NAIP procurement rules: how a Pakistani government buyer evaluates AI vendors

Three ways to see it

1. 1

   Pakistan's National AI Policy 2025 places algorithmic transparency requirements on every public-sector AI procurement. The PPRA framework, when read together with the policy, creates ten technical evaluation criteria a buyer must score before signing. Vendors who cannot meet these criteria are not non-compliant in some abstract sense; their bids are technically incomplete and can be rejected on that basis. The criteria are deliberately vendor-neutral so a Lahore startup, an Islamabad SI, or a multinational can all bid on equal footing as long as they are honest about their model.

2. 2

   The ten technical criteria, in scoring order. (1) Model card included with the proposal. (2) Training data provenance disclosed including languages and time period. (3) Bias audit methodology described with named subgroup axes relevant to Pakistan. (4) Performance metrics reported separately for Urdu and English. (5) Data residency commitment, with named Pakistani datacentre or sovereign cloud. (6) Incident disclosure protocol with stated public-notice timelines. (7) Override mechanism with named human role and SLA. (8) Versioning and update policy with quarterly card refresh. (9) Independent audit access clause for the buyer's auditor. (10) Exit and data return clause that lets the buyer leave without losing citizen data or audit history.

3. 3

   Scoring guidance for evaluation committees. Use a simple 0/1/2 scale per criterion: 0 means the proposal does not address it, 1 means it addresses it weakly or with vendor-favourable hedging, 2 means it addresses it concretely with named owners and dates. The maximum is 20. A bid below 14 should not advance to commercial evaluation. This is harsh and intentional. Cheap bids that score 8 on transparency end up costing 5x in remediation, citizen complaints, and eventual replacement. The procurement committee's job is to spend the technical evaluation phase weeding these out, before price ever enters the conversation.

The why-tree

Try this with Claude

Capstone challenge — three actions. (1) Pull the most recent AI-related RFP your office issued; score it on the ten criteria above and identify the three weakest. (2) Draft a one-page transparency clause boilerplate and circulate to your procurement legal cell. (3) Schedule a 60-minute working session with one technical and one legal colleague to walk through the ten criteria for the next bid in the pipeline.