Why AI procurement is where the real fight is

Three ways to see it

1. 1

   Pakistani public sector AI procurement in 2026 has four structural problems. The first is vendor concentration: eighty percent of significant deployments today are powered by three or four foreign clouds and model providers. The second is sunk cost: once your data is in a foreign cloud, switching costs grow exponentially. The third is sovereignty: the contracts often include clauses where data, models, and even configuration choices effectively leave Pakistani jurisdiction. The fourth is capacity: the procurement officer in BPS-19 was trained on stationery and vehicles; she has not been trained on tokens, GPUs, and model lifecycles.

2. 2

   Way one to take procurement seriously: treat the RFP as policy in miniature. Every clause is a future right. Data residency is the right to keep records on Pakistani soil. Exit obligations are the right to leave. Source code escrow is the right to read what runs on your servers. SLA penalties are the right to recoup. Procurement officers who think of themselves as policy authors do work that lasts five years. Officers who copy boilerplate do work that lasts five hours.

3. 3

   Way two: separate the buy from the build decision. Most Pakistani procurement starts with a vendor pitch. Better practice is to first decide, for each AI capability, whether the country or organisation should buy, build, or open-source. Buy makes sense for commodity capabilities. Build makes sense for capabilities that are differentiating and strategic. Open-source plus local engineering makes sense when sovereignty matters and the talent exists. The MoITT NAIAI tender is in part a test of whether Pakistan can make this trifurcation on a serious scale.

The why-tree

Try this with Claude

AI-edge prompt to try: 'You are a Pakistani federal procurement adviser. List ten contract clauses that should appear in every public-sector AI tender of value over PKR 500 million. For each clause, write the one-sentence rationale and one example wording.' Use the output as a starter pack, then localise.