From a gallery to a living knowledge centre
The project began as a proof gallery showing what Cowork produced on real tasks and how many credits each run consumed. As the product evolved, it became clear that users needed more than examples. They also needed grounded answers to what Cowork is, how to access it, where it fits, how to control its cost, and how it differs from Copilot.
The product now connects three layers:
- Current, sourced guidance for Microsoft Copilot and Cowork.
- Measured Cowork studies built from real task runs.
- Repeatable Copilot recipes with video, copyable prompts, and downloadable outputs.
This closes the gap between product marketing and practice. Visitors can understand the operating model first, then open a study close to their role and inspect the chain of data, process, output, and cost.
Two recipe libraries
Cowork Recipes
The 19 published studies range from financial close and revenue recovery to human-agent work design, organisational learning, and agent evaluation infrastructure. Every study follows the same evidence contract:
- clearly labelled synthetic Turkish data,
- task setup and human approval gates,
- the steps Cowork actually completed,
- downloadable files and decision outputs,
- credit consumption measured through
/cost, - sourced human-effort assumptions and comparable return,
- limitations, failures, and an honest assessment.
Copilot Recipes
The 11 published recipes explain practical Microsoft 365 Copilot workflows through short videos, copyable prompts, and downloadable sample outputs. The first collection focuses on Excel, while the “Copilot Memlekette” series applies the same repeatable format to recognisably Turkish scenarios such as holiday planning, football-fixture analysis, and household economics.
The product guide
The homepage now does more than route visitors into case studies. It explains the progression from chat and reasoning to agent harnesses and Autopilots, locating Cowork within that wider shift.
The guidance layer also answers operational questions with primary sources:
- the Microsoft 365 Copilot licence and admin configuration required for access,
- credit cost across model use, organisational context, tool calls, and run time,
- planning ranges for light, medium, and intensive tasks,
- Work IQ, background execution, human approval, and the enterprise trust boundary,
- new models, security changes, and product updates.
Flagship study: revenue loss and customer recovery
In one of the most substantial runs, Cowork joined 253,500 related records from 13 files at account level. It prioritised 492 high-risk accounts out of 5,000 and surfaced TRY 1.66 billion in expected revenue exposure.
One task produced an Excel workbook, executive presentation, Word and PDF reports, an offline HTML dashboard, a communication pack for the top 25 accounts, a war-room brief, and a reusable skill-quality report. The run consumed 3,344.2 credits, while three human approval gates ensured that no customer communication was sent automatically.
The study captures the project's central principle: do not publish a headline outcome alone. Publish the data, decisions, deliverables, cost, and limits that made the outcome possible.
Evidence and measurement
Real task + labelled synthetic data
→ human-supervised Cowork run
→ downloadable deliverables
→ measured credit consumption
→ sourced human-effort baseline
→ recalculable cost and return
Credit cost is converted into TRY using the daily Central Bank of Türkiye USD/TRY rate. Human cost is derived from sources shown for the relevant role and period. Visitors can change the time and hourly-cost assumptions to recalculate the comparison for their own context.
The system
| Layer | Implementation |
|---|---|
| Experience | Fast, static, Turkish knowledge centre built with Astro |
| Content | Product guidance, Copilot recipes, Cowork studies, and an update log |
| Evidence | Synthetic data, screenshots, real outputs, and human approval trails |
| Measurement | TypeScript cost engine, /cost data, and daily exchange rates |
| Production | Azure AI Foundry-assisted content pipeline with a critic quality loop |
| Automation | Human-supervised Playwright runs and publishing checks |
| Delivery | Azure Static Web Apps, Functions, Storage, and Bicep IaC |
Scope and boundaries
This is an independent personal project, not an official Microsoft property. No real customer data is used in the studies. Export files are not described as live system connections, estimates remain distinct from measurements, and no external action is presented as completed without human approval.
Because product capabilities and pricing can change, the guidance is maintained alongside primary-source links and an update log. The goal is not a flawless success gallery, but an inspectable workspace showing what another team could reproduce, under which conditions, and where the approach stops.
My role
I own the product end to end, from the concept and information architecture through research, scenario design, synthetic data, measurement, content and automation, the Astro frontend, Azure infrastructure, Turkish quality gates, and production release.