Hospitality Operations · B2B · AI Systems
A front-desk console for the agentic era. Agents handle the robot work, people own the relationships.
research-driven proof of conceptA front desk handles two very different kinds of work. High-volume and low-emotion: towels, Wi-Fi passwords, the same checkout question asked forty times a day. And low-volume, high-emotion: a midnight arrival, an upset guest, a regular coming back to a place they think of as familiar. Today both land on the same person, in the same queue.
Seven interviews across three hotels kept circling one thing. Staff found meaning in connecting with guests, not in completing transactions, and monotony was what pushed them out of hospitality. So the question was never how much human labor could be automated away. It was which work should disappear, so people have capacity for the work only they can do.
Hearth is a console for managing AI agents inside hotel operations. Agents absorb the transactions. Humans receive the moments that can become relationships. A hearth is the warm center of a home, which is the feeling hospitality is ultimately trying to create.
Solo, from research through prototype. Hearth began as a Stanford needfinding project in fall 2025 and grew into a full interaction prototype by summer 2026. I ran the interviews, set the thesis, defined the product model, designed the interface, and built the frontend.
Tier 1
Routine, low-risk work. A Wi-Fi question, a request for towels. The agent simply handles it.
Tier 2
Small, reversible actions the agent takes on its own, with the result left visible to staff. Independent, but never invisible.
Tier 3
Emotion, ambiguity, meaningful spending, safety, or opportunity. These go to a human, every time.
Something has gone wrong, and a person needs to own the repair.
Something could go especially right: an anniversary, a returning guest, a shared language far from home. Escalation should not only fire when the system fails.
What needs a person lives on the left. Proof that everything else is handled sits in a quieter rail on the right. When a task turns ambiguous or emotional, it does not stay put and grow a red badge. It moves into Needs You. That gives a tired employee one reliable place to look, so the agent rail can fill during a rush without making the interface louder. When nothing needs intervention, the board simply says: the desk is yours.
It carries the guest's name, the situation in one sentence, and the best suggested way to begin. The employee never has to make a guest repeat themselves, or read a transcript before they can respond.
One recommended opening rather than a menu. The source of every fact, and whether it is safe to say out loud. What the agent already finished. And a way to flag that this was not a moment at all, which corrects bad routing. No timers, no badges, nothing that gamifies a human interaction.
Managers have never configured a webhook, but they have spent years hiring and reviewing employees. So agents get names, shifts, permissions, spending authority, and reviews. A new one is onboarded like a new hire, held in a training state until it earns the next level of authority. They carry abstract sigils rather than faces: enough identity to manage clearly, not enough to mistake them for people.
Configuration means moving a responsibility between owns quietly, does and reports, and always a person. Keys and safety decisions stay visibly locked to humans, and goodwill spending is an open ledger rather than a hidden cap. It becomes a question operators already answer: what is this member of the team allowed to do in the house?
An agent's review reads like a performance review rather than a system log. What it handled, and where it should have stopped and asked.
Staff can reorder the agent's queue, and the interface labels the change as a human decision. Automation stays accountable instead of quietly absorbing authority.
A request is explained in four plain-language sentences instead of a technical event log, so nobody has to interpret a table of system events.
Condenses what actually matters about today's arrivals, and raises one warm signal for the guest most likely to benefit from a human connection.
Time returned to staff, requests absorbed, moments delivered. No employee leaderboard and no per-person tracking. The screen evaluates the agents, not the people.
Designed for a tired employee on a short-staffed Saturday midnight shift. Priority badges, countdowns, and red-yellow-green indicators were rejected on purpose: when everything is urgent, color stops meaning urgent. Hearth runs mostly monochrome and spends its one warm accent only on what needs a person.
Motion is minimal and the typography is editorial. Even room numbers follow the way staff say them out loud, so 4·12 reads as four-twelve. The interface is not trying to look futuristic. It is trying to stay understandable during the worst hour of the week.
Speech systems mishear people. A production version needs a visible uncertainty path that says, in effect, I did not understand this, so I am routing it to you. Agents should degrade to honesty rather than to confidence.
Holding a fallback room carries an operational cost the prototype does not surface. And the board holds four detailed Needs You cards before it would need another layer of triage, so a real deployment needs overflow behavior where recovery takes priority.
Two rules anchor the system.
No synthetic empathy. An agent can acknowledge a request, but it should not perform emotional ownership of a genuine service failure. When the hotel needs to apologize, a person owns the apology.
Keys always require a person. Identity verification is a security boundary, not administrative busywork. An agent cannot issue access simply because doing so would make the workflow faster.
Hearth changed how I think about agentic products. The most important design problem is not making the agent appear intelligent. It is making its authority visible.
People need to understand what the system is doing, why it made a decision, what it is permitted to do, where its information came from, and when it will stop and ask for help.
The best automation does not erase the human from hospitality. It creates more room for hospitality to be human.
Hearth is my B2B and systems-design case study, an exploration of how agentic technology can be made legible and trustworthy to a nontechnical operator. It is not production software. There is no backend, no live property-management integration, and no functioning phone system. Agent actions are scripted, because the objective was to design and test the interaction model.
I used Claude and other AI-assisted tools during implementation, but the research, information architecture, product boundaries, visual system, and interaction decisions were my own.