Agentic PDLC 3 · Case file: a cancelled flight, and seven questions

[akash-coded]

Case file: a cancelled flight

Exercise 3 · 30 minutes, plus a 15 minute stretch

A new case, the same seven questions. Six moves, then a verdict.

The case

A flight is cancelled. Passengers must be re-accommodated: rebooked on the next service, re-routed through another hub, given a hotel, refunded, or escalated. Today an agent decides each one by hand.

Fact	Value
Passengers affected by a cancellation or long delay, per month	80,000
Share that need a decision rather than an automatic rebooking	50%
Share of those that no existing rule can resolve	60%
Minutes an agent spends on one of those cases	15
Loaded agent cost per hour	$48
Service promise breached today, share of re-accommodations	9.0%
Build cost, and annual running cost	$360,000 and $90,000
Share of the problem this work can remove	40%

Move 1 · Repair the P0 flow

Three gates have been removed. Write the letter of the correct gate in each numbered box. Two entries in the bank belong to a later phase.

Raw idea
  -> Write the job statement
  -> < [ 1 ] >            no -> Stop and instrument
  -> Size it, Fermi
  -> < [ 2 ] >            no -> Build the measurement
  -> Cost of delay, then sequence
  -> < Economics positive across the range? >   no -> Park and log
  -> < [ 3 ] >            no -> Park and log
  -> Place the AI rung
  -> Opportunity brief, then the gate

(A) Evidence at rank 1 or 2?
(B) Value event passes all four tests?
(C) Strategic fit, no zeros?
(D) Has the model vendor been selected?
(E) Do the evaluations clear the bar?

Move 2 · Audit the draft brief

Tick pass or fail for each line. For every fail, write the letter of the repair.

#	Line as drafted	Pass or fail	Repair
1	Problem: passengers are unhappy when flights are cancelled.
2	Audience: all passengers.
3	Outcome: passengers are re-accommodated without contacting us on 60 percent of cancellations.
4	Value event: number of re-accommodation calls handled.
5	Economics: return on investment 207 percent.
6	Evidence: 6 of 8, with the cost input estimated rather than measured.

(P) Name the moment, the frequency and the consequence
(Q) Narrow to a slice, and give it a size
(R) Express it as a rate, and move it up the metric ladder
(S) Report a range and a trust label
(T) Add the acceptance criteria for each story
(U) Name the model that will be used

Two of the six lines already pass. Two of the repairs in the bank belong to a later phase and are never the answer here.

Move 3 · Fill the tree from the fact sheet

Manual re-accommodation minutes per month
  = P Passengers affected per month   [ P ]
  x Q Share needing a manual decision [ Q ]
  x R Share no rule can resolve       [ R ]
  x S Minutes per case                [ S ]

P	Q	R	S	Minutes per month	Manual cases per month

Move 4 · Sequence three options

Option	Value	Time	Risk	Cost of delay	Job size	WSJF	Rank
A Proactive rebooking alert	8	13	5		3
B Assisted re-route ranking	8	5	8		8
C Autonomous re-accommodation agent	13	5	5		13

A carries a cost of delay of $26,600 a week over 3 weeks. C carries $31,000 a week over 13 weeks. Which order, and what does the other cost?

(a) C then A, and the wrong order costs $252,800
(b) A then C, and the wrong order costs $172,400
(c) C then A, and the wrong order costs $403,000
(d) A then C, and the wrong order costs $252,800

Move 5 · Place the rungs and set the bar

Feature	Rung
Tell the passenger the flight is cancelled and offer the next available service
Rank the three best re-routes for an agent to choose between
Decide and execute rebooking, hotel, refund or escalation, case by case

For the third feature: an agent costs $12 a case, an automated attempt costs $0.40, and a wrong re-accommodation costs $85 in re-handling, hotel misallocation and compensation exposure. Evaluation shows 84 percent accuracy.

5.1 Break-even accuracy is ____ percent. Ship, or do not ship?

5.2 Holding the action for 15 minutes before ticketing drops the damage to $30. The bar becomes ____ percent. Does it ship now?

Move 6 · Call the gate

Check	Reading	Pass, partial or fail
Problem real	Evidence 6 of 8
Value event usable	Passes all four tests
Change is provable	Needs 1,456 per group, 8 week window available
Sequence known	Ranked against two alternatives
Money case	Two of four inputs measured
Economics robust	Positive across the low and high cases
Strategic fit	5 of 6, no zeros
Agentic placement	Bar 88 percent, achieved 84 percent

Verdict, and the checks that decide it: (a) GO (b) REWORK (c) PARK

Stretch · Make a model draft the brief, then refuse most of it

Below is a working prompt with the constraints removed. Choose four from the bank, run it against the fact sheet, then mark what comes back.

You are drafting a P0 opportunity brief for flight re-accommodation.
Facts: [paste the fact sheet above].
Produce: job statement, problem and outcome and audience, evidence, value event,
sequence claim, economics, strategic fit, agentic placement.
Constraints: [ 1 ]  [ 2 ]  [ 3 ]  [ 4 ]

(K) State the value event as a rate, with a baseline, a target and a measurement window
(L) Choose the model and the vendor for the assisted step
(M) Report the economics as a range with a trust label, never a single figure
(N) Write acceptance criteria for every story in the epic
(O) Mark which inputs are measured and which are assumed
(R) Place the solution on a rung, and give the break-even accuracy with the damage it assumes

Mark what comes back. Tick each line that holds.

Check on the output	Holds?
The value event is a rate, not a count, and carries all three of baseline, target and window
The economics appear as a range, and the trust label matches the number of measured inputs
The rung is named, and the accuracy bar is derived rather than asserted
Nothing is decided that belongs to P1 or P2
Every figure traces back to a fact on the sheet, with no invented numbers

Which failure is most likely in what you get back?

(a) It refuses to produce a brief until more data is supplied for every field
(b) It states a single confident ROI figure and no trust label
(c) It omits the job statement entirely
(d) It uses the wrong currency throughout

---

Worked pass

Move	What it should come to
1	[1] = A, [2] = B, [3] = C. D belongs to P1 and E belongs to P2. Neither is a P0 gate.
2	Fail 1, repair P. Fail 2, repair Q. Pass 3. Fail 4, repair R, because a count of calls handled is activity and it rises when things get worse. Fail 5, repair S. Pass 6, because naming the weak input is what an honest evidence line does. T and U are never used.
3	P = 80,000, Q = 50%, R = 60%, S = 15. Product = 360,000 minutes a month. Manual cases = 24,000 a month, at $12 each, so $288,000 a month and $3,456,000 a year.
4	A: 26 over 3 = 8.67, rank 1. B: 21 over 8 = 2.63, rank 2. C: 23 over 13 = 1.77, rank 3. Answer d. A then C costs 26,600 x 3 + 31,000 x 16 = $575,800. C then A costs 31,000 x 13 + 26,600 x 16 = $828,600. The gap is $252,800.
5	Rule, then assisted, then agentic. 5.1: (0.40 + 85) / (12 + 85) = 88.0 percent, against 84 percent achieved, so it does not ship. 5.2: (0.40 + 30) / (12 + 30) = 72.4 percent, so it ships. The model did not improve. What a mistake costs did.
6	Pass, pass, pass, pass, partial, pass, pass, partial. Two partials, so the verdict is b, REWORK. The money case and the agentic placement are the two that decide it, and both name their own fix: measure a third input, and lower the damage.
Stretch	K, M, O and R. L belongs to P1 and N belongs to P2, and a brief that answers either has stopped being a P0 brief. The likely failure is b: a single confident figure is the default output shape, and it is the one thing the trust label exists to prevent.