48 hours Hackathon 2026 · AI agent

“Is this me, or the market?”

An RM AI agent that tells hosts why bookings are slow, with a deterministic verdict, a confidence score and the one lever to pull. Planned, specified and handed to engineering in two days.

Role
Designer · planning, spec & response UX
Team
Designer, RM, 2 SCs, engineers
Timeline
2 days
Output
Verdict logic spec v2.1
H

Are my bookings slow for October? Is it me or the market?

Mixed Market down 6 pts, you’re down 19; the 13-pt gap is yours.
  • 3 of 5 comps are already above 65% occupancy for October.
  • Your far-out premium is pricing you out before the booking window opens.
Lever: DSO −8% on Oct 10–20 · relax min-stay on weekdays
Medium confidence5 well-matched compsShow the numbers ▾
Part 1

The case study

Two days, summarised.

Problem

A slump has many causes. A naive “me vs market” answer is wrong about a third of the time, and support can’t forward guesswork to customers.

What I did

Planned the two days, synthesised an RM’s and two consultants’ logic, and specified a 12-stage pipeline, 8 verdict states and the response card.

Outcome

A build-ready spec in 2 days, with 7 open decisions and their owners surfaced up front. The spec became the basis for the RM AI agent on the product roadmap.

01 · Problem

“Me or market” isn’t a yes/no question, and an LLM that decides it freely gives a different answer every time.

Many causes
  • Market softness, settings, price, timing and data quality all look alike in a single occupancy number
Consistency
  • The eval required the same result 4–5 times in a row, which a free-form LLM verdict fails
Customer-safe
  • SCs forward the output, so no competitor identities and no guesswork

02 · Two days, planned backwards

With two days, the plan started from the output (a build-ready spec) and worked back.

Day 1 · Understand & frameSessions with an RM and two SCs; study the shipped Account Review classifier; agree design principles; draft the 12-stage pipeline
Day 2 · Specify & hand offVerdict states, validity gates, threshold block and response card; 10 worked cases; open decisions with owners; handoff brief
  • Reuse, don’t rebuildThe shipped Account Review classifier became the core check, verbatim. The agent is the layer before and after it.
  • Scope edge decided earlyMechanism diagnosis was marked as a stretch: drop it under time pressure and the agent is still complete.
  • Unknowns made visibleEvery blocking question got an owner, so nobody built on an assumption.

03 · Design principles

Rules decide, the LLM only narrates.

1Rules decide, the LLM narrates

Verdict and confidence come from fixed thresholds; the LLM turns them into customer-safe language.

2Eight states, not two

No Slump, Market, Mixed, Me, Winning, plus two overrides and an honest “Inconclusive”.

3Pace-matched, always

Compare on-the-books now to the same lead time last year, never to final results.

4Customer-safe by default

Output survives a paste into a customer email.

5Diagnose, don’t solve

Name the lever, not a solution engine.

04 · The pipeline

Twelve deterministic stages from question to card. About a third of questions exit early as “No Slump”.

0Resolve scope
1Pace-matched data
2Validity gates
3Validate the claim~⅓ exit: No Slump
4Core check
5Market overlay
6RevPAR overlay
7Booking-window veto
8Mechanismstretch
9Segment
10Confidence score
11Narrate (LLM)
Shipped helper, reusedNew RM logicOnly step the LLM touches

05 · Verdict states

Eight outcomes, so the agent can say “you’re fine” or “we can’t tell yet”.

No Slump

Pacing at or above last year and market. Return proof, stop.

Market

You fell about as far as the market. Wait it out.

Mixed

Both down, you worse. Split the gap into market and self-inflicted.

Me

Market held or rose, you fell. A listing-side gap.

Winning

You’re up, market’s down. Check for underpricing.

On Strategy

Looks like “Me”, but it’s a deliberate hold-high, sell-late pattern.

Late Not Lost

Pickup is healthy; demand is just arriving later.

Inconclusive

Data too weak to call. Say so honestly.

06 · Outcome

2 daysfrom interviews to handoff
12pipeline stages, 3 reused as-is
8verdict states instead of 2
7open decisions, each with an owner
Part 2

The full spec

Verdict Logic Spec v2.1, exactly as handed to the team.