Capstone: a one-page AI deployment memo for your own department

Three ways to see it

1. 1

   The deliverable for this capstone is a single page. One side of A4. No annexes. No diagrams. The audience is your own secretary, your own chief secretary, or your minister, depending on your grade. The memo proposes one concrete AI deployment your department should undertake in the next six months. It names the problem, the data, the model class, the procurement route, the risk register, the budget, the success metric, and the kill switch. Eight items, one page. The discipline is the point.

2. 2

   The eight items in order. (1) The problem in one sentence, naming the citizen or stakeholder affected. (2) The data: what already exists, who owns it, what classification, where it sits. (3) The model class: rule-based, retrieval-augmented, predictive, generative, or hybrid. Reject anything you cannot describe in plain Urdu to a deputy commissioner. (4) The procurement route: in-house build, panel vendor, open tender, donor-funded sandbox, or shared service from PITB/KPITB. (5) The risk register: top three risks with mitigations. At least one must be a citizen-harm risk, not a delivery risk. (6) The budget: one-time and recurring, in PKR, named under an existing development or non-development head. (7) The success metric: a single number, measurable within the pilot horizon, agreed in advance with the auditor. (8) The kill switch: the named threshold or date that triggers an honest shutdown.

3. 3

   Worked example, in summary. An additional director at the excise and taxation department of Sindh proposes an AI-assisted property-tax dispute triage pilot. Problem: 18,000 dispute applications per year, average 14 weeks to first response, 41 per cent abandoned. Data: existing PT-1 forms in the excise database since 2019, all in Urdu and English, no PII outside name and address. Model class: retrieval-augmented classifier with rule-based escalation; rejects anything generative for an audit-bearing record. Procurement: PITB shared services, no fresh tender. Risk register: wrongful auto-rejection of an elderly widow's case (mitigation: human-in-loop for all rejections), data leak through API (mitigation: VPN-only access), bias against rural addresses (mitigation: monthly accuracy audit by district). Budget: PKR 6.4 million one-time, PKR 1.1 million annual. Success metric: average time to first response reduced from 14 weeks to 6 weeks within nine months, measured against a 2026 baseline. Kill switch: if rural-address acceptance rate falls more than 10 points below urban-address rate for two consecutive months, the pilot is suspended and reviewed by the secretary.

The why-tree

Try this with Claude

Capstone challenge, three actions for this week. (1) Pick one real problem in your department where AI plausibly helps. Write the eight items in rough form on the back of a meeting agenda. (2) Walk the rough memo past your most sceptical BS-19 colleague. Note every objection without defending. (3) Refine the eight items in light of the objections and submit the one-page memo. The grade is for clarity, honesty, and survivability, not for ambition.