Deep Hospitality · a design study · 13 Nov 2026

The dashboard of the future is mostly silence.

You asked what a dashboard looks like when AI sits at the core of a hotel and people stand at the edge. My answer, after the research: the wall of tiles goes away, the screen stays for a smaller job, and human supervision shrinks to four things that do not disappear. Below is the evidence, a schematic of how we get there, and a prototype you can run.

AIcore hostsengineersattendantscooks guests

Your future-state ring, with the one thing it leaves out. The guest is outside the ring, so the edge is where the hotel meets the guest. Being on the periphery puts people in the most important seat in the building.

The view in six positions

An independent reading of your brief

You told me not to agree with you. On four points I do not, and the research is why.

01

Watching is the job to abolish

Lisanne Bainbridge wrote in 1983 that a person cannot hold attention on a display where little happens "for more than about half an hour".1 A manager watching green tiles for an AI that rarely fails is that experiment with a salary. A better dashboard makes it worse.

02

The approve button is the same trap

The 21-agent picture keeps a person in charge through "Proposes. I approve." Bainbridge covered that too: people "can write down numbers without noticing what they are".1 A review of 41 oversight policies found they give "a false sense of security".2

03

The machine should call the person

Waymo's car stays in control and asks a human a specific question when it needs one;3 about 70 people cover some 3,000 vehicles.4 Toyota's line stops itself and signals for help.5 Initiative moves to the system. That inversion is the real change.

04

The screen survives

Humane's pin was shut down and sold for $116 million.6 OpenAI's screenless device is delayed to 2027 and the company is reported to be building a phone.78 Ears and wrists carry "come now". Eyes still do the comparing.

05

Supervision shrinks to four things

Setting terms in advance, answering what the core asks, auditing a sample, and holding the stop. When Anthropic let one AI manage another in a real shop, the manager approved leniency eight times as often as it refused, because it shared the worker's blind spots.9

06

The edge is most of the hotel

Labour is 56% of hotel operating cost10 and guests in a 3,250-person survey preferred a human at all nine touchpoints tested.11 The AI core is proven for pricing and thin elsewhere. Nobody has built the right-hand picture yet, in hotels or anywhere.

The schematic you asked for

Seven forms, from the printed report to the house itself

Fifty years changed the tools and left one thing alone: a person goes to a screen and reads it. Stephen Few's 2004 definition still ends with information "monitored at a glance".12 The next three stages are where that assumption breaks. Step through them.

Bars are a schematic of direction, drawn by me. They are not measured data. Stages 5 to 7 and their dates are extrapolation. Watch the third bar: watching falls to nothing while judging barely moves.

Why the first four stages look alike

Each one fixed the previous bottleneck and kept the reader. Surveys put business-intelligence adoption at roughly a quarter of employees, flat for years while vendors reported rising use.13 Benn Stancil's line is the best description of what a dashboard is: BI tools are "borders, handing off data between machines and humans".14 When the machine also acts, the border moves.

Why stage 4 is a waypoint

Microsoft calls Agent 365 "the control plane for AI agents";15 Salesforce's command centre reports "error rates, escalation frequency, agent latency".16 These are IT dashboards pointed at agents. None reports the work in the hotel's terms: what was decided, what was promised, what is waiting on a person. Gartner expects over 40% of agentic projects to be cancelled by the end of 2027.17

Working prototype

One day in a 180-room hotel, 13 Nov 2036

Drag the clock or play the day. The core takes thousands of decisions and writes each one down. Watch how few reach a person, who they reach, and in what form. When a call has buttons, answer it.

09:00

The house

0decisions taken by the core today
0calls that needed a person
0minutes of human attention asked for

What reaches a person

Promises · allowance left this month

The one dial that matters

How much may the core spend on a guest gesture without asking? Ritz-Carlton answered this for people decades ago: up to $2,000 per guest, per incident, without approval.18 The same question now applies to software. Move the limit and see what it costs.

