AI Operational Supervisor

What you asked for

One layer that summarises the operation and tells you only what requires attention, so management stops checking the PMS, guest messages, LINE, maintenance, cleaning, inventory and pricing separately.

Today: 82% occupancy / 9 check-outs / 11 check-ins
Cleaning:    7/9 completed, 2 outstanding
Maintenance: 2 unresolved issues
Guest:       3 responses awaiting approval
Inventory:   2 items need ordering
ADR:         ¥26,400 vs market ¥28,100 → suggested ¥27,500

"These are the 5 things requiring your attention today."

And over time, pattern detection: recurring maintenance problems, cleaning complaints, abnormal consumption, pricing opportunities.

The honest framing

This is not a module. It is a report.

Everything in your example digest is a query against data the previous four chapters produce. Occupancy and check-in counts come from the reservation spine. Cleaning status comes from the task engine. Awaiting-approval counts come from the guest inbox. Inventory and ADR come from their own modules. There is no new capability required — which is exactly why it must be built last, and why it is nearly free when it is.

Build it first and it summarises nothing, because there is nothing to read. This is the most common way ambitious operations projects fail: the dashboard is the most visible piece, so it gets built first, and then it displays four numbers pulled from spreadsheets and quietly stops being opened.

The good news: the digest is perhaps two weeks of work once the modules exist. It is the cheapest thing in this handbook and the most visible. That is a very good place to end a build sequence.

Three layers, and only the third needs AI

It is worth separating these because they have completely different reliability characteristics, and conflating them is how you end up with a system nobody trusts.

Layer 1 — The deterministic digest

Plain queries. No AI. Fully reproducible, testable, and correct by construction.

── Monday 14 September ─────────────── Japan · 14 units ──

  Occupancy      82%  (11/14 occupied tonight)
  Check-outs      9   ·  Check-ins  11   ·  Same-day turns  7
  Revenue MTD    ¥4,182,000   (+11% vs Sep 2025)

  CLEANING        7 of 9 verified · 2 outstanding
                  ⚠ Osaka-04 checkout 10:00, check-in 15:00, unassigned
  MAINTENANCE     2 open  ·  1 overdue (Kyoto-02 water heater, 4 days)
  GUEST           3 replies awaiting approval  (oldest 41 min)
                  1 escalated — negative sentiment, Namba-01
  INVENTORY       2 items below reorder point (Bali · Canggu store)
  PRICING         14 rate recommendations pending
                  portfolio ADR ¥26,400 vs market ¥28,100
  COMPLIANCE      Shibuya-03 at 164/180 minpaku days — cap ~14 Nov
                  Canggu-02 NIB expires in 38 days

This layer alone delivers most of the value you described. It should be built first within the supervisor, and it should be right one hundred percent of the time, because everything else you build on top inherits its credibility.

Layer 2 — Deterministic triage

Ranking, not generating. A scored rule engine that answers "what are the five things."

Score each open item on consequence × urgency × decay:

Then take the top five. Rules, not a model — because "which five things matter" is a business policy you should be able to read, argue with, and change, not an emergent behaviour you have to probe. When you disagree with the ranking, you edit a weight. That is the whole point.

Layer 3 — The narrative and the patterns

Here an LLM earns its place, and only for two jobs:

Writing the summary. Turning the ranked list into three sentences a person reads on a phone at 7am. Purely presentational, grounded entirely in the structured digest, and — critically — the model never computes a number. Every figure is passed in and quoted. If the digest says 82%, the narrative says 82%. A supervisor that occasionally invents an occupancy rate is worse than no supervisor, because it poisons every number next to it.

Finding patterns. This is the genuinely interesting part and it is what you were reaching for:

• "Kyoto-02's water heater is the third HVAC/plumbing ticket there in 90 days — total labour ¥84,000 against a ¥140,000 replacement." Recurring-issue detection with an economic conclusion.
• "Cleaning defect reports on Namba units are 3× portfolio average, concentrated with one cleaner on Saturday turns." Correlation across staff, unit and day-of-week.
• "Bath towel consumption at Canggu-03 is 40% above CPOR for four weeks with no occupancy change." Anomaly, likely theft, leak or miscount.
• "You have declined 71% of upward rate recommendations for Golden Week and are now 18% below comp median on those dates." The system noticing you, which is the most valuable pattern of all.

Structure this as weekly, not daily. Patterns need a window, and a daily pattern report becomes noise within a fortnight — at which point the whole digest loses its authority.

Delivery

Match the channel to the population, as everywhere else in this handbook:

• Morning digest — one push, 7am local, to LINE for ops managers, email for you and any owners. One tap opens the web view.
• Web dashboard — the same digest, drillable. Every number clicks through to the underlying list. Non-negotiable: an unclickable number is a number nobody trusts twice.
• Interrupts — a very short list, pushed immediately, that must be genuinely rare or it destroys the value of everything else: guest escalation, unassigned same-day turn inside four hours of check-in, integration sync failure, licence lapse.
• Weekly pattern review — email, Monday, for you and ops leadership.

One rule that determines whether this system survives its first month: every alert must be actionable and dismissable. An alert you cannot act on is noise, and an alert you cannot dismiss trains people to ignore the channel. Track your own alert-dismissal rate; if it climbs above about a quarter, the thresholds are wrong and the fix is to tune them, not to send more.

The owner report, since you mentioned the management fee

You framed all of this against a 10–15% management fee on gross profit including marketing. That reframes the supervisor: the same data spine that produces your morning digest produces a monthly owner report, per property, automatically — occupancy, ADR, RevPAR, RevPAR index against comp set, maintenance spend, cleaning cost per turnover, review score trend, and the pricing decisions taken with their evidence.

That report is a commercial asset, not an internal tool. It is the artefact that justifies the fee, differentiates you from a manager who sends a spreadsheet, and — because it is generated from the same audited spine — costs almost nothing once the modules exist. Build it as part of this chapter, not as a later project.

Effort and risk

Effort: low, and entirely dependent on what precedes it. Two weeks for layers 1 and 2. Another week for the narrative and pattern jobs. Add a week for the owner report. If it looks like more than that, the underlying modules are not finished and building the digest will paper over it.

Risk: low technically, moderate in credibility terms — and credibility risk is the one that actually matters here. A digest that is wrong once gets checked twice; wrong twice and it gets ignored permanently. So: layer 1 must be deterministic and tested, layer 3 must never compute, and every number must be traceable to a row someone can open.

Decisions for you

1. What are your five? The ranking table above is a proposal. Your actual priorities may differ — some operators would rank pricing opportunity above maintenance. Tell me the order and it becomes the weights.
2. Digest timing and audience. One 7am digest, or a morning and evening pair? Who receives which — you, ops leads, owners?
3. The interrupt list. What genuinely justifies a push at 2am? Keep it to three or four items.
4. Do owners see anything directly? An owner portal is a natural extension of the monthly report and a real differentiator for winning management contracts. Worth deciding early because it changes the data model's access rules.
5. How do you want to be told you are wrong? The pattern engine will eventually tell you that your own overrides cost money. Decide now that you want that, because it is the highest-value output in the whole system and the easiest to switch off.