Start with the problem, not the tool

Three ways to see it

1. 1

   A worked example. The tax-policy wing at FBR has eight officers who summarise tax returns flagged by the automated risk engine. Each return is twenty pages of mixed Urdu and English, with attachments. An officer spends three hours per return; each officer handles two per day; the wing closes sixteen returns daily. That is the activity. Now the question is whether AI fits. The clue is in the verb: summarise. Summarisation is the single task LLMs do well. The task fits. The numbers fit. Even a 50% time saving across eight officers is four hours per officer per day, the equivalent of a full new headcount the wing cannot otherwise hire.

2. 2

   The five-question filter. Run any candidate problem through these before committing time. One: what is the input and what is the output, in concrete artefacts, not adjectives? (A twenty-page PDF in, a two-page brief out.) Two: how often does the work happen, and what is the unit cost? (Sixteen times a day, three person-hours each.) Three: is there a measurable definition of done? (The auditor reviewing the brief in fifteen minutes can decide audit or no-audit.) Four: who currently does this work, and what would they do with the freed time? (Eight officers, freed up for deeper investigations.) Five: what is the cost of an AI error? (A briefer that misses a fraud signal causes a missed audit; a briefer that flags a non-fraud causes one extra review.) If any of these answers is hand-wavy, refine the problem before touching tools.

3. 3

   Three Pakistani problem archetypes that fit AI well. Document summarisation: tax returns, board minutes, court orders, audit reports. Language translation and bridging: Urdu citizen complaints into structured English tickets. Drafting under constraints: regulatory letters, internal memos, vendor responses where the tone matters and the facts come from a small reliable source. Three archetypes that fit poorly today: anything requiring guaranteed legal advice; anything where a single wrong answer carries criminal liability; anything where the underlying data is too sparse or too sensitive to leave your perimeter. Knowing which archetype your problem belongs to does half the work of choosing the right tool.

The why-tree

Try this with Claude

AI-edge prompt to try with Claude or ChatGPT: 'I run operations at a Lahore textile exporter. Brainstorm ten recurring activities in my team that take more than eight person-hours per week. For each, mark whether the activity is a strong, medium, or weak fit for an LLM today, with one Urdu sentence of justification. Then rank them by ROI in person-hours saved if AI cut them by half.' Use as a starting list, not a final one.