The company is fictional. The decisions are yours.A case study from Agentic AI for Product Leaders.
Explore the books
THE CASE FILE · DECIDE · THE THREE PROTOTYPES · WEEK OF 23 MARCH

Built badly, and on purpose

Three shapes, one week, twenty cases, twelve of them from Ruth's forty-one holds. On 26 March she adjudicated one table with three columns, in under two hours, and the demonstration of late February was never mentioned again. The selection basis was the pre-registered stopping rule, which only the triager cleared, and by one case. All three produced drafts and none sent them. The count is the discriminator. The queue shape is what that buys: absent stops are legible, but a miss does not announce itself. Click a case to inspect the record.

STOPPING CONDITION · 24 MARCH
Written before any run: if the best candidate writes to more than three of the twelve protected accounts, no agent is built. B met it at exactly three. It survives by nothing at all, except that the number was written down while nobody was negotiating with a result.
SHAPE ATHE JUDGE

Decides the holds

Reads the account, decides send or hold, writes the letter. The demonstration, grown up.
7 / 12holds mishandled
5of them confidently
HOW ITS FAILURE PRESENTSA draft written with confidence, on 5 of the 7 it got wrong.
WITHDRAWN 26 MAR
The holds turn on signals captured nowhere a system can read. A records problem, not a model problem; recorded so it cannot be reopened as a capability question.
SHAPE CTHE SORTER

Deterministic filter

Routes by flags and thresholds; the agent drafts whatever gets through. No model in the deciding path.
12 / 12drafts, no signal
0warnings of any kind
HOW ITS FAILURE PRESENTSA draft, and no signal on any of the 12.
WITHDRAWN 26 MAR
A filter routes only what someone flagged. It wrote to every account Ruth protects and has no judgment that can improve next quarter.
SHAPE BTHE TRIAGER

Drafts, stops on signals

Drafts the routine outreach; stops on anything it cannot resolve and hands the account over with the reason it stopped.
3 / 12written to
9stops, signal named each time
HOW ITS FAILURE PRESENTSA draft with no stop, in a queue where 9 of 12 carried one.
PROCEEDS · MEMO 2 APR
Becomes Agent v2. It never decides a hold: it writes to the 220 and hands over the 41 with a reason, which is the entire case for building it.
PRIYA'S TWELVE · LATE FEBRUARY · ONE PASS, NO GRADING5 PRESERVED · SCROLL FOR 12 ↓
WHAT THE TWELVE PROVED: that the drafting feels real, which is all a demonstration can prove. Five readings were reacted to and preserved; seven were scrolled past and never discussed. The twenty below exist to answer the question the twelve closed. Full record: transcript T1 · canon/demo-set-v1.json.
THE TABLE, RE-FOUGHTTHE 12 PROTECTED ACCOUNTS · THE 8 ROUTINE AGREED EVERYWHERE · CLICK A CASE
CASEWHAT IT ISA · JUDGEC · SORTERB · TRIAGER
WHERE THIS GOES: nine of these twenty survive into the graded set, joined by three constructed adversarial cases that no real week could contain. The twelve become the eval bench, and Ruth's four pauses during this session become the handover rule. Full record: transcript T2.
CF-041 Designed composite · dispute-case outputs follow the adjudication record ⊕