Shahzad Ali · Learning notes
Work has changed.
How AI changes the way I move from a question to a result.
I use AI to help gather information, analyse options and prepare drafts. My attention goes into setting the direction, judging the evidence and deciding what happens next.
The work changes.
I own the outcome.
How I organise a research task, from the first question to the final decision.
Define
Before AISet the question and requirements.
With AISet the question, context and requirements.
Gather
Before AISearch, read and organise the material.
With AIAsk AI to gather material with sources.
Develop
Before AIAnalyse options and prepare a draft.
With AIDirect the analysis and refine an AI draft.
Review
Before AICheck evidence and revise the work.
With AICheck evidence, assumptions and the draft.
Decide
Before AIChoose the next action.
With AIChoose the next action.
Set the direction→Judge the evidence→Choose what happens next
My perspective
The goal still needs an owner
I can ask AI to search, organise, draft and check. I still need to decide what matters and whether the result serves the problem I set out to solve.
For me, the interesting shift is spending more attention on questions, judgement, relationships and decisions.
How delegated work happensJudgement
Keep subject knowledge in the loop
Understanding the work helps me recognise missing information, weak assumptions and fluent mistakes. It also helps me give a useful brief.
Subject knowledge helps me give a clear brief and recognise weak assumptions.
How to judge the resultMeasured evidence
Productivity is specific to the task
In a 2023 randomised experiment with 453 college-educated professionals, ChatGPT reduced average completion time by 40% and increased assessed quality by 18% on selected professional writing tasks.
The participants completed selected professional writing tasks. The effect on another kind of work needs its own measurement.
What an evaluation measuresA contrasting study
Tools can add work as well
A 2025 randomised study of 16 experienced developers working on 246 real tasks found that AI access increased completion time by 19% in that setting.
The tools and tasks were from early 2025. METR’s February 2026 follow-up describes selection and measurement problems that made its newer estimate unreliable. These findings cannot settle today’s results for every workflow.
Evaluating the actual taskMy practical takeaway
Measure the work I actually care about
For repeated tasks, I track time, accepted output, quality, cost and rework.
That tells me where AI is helping and where I need to change the process.
From a claim to a test
Sources behind the technical details
Sources for these notes.
Reviewed . Technical details are linked to their sources. My perspective on how I use AI is personal.
- Developer engineering account · First released
Building effective agents ↗
Anthropic. Engineering patterns for workflows, autonomous loops and orchestrator-worker systems.
- Primary research · First released
Experimental evidence on the productivity effects of generative artificial intelligence ↗
Noy and Zhang. Randomised professional-writing experiment. The publisher abstract reports the sample and main results.
- Research preprint · First released
Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity ↗
Becker and colleagues, METR. Randomised access to AI across 246 tasks. Results concern experienced developers and early-2025 tools.
- Researchers’ follow-up · First released
Update on our developer-productivity research ↗
METR. Explains selection effects and measurement difficulties in the follow-up study.