AI pricing models and PKR projections

Three ways to see it

1. 1

   There are five common AI pricing models. Per seat: a fixed monthly fee per licensed user, common for copilots and productivity AI. Per transaction: a fee per call or per resolved case, common for chatbots and verification. Per token: a fee per thousand input or output tokens, the standard for direct LLM API access. Per API call: a flat fee per request, common for managed AI services. Outcome-based: percentage of value created or fraud recovered. Each fits some use cases and ruins others.

2. 2

   Way one to model PKR cost: build a three-year forecast with three demand scenarios. Low, expected, high. For an Urdu-English chatbot at a federal complaint cell, assume 10,000 daily conversations expected, 30,000 high, 3,000 low. At a typical 2,000 tokens per conversation and a frontier model at PKR 0.5 per 1,000 input + PKR 1.5 per 1,000 output, expected daily cost is PKR 60,000, monthly PKR 1.8 million, annual PKR 22 million. The high scenario triples that. Now compare against per-seat or per-transaction quotes; the picture flips multiple times.

3. 3

   Way two: hunt for the hidden costs. Egress fees when you move data out of a cloud. Premium support beyond business hours. Fine-tuning compute on top of base licensing. Data labelling fees. Compliance attestations. Penalty waivers. Each of these can be 10 to 30 percent of the visible contract. Ask the vendor to itemise them up front; ask twice if they say there are none.

The why-tree

Try this with Claude

AI-edge prompt to try: 'Acting as a Pakistani CFO, build a three-year PKR cost model for the AI chatbot I describe (paste use case). Assume PKR depreciation of 8 percent annually. Show low, expected, high demand scenarios. Output as a markdown table.' Use it as a draft for negotiation.