Operating model · internal

How Quell runs.

42 humans. 1,200 agents. One shared inbox.

01 · Headcount

The whole company fits in one room.

Engineering
18
Field Ops
6
Product
5
GTM
7
Design
3
Founders + Ops
3
Total humans
42
Revenue per employee
€2.1M

Every Quell employee generates €2.1M in annual revenue. We don't scale by hiring — we scale by shipping agents.

For reference: Booking.com carries roughly 27,000 employees across a stack that still routes search, ranks inventory by hand, and runs loyalty as a separate org. We run the same workflow with 42 people because the agents were there from day one — not bolted on.
02 · The agent stack

Six core agents. Every trip makes the next one smarter.

Reality-check agent
001

Scans the last 30 days of reviews, photo-vs-listing drift, and cancellation patterns on every booked property. Flags when lived experience diverges from the pitch.

Swap agent
002

Holds a ranked Plan B within 2km of every booked stay, pre-vetted against the traveler's memory. One tap ships it. Median swap-to-confirm: 12 seconds.

Refund agent
003

Drafts the claim against the failed property, runs the back-and-forth, escalates to the card network after 72 hours of silence. Zero emails from the traveler.

Memory agent
004

Stitches preferences and past failures across trips — no early flights, quiet floors, accessible terraces — and silently filters every future recommendation.

Disruption agent
005

Watches flights, weather, strikes, local events against every active itinerary. Fires the first warning before the airline does; hands to the swap agent.

Follow-up agent
006

Day-after debrief — "how was it, what broke?" — and writes findings straight to the memory agent. Closes the loop. Every trip trains the next.

Pricing agent
007

Rescores inventory every 90 seconds against live signals. No human curator, no category manager. Taxonomy is emergent.

Comms agent
008

Writes the 2am text to the traveler and the 2:01am text to their partner, in the tone each prefers. No drafted templates; every message is written fresh.

Eval agent
009

Runs the agent-on-agent eval suite continuously. If any agent regresses against last week's failure cases, traffic is rolled back automatically before users see it.

03 · How we charge

You pay when the trip is delivered. €0 on failure.

Event Charge On failure
Traveler browses, plans, asks questions €0 — no lookup fee, ever
Trip booked and delivered as pitched €47 / trip
Disruption resolved by agent (one-tap swap) €18 / resolution €0 if the traveler isn't satisfied
Refund recovered from a failed property 12% of recovered €0 if we recover nothing
Inventory commission, pay-for-placement, loyalty skim €0 — never, by design

We make money when the traveler gets what they paid for, or when we get their money back. Those are the only two events. No ads, no placement, no commission floor.

04 · Unit economics

The hero number.

Hero metric · live
87%

of trips close without a human touching them.

That's the one number every team at Quell watches. When it drops below 83%, we roll back the latest agent release. When it holds, we ship.

Revenue per completed trip
€64
trip fee + avg. resolution + recovered-refund cut
Gross margin
81%
inference + comms + refund workflow costs
CAC payback
4.2 mo
second trip triggers organic recovery
05 · What we don't have

The teams we skipped on purpose.

Not at Quell
Search, SEO & affiliate
~1,800 people at Booking.com
Replaced by

There is no SERP in Quell. You name the trip; the reality-check agent returns one vetted option already matched to your memory. No list to rank, no ads to place, no pay-for-placement. The interface isn't a search page anymore — so the org disappears.

Not at Quell
Loyalty
~200 people at Booking.com
Replaced by

The memory agent is the loyalty program. It remembers you hate early flights, need accessible terraces, and why the last anniversary property failed. That's worth more than any point balance. Loyalty is a byproduct of the product working, not a separate team running points and tiers and win-back campaigns.

Not at Quell
Category managers & content ops
~600 people at Booking.com
Replaced by

The pricing agent rescores 2M properties every 90 seconds against live reviews, photo-vs-listing drift, and cancellation patterns. Taxonomy is emergent, not edited. Zero manual curation queue, zero SKU meetings, zero editorial backlog.

Not at Quell
Customer service
~10,000 people at Booking.com
Replaced by

The refund agent handles disputes end-to-end in 10 languages. When a human judgment is required, the Field Ops pair (6 people total) reviews in under 3 minutes. We don't have a call center. We don't have a ticket queue. We have eval scores.

06 · How a decision flows

One real night. No human in the loop.

A real Cappadocia stay goes wrong at 11pm. Here's every action Quell takes, minute by minute, with the traveler consenting once.

  1. 23:04 · signal
    Reality-check agent flags the stay.

    Rooftop terrace closed 8 of last 10 nights · front-desk review sentiment dropped 41 points in 72 hours · both match traveler's memory-flagged failure modes.

  2. 23:04 · agent
    Swap agent drafts three options.

    Within 2km · pre-vetted against memory · ranked by match score. Top option: Sakli Konak Cave House (92% match).

  3. 23:05 · agent
    Pricing agent confirms availability and holds the room.

    90-second rescore against live inventory. Room held for 10 minutes pending traveler tap.

  4. 23:05 · human · ONE TAP
    Traveler taps "Yes. Move us."

    The only human action in the entire flow. Consent gate preserved — the agent proposes, the human confirms.

  5. 23:06 · agent · auto-execute
    Swap executes. Refund agent files claim.

    New room confirmed · checkout modified on old property · comms agent texts traveler + partner · refund claim queued against failed hotel.

  6. + 3 days
    Refund recovered. Memory updated.

    €612 returned without a traveler email. Follow-up agent writes the failure pattern to memory. This chain will not surface for this traveler again.

One human tap. Six agents. Zero Field Ops involvement.
Above €500 or when a trip memory is inconsistent, the Field Ops pair reviews before the tap is offered. That accounts for 13% of flows. The other 87% close on their own.