Limit:

What that setting buys

calls a day to the duty host
decisions a month a host would have reversed

Everything in this prototype is invented: the hotel, the guests, the numbers. The calls-per-day figures are illustrative arithmetic. The threshold they are judged against is real: process-industry guidance treats about one alarm per ten minutes as the long-run load a trained console operator can handle, and about twelve an hour as the ceiling.19 A host on a hotel floor should get far fewer.

Your question: how do we know the processes are working?

Six signals that need nobody to watch

Industries that already live with automation settled on the same handful of devices. Each one is in the prototype above.

1. Silence is the normal state

An Airbus flight deck shows no lights when everything is set and fit to fly.20 The known weakness is that dark can also mean dead, which is why a quiet system needs an independent heartbeat proving its monitors are alive. Bainbridge asked it first: who notices when the alarm system is not working?1

2. Promises with an allowance for failure

Google's reliability engineers agree in advance "how unreliable the service is allowed to be" and track what is left of that budget.21 A hotel version: room ready when promised, dinner in thirty minutes, no guest asked twice. The instrument shows allowance remaining. Activity is left in the ledger.

3. An attention budget

At Three Mile Island about a hundred alarms arrived within minutes and an operator said he "would have liked to have thrown away the alarm panel".22 Knight Capital received 97 warning emails before losing more than $460 million in 45 minutes.23 Google's rule is "Every page should be actionable".24 Interruptions are rationed like cash.

4. The ask rate

How often does the core ask a person, and how often does the answer change what it does? It is the most informative number about an autonomous system. Robotaxi operators declined to give it to a US Senate inquiry.25 No agent platform I found reports it. It should sit at the top of the owner's letter.

5. A second core that disagrees

AI can take the first layer of oversight if the checker is built differently from the thing it checks. Control research gets its safety from a separate trusted model plus a small budget of human audit.26 Where a weaker system oversees a stronger one, success in one study ran from 9% to 52%.27 So the audit core plants test cases and a person samples its findings.

6. Lines the core cannot cross, and a cord

Toyota's line halts itself on an abnormality and any worker can stop it.5 EU law will require that overseers of high-risk AI can "interrupt the system through a 'stop' button".28 In the prototype three decisions are never the core's: walking a guest, life safety, and delivering a gesture.

What replaces each report a general manager reads today

My synthesis. No hotel was found operating this way.

TodayWhat it showsIn an AI-core hotelBorrowed from
Daily flashYesterday's occupancy, rate and revenuePromises kept, with allowance left for eachError budgets in software reliability21
STAR reportShare against competitorsKept as it is. It is outside evidence the hotel's own agents cannot markMeasuring a dark factory by what leaves the gate
Pace and rateBookings on the books, price positionException dates raised by the pricing engine, limits in force, an override logRevenue systems on autopilot: about 70% of Atomize users, by the vendor's account29
Review scoresSampled opinion, weeks lateA recovery queue: failures that need a person's warmth nowService-failure research; Klarna's reversal30
NothingNot shownThe ask rate, and how often a person's answer changed the outcomeWaymo fleet response3
NothingNot shownHeartbeat and fallback: proof the monitors are alive and people can still take overThe dark-cockpit weakness; Sonder's collapse31
NothingNot shownThe authority ledger: what each agent and person may decide alone, and how they used itRitz-Carlton's $2,000 rule18
Your question: will it even be on a screen?

Yes, for a smaller job

Conscious human throughput is about 10 bits a second.32 A dashboard of 200 numbers wastes nearly all of them, which is the case for the machine doing the scanning. It is also the case for keeping vision where it is strongest: side-by-side comparison, magnitude, history. Each sense gets the job it is good at.

