes at 10pm; she has no reason to know it closes at 9pm here, and guessing wrong sends a family down two floors to a locked gate. She checks the system, which returns the airport property's current pool hours specifically — not the brand-wide default, not the downtown property's hours — because the index is built per property, not as one shared brand document standing in for all of them. She gives the guest the correct time on the first try, at a property she is covering for two days.

What this doesn't fix

A knowledge system is only as reliable as the documents behind it, and that matters more in a kitchen than almost anywhere else. If a recipe changes and nobody updates the allergen sheet, the system will confidently return the old, wrong version — which is exactly why the allergen flow is built to surface the controlled document itself, with its date, rather than paraphrase it, so a person checking it has a real chance of catching a stale entry. Keeping that source document current is a kitchen and management discipline the system supports but cannot enforce on its own.

It also does not replace a person's judgement on anything that needs one. A server still confirms the allergen answer with the guest before the order goes in. A maintenance tech still decides whether a repair is within their skill or needs a call to the supplier. The system's job is to put the correct document in front of the right person fast enough that they can use it — human confirmation stays in the loop on anything where being wrong carries real weight, which in this industry starts with anything a guest is going to eat.

Connecting to your PMS, POS, and property documents

This runs on the same approach behind Calfy's knowledge systems work generally, scoped to what a hotel or restaurant group actually has on hand.

Connection. We index what already exists — a shared drive of SOP documents, a POS system's menu and allergen data, a wiki, scanned supplier manuals, a PMS's property-specific content fields. Nothing needs to move into a new system before this works.

Per-property structure. For a group, the index is built to respect that no two properties are the same. A question gets answered for the property the person asking is actually at, not a brand-wide average that is subtly wrong everywhere.

Permissions. Access mirrors what your systems already enforce — a maintenance manual is visible to the people who could already see it, not opened to everyone with a login.

Retrieval, not invention. The underlying method — search the real documents first, then answer from what was found, with the source shown — is the same retrieval-augmented generation approach behind any properly grounded system, and the knowledge base search use case covers the general pattern in more depth. Where the need is a guest's call being answered rather than a staff member's question being looked up, that is the territory covered by voice AI for hospitality instead — a related but separate system.

Frequently asked questions

How does this handle allergen questions safely?

For allergens, the system does not write an answer — it surfaces the current, controlled allergen document for that specific menu item, with its source and effective date shown, and a staff member reads and confirms it before telling the guest. That is a deliberate design choice: a wrong allergen answer is a safety incident, not an inconvenience, so this flow sources rather than composes.

How do you keep the information current when recipes and SOPs change?

The system indexes whatever your kitchen and management team maintain as the source of truth, and an update to that document flows into the index on a schedule matched to how often it changes. It cannot catch a recipe change nobody recorded anywhere — keeping the source documents current stays a kitchen discipline the system supports, not one it replaces.

Does this work across a group with several different properties?

Yes. The index is built per property, so a question about pool hours, parking, or accessible rooms returns the answer for the specific site someone is asking about, not a brand-wide default. That matters most for staff covering a shift outside their usual property.

Will our team need training to use it?

Barely. It is built to answer a plain question on a phone or floor tablet, not to hand staff a new system to learn during a shift that gave them no spare time to learn it. That matters in an industry where the person on shift today may not be the person on shift next season.

How much does this cost, and how long does it take to go live?

It depends on how many source systems it connects to — a POS, a PMS, a shared drive of SOPs — and whether it spans one property or a group. We scope every engagement and agree a clear price before any build work starts, and most properties see the system answering real questions within weeks.

Bring the question your floor staff ask a supervisor most often, or the allergen check your team still does from memory. Thirty minutes on a free strategy call is enough to tell you whether a knowledge system fits your property, and roughly what it would take to build.

See what this looks like for your business

Thirty minutes, your actual workflow, and a straight answer on whether this is worth building.

Free 30 minutes. No pitch deck. You leave with a plan either way.