At the multi-unit groups I have run, the area director's phone starts vibrating around 9pm. A general manager in the middle of a Friday close needs to know if they can comp a 68 dollar guest check. Another one at 10:40pm wants to know if a minor can close on a school night. A third at 11:15pm cannot remember which vendor covers the walk-in compressor after hours.
None of those questions requires judgment. All of them are already answered somewhere in the ops manual, the HR handbook, or a Slack thread from six months ago. The area director answers them anyway, because the general manager has no faster path to the answer than the phone call.
The chatbot is that faster path. Not a moonshot. Not an agent that runs the restaurant. A reference bot that reads the same documents a new area director would read on day one, and answers with a citation the general manager can verify.
What the bot handles well
Across two multi-unit groups I have shipped this in, the bot handles four categories cleanly. These four cover about 70 percent of the after-hours volume.
Policy questions
Comp and void thresholds. Refund limits by manager tier. PTO accrual rules. Uniform and grooming standards. Delivery and third-party marketplace pricing rules. Cash handling limits at close. These are questions where the answer is a number or a paragraph and the whole point is that there is a right answer. The bot pulls it back with the source section attached.
Scheduling and labor rules
Overtime approval thresholds. Minor labor rules by state, which was a real pain point at Hana Group with 21 franchise units across six states. Shift trade approvals. Split shift premium eligibility in California. Meal break waivers. The general manager types the state and the situation, and the bot returns the rule with the code reference.
Vendor contacts and reorder cutoffs
Who to call for a walk-in that stopped running at 11pm. What time the produce order cuts off on Saturday. Which POS vendor to escalate a Toast printer failure to. The bot returns the phone number, the escalation email, and the cutoff time in one message. This alone eliminates half the after-hours pings at any group with more than five units.
Food safety references
Cook temperatures. Cooling times. Sanitizer PPM ranges. Allergen protocols. Illness reporting steps under FDA 2022 Food Code. Every general manager and shift lead has ServSafe. Nobody remembers the specific number under pressure. The bot returns it in three seconds, with the section of the food safety manual noted.
One test I run before rolling the bot out. I pull the last 90 days of after-hours DMs to the area director, categorize them, and count. If more than 60 percent fall into those four buckets, the bot pays for itself. In every group I have measured, the share is closer to 70. At Zareen's it was 74 percent, and the biggest category by volume was scheduling rule questions from shift leads covering the manager on close.
Fig. 1 · The retrieval architecture, source to Slack.
What the bot should never answer
This is the part most operators skip and then regret. The bot needs a hard list of topics it refuses.
Personnel. If the question names a specific employee, the bot answers with the escalation path, not with policy interpretation. "Should I write Maria up for being late" gets routed to a human every time. The bot does not diagnose people.
Discipline. Progressive discipline steps are a policy question the bot can answer in the abstract. Applying them to a real situation is a human call. The bot draws that line explicitly.
Health and safety escalations. Guest injuries, allergic reactions, employee injuries, alleged harassment, and health department visits do not go to a bot. They get a page to the on-call area director and a text to legal. The bot's job here is to route, not to advise.
Judgment calls. Should we close the dining room early because of the weather. Should I send a line cook home after they showed up angry. Should I fire this vendor. Those are the calls the general manager is paid to make with a person on the other end of the phone, not with a chatbot.
How you stop it from hallucinating
The single most important rule. Every answer has to cite the specific source document and section. If the retrieval layer does not surface a strong match, the bot returns a short refusal and offers the escalation contact. It does not fill the gap with a plausible-sounding guess.
That one rule kills roughly 90 percent of the failure modes. The remaining 10 percent are cases where the source document itself is ambiguous or contradictory. Those are the answers you want to see, because the fix is not on the model side, it is in the manual. You go back and rewrite the section.
Weekly review of the bot's transcript for the first eight weeks. Not a compliance exercise. An editing exercise. Every time the bot got something confused, someone updates the source doc so it will not happen again. By month three the review cadence drops to monthly, and the bot is measurably more accurate than the 2019 PDF nobody read.