ChannelWhat it carries wellWhere it failsIts job in the house
The placeOne or two variables, to everyone at once, without anyone looking. Weiser and Brown called it technology that moves "from the periphery of our attention, to the center, and back".33Capacity. The famous claim that an ambient orb cut peak energy use 40% could not be traced to any primary source.State of the house: quiet, busy, or needs someone
The wrist"Attend now", privately, in a noisy roomAnything with detailThe summons
The earA sentence of context while hands and eyes are with a guestMagnitude. Anaesthetists could hear which way a patient's oxygen was moving from the monitor's tone and could not tell how far.34 Voice-first devices have a poor record.6The brief, and the call
GlassesA glance without reaching for a device. The form with momentum: IDC forecasts about 39 million units in 2030, most of them without a display.35Still accessory scale beside a billion phones a year. Display glasses are a screen that moved.The glance card, for those who choose to wear them
The screenComparison, history, evidence, anything you need to point at with someone elseAs wallpaper. Nobody can watch it.1Inspection after a call; the owner's charter; the audit
PaperAccountability and signature. Air traffic control kept paper flight strips for decades for good reasons, and the US was still replacing them in Nov 2026.36Anything liveThe monthly letter and the charter

On Jarvis. It is a voice that serves one person standing at the centre of the action. In your future-state ring nobody stands there. The useful part of the idea is the word in the ear. The floating wall of holograms is a dashboard that learned to fly, with the same reader the research says cannot keep watching. Brain interfaces are a clinical technology for paralysis, with a few dozen trial participants, and nothing on any roadmap I found writes information into a healthy manager.

As instructed

Where I disagree with the brief

You suggestedDesign a dashboard of the future, far different from today's.
My viewThe far-different thing is that most of it is absent. What remains is a charter, a short queue of calls, a ledger read mostly by another machine, and a screen you open after you have been called. If I had drawn a beautiful 2036 dashboard I would have drawn stage 4 in better clothes.
You suggestedSomething like Jarvis.
My viewKeep the earpiece, drop the hero. Jarvis assumes a human at the centre issuing commands. In the house, direction runs mostly the other way, and at 21:10 in the prototype it runs back: a host tells the core "look after 207". A person in the room is still the best sensor a hotel has.
You suggestedThe 21-agent dashboard looks contrived.
My viewAgreed, and I can say why. It is an org chart with software in the boxes and an approve button on each. The results its author reports for a 30-room hotel are his own and unverified.37 It is an honest picture of 2026. Its weakness is the human role it implies: approving things all day.
You suggestedIn ten years it may not be on a screen.
My viewThe base rate says it will. Every screenless AI device since 2023 has failed, been absorbed or been delayed.68 Even kitchens moved from paper rails onto screens. What changes is who starts the conversation. You stop going to the screen. Something taps you, then you look.
You suggestedUltimately we may not need a human to supervise AI.
My viewOn current evidence, no. Capability is doubling roughly every six to seven months on METR's measure,38 yet in Anthropic's own data a human was in the loop on 73% of nearly a million tool calls, and its conclusion was that oversight means "being in a position to intervene when it matters".39 The law agrees: a tribunal held Air Canada liable for what its chatbot said.40 Someone signs.
Your diagram showsExperts and levels 1 and 2 as letters on a ring around the AI core.
My viewTwo amendments. Put the guest in the picture, outside the ring, and the periphery becomes the front. And budget for Bainbridge's last irony: "the most successful automated systems, with rare need for manual intervention, … may need the greatest investment in human operator training."1 The ring is thinner and each person on it is worth more.

What would change my mind

Two kinds of evidence that do not exist today. An independently audited operating result from an agent-run hotel: every one I found is self-reported. And controlled data showing AI monitors holding up against more capable systems with no human audit behind them. Either would move the dial in the prototype to the right.

What I would do with this at Deep Hospitality

Ask any agent vendor three questions no screen in this research answers. How often does the core ask a person? How often does the answer change what it does? How would we know if the monitor had died? Then write one hotel's charter on a single page, before buying anything.

Sources

What this rests on, and how firmly

Marked primary where the research read the page itself, secondary where it relied on coverage, an abstract or a summary, and self-reported where the only source is the party making the claim. Quotes from secondary sources should be checked against the original before this is shared outside.

