Turning frontier AI capabilities into customer-ready products
Product lead for Frontier Tuning, an AI capability within M365 Copilot — taking frontier model capability from research promise to something customers can adopt, evaluate and rely on.
AI task quality improvement with a strategic customer
customizable Copilot agent task templates launched
productivity scenarios in the RL-powered flagship portfolio
Context
Frontier model capability moves faster than product definition. A capability can be genuinely new and still be unusable, because nobody has said which customer workflow it serves, what better means for that workflow, or how anyone would know it worked.
I lead product for Frontier Tuning inside M365 Copilot, announced at Microsoft Build 2026: the work of connecting emerging model capability to real customer workflows, with evaluation and measurement built in from the start rather than bolted on before launch.
Visual model
- 01Customer workflow
- 02Evaluation criteria
- 03AI experimentation
- 04Product / model iteration
- 05Measurement
- 06Customer impact
↻ Customer impact feeds the next workflow — the loop is the product process
The product development loop I run for frontier AI capability.
How I work the problem
- 01
Start from a named customer workflow, not a capability demo — the workflow decides what quality means.
- 02
Define evaluation criteria with data science and research before experimentation starts, so results are comparable and arguable.
- 03
Run experimentation as a product loop: hypothesis, tuned iteration, measured delta, decision.
- 04
Partner closely with AI research and engineering so trade-offs are made once, explicitly, and in the open.
- 05
Ship customizable templates rather than bespoke one-offs, so the capability compounds across customers.
- 06
Design governance, compliance and regional data residency into the product and go-to-market plan, rather than treating enterprise readiness as launch cleanup.
How I approach AI products
Customer problems set the target. Model capability sets what is currently reachable. Evaluation methodology is the bridge between them — it converts a subjective sense that an answer is better into evidence a team can act on.
Experimentation then becomes cheap to reason about: each iteration is a claim about the workflow, tested against criteria agreed in advance. Measurement closes the loop, and customer impact is the only result that ends it.
The discipline is refusing to move forward until every stage of that loop has an owner and a number.
Themes
- Frontier AI
- AI customization
- Customer workflows
- Evaluation
- Model & product quality
- Experimentation
- Customer discovery
- AI research collaboration
- Data science
- Engineering
- Product strategy
Details of Microsoft's product architecture, technologies and customers are deliberately not described here. Happy to talk through the approach in conversation.