The bot does not know your operation. It knows your documents. If the documents are bad, the bot is bad. That is a feature, not a bug.
Where it lives matters more than which model runs it
General managers do not open new apps. They open Slack. So the bot lives in a Slack channel called #ask-ops, or as a slash command inside the channels the general managers are already sitting in. Ask in the channel, get a threaded reply with the answer and the citation. That is the entire surface.
A few operators I have talked to built a beautiful web app for their bot. Six weeks in, nobody was using it. They rewrote it as a Slack integration in a weekend and adoption went from 12 general managers a week to 47. The lesson is the same one I learned building operating dashboards. The tool has to be one keystroke from the workflow the field is already in.
The tooling, concretely
For a group under 25 units, this is the stack I would build tomorrow. None of it is exotic.
- Notion or Google Drive as the source folder. One place. Curated. Version dated.
- Custom GPT inside ChatGPT Team, or a Claude Project. Either one indexes the folder and enforces the citation rule in the instructions.
- Slack for the surface, wired through Zapier or Make if the bot is not already a native app in the model provider you chose.
- Airtable for the vendor contact list, if you want the bot to answer with a live phone number rather than one buried in a PDF.
- A shared review doc where a human logs the questions the bot got wrong that week. Ten minutes on Monday. Zero fancy tooling.
Total license cost under 200 dollars a month for a mid-sized group. The variable cost is human curation time. In the first two months plan on ten to twenty hours a month. After that it settles to two or three hours a week.
Human-in-the-loop is the whole game
The bot is not there to replace the area director. It is there to protect the area director's attention for the questions that actually need judgment. Reference questions get answered in Slack in three seconds. Judgment questions get the phone call they deserve, and the area director is now awake enough to answer them well because they did not spend the last two hours on comp policy trivia.
That framing changes how the field feels about the bot. Nobody worries the bot is going to take their job. The area director likes it because they sleep more. The general manager likes it because they get answers faster. The operator likes it because the group starts running the same way in unit 3 as it does in unit 17.
How to roll it out without losing the room
The rollout matters more than the model choice. If you drop a chatbot in front of 21 general managers with an all-hands email, half of them will try it, three will find a weird answer, and by week two you have a mixed reputation you cannot walk back.
The sequence I use now. Pick two general managers who are already comfortable with software and who will tell you the truth. Give them the bot for two weeks. Sit with them once during that window. Watch what they ask. Fix the top three annoyances. Then expand to five general managers and repeat. Only after that go wide. The whole rollout takes six weeks and by the end the bot is being defended by the field, not by corporate.
One more piece. Announce the bot as an area director tool, not a replacement for the area director. The general managers understand that framing. The area directors relax about it. Nobody gets territorial about who "owns" the policy answer. The channel just starts being useful.
Mistakes I made the first time
I loaded too much into the source corpus on version one at Zareen's. Every deck, every PDF, every training video transcript. The bot got noisier and less accurate. Version two used only the current handbook, the current SOP set, and the last twelve months of resolved Slack threads. Accuracy jumped.
I skipped the citation rule for two weeks because the answers looked good in testing. Then it hallucinated a comp threshold that did not exist and a general manager comped a 300 dollar guest check on the bot's authority. The rule went in the next day and never came out.
I built the first version as a standalone web app because it demoed well. Nobody used it. I rebuilt it as a Slack channel and it became the second most active channel in the workspace inside a month.
The point
A general manager who can look up a policy in three seconds runs a different restaurant than a general manager who guesses or waits until Monday. Consistency across units improves before the labor savings show up on any report. The area director's phone rings less at 11pm.
The bot is not a magic operator. It is a well-organized copy of documents you already had, made searchable in the surface your field team already uses. That is what makes it work. That is also why it is easy to underestimate until you actually ship it.
Build it small. Ship it into Slack. Force the citations. Review the transcript every Monday for eight weeks. That is the whole recipe.