  1. Lisanne Bainbridge, Ironies of Automation, Automatica (1983). ckrybus.com [primary]
  2. Ben Green, The Flaws of Policies Requiring Human Oversight of Government Algorithms. arxiv.org [primary, abstract]
  3. Waymo, Fleet response (May 2024). waymo.com [primary]
  4. Reuters via The Star on Waymo's remote assistance staffing (18 Feb 2026). thestar.com.my [secondary]
  5. Kaizen Institute, Jidoka. kaizen.com [secondary]
  6. Fortune on HP buying Humane and the AI Pin shutting down (19 Feb 2025). fortune.com [secondary]
  7. Gizmodo on OpenAI's reported phone (27 Apr 2026). gizmodo.com [primary]
  8. MacRumors on the OpenAI device delay to 2027 (10 Feb 2026). macrumors.com [secondary]
  9. Anthropic, Project Vend phase two. anthropic.com [primary]
  10. LODGING Magazine, A Lens on Labor. lodgingmagazine.com [secondary]
  11. Mews guest survey press release (7 Oct 2026). mews.com [self-reported vendor survey; method undisclosed]
  12. Stephen Few's 2004 definition of a dashboard, as reproduced by Tableau. tableau.com [secondary]
  13. TechTarget on the BARC/Eckerson BI adoption survey. techtarget.com [secondary]
  14. Benn Stancil, Is BI dead? (17 Sep 2021). benn.substack.com [primary]
  15. Microsoft, Agent 365: the control plane for AI agents (18 Nov 2025). microsoft.com [primary]
  16. Salesforce, Agentforce Command Center. salesforce.com [primary]
  17. Gartner press release on agentic AI project cancellations (25 Jun 2025). gartner.com [secondary, via trade coverage]
  18. Hospitality Net, An Empowerment Lesson From The Ritz-Carlton. hospitalitynet.org [secondary]
  19. Emerson, Alarm management by the numbers (EEMUA 191 and ISA 18.2 benchmarks). d1-live.emerson.com [secondary; the paid standards were not read]
  20. Airbus presentation on flight-deck philosophy, hosted by ICAO (2017). www2023.icao.int [secondary]
  21. Google SRE book, Embracing Risk. sre.google [primary]
  22. Kemeny Commission report on Three Mile Island, 28 Mar 1979 chapter. pddoc.com [primary]
  23. US SEC press release 2013-222 on Knight Capital. sec.gov [primary]
  24. Google SRE book, Monitoring Distributed Systems. sre.google [primary]
  25. TechCrunch on robotaxi firms declining to disclose remote-help rates (31 Mar 2026). techcrunch.com [secondary]
  26. Greenblatt et al., AI Control: Improving Safety Despite Intentional Subversion. arxiv.org [primary, abstract]
  27. Engels et al., Scaling Laws for Scalable Oversight. arxiv.org [secondary]
  28. EU AI Act, Article 14, human oversight. artificialintelligenceact.eu [primary; application dates since deferred, check the Official Journal]
  29. Hospitality Upgrade on Mews and Atomize autopilot use. hospitalityupgrade.com [self-reported by the vendor]
  30. Finextra on Klarna rehiring human agents (May 2025). finextra.com [secondary]
  31. Hotel Dive on the Marriott and Sonder court filing. hoteldive.com [secondary]
  32. Zheng and Meister, The Unbearable Slowness of Being. arxiv.org [secondary]
  33. Weiser and Brown, Designing Calm Technology. calmtech.com [primary]
  34. Queensland Health research record on pulse-oximeter tones. dora.health.qld.gov.au [secondary, search summary only]
  35. IDC, Smart glasses surge (15 Jun 2026). idc.com [secondary]
  36. Wendy Mackay, Is Paper Safer? (1999), and the FAA electronic flight strips programme (https://www.faa.gov/air_traffic/technology/tfdm/efs). lri.fr [secondary]
  37. Piers Thackray, LinkedIn profile. linkedin.com [self-reported; no independent verification found]
  38. METR, Time Horizon 1.1 (29 Jan 2026). metr.org [primary]
  39. Anthropic, Measuring AI agent autonomy in practice (18 Feb 2026). anthropic.com [primary]
  40. American Bar Association on Moffatt v. Air Canada (Feb 2024). americanbar.org [secondary]