How I work
Redesign the work, then add the agents.
AI transformation is a business transformation, not a tool rollout. I employ a consistent five-step framework on every initiative, from a single Copilot skill to a multi-agent platform, and they follow the Frontier Firm approach: set the ambition, redesign the work, codify the advantage and controls, scale with the right level of AI, and invest in people.
Five steps on every AI initiative
-
01 Outcome
Anchor to a business outcome
Agree with leadership on the objective and how it's measured: a better employee experience, stronger stakeholder engagement, a faster process or new innovation. Efficiency is the floor, not the ceiling, and every goal has a named owner.
-
02 Workflow
Redesign the work before adding agents
Analyze the end-to-end workflow and optimize by removing redundancies before introducing AI. Then pick the recipe: accelerate a persona with an assistant, redesign the process with AI, or start from a blank page with an AI-first build.
-
03 Rules
Codify the advantage and the controls
Turn standards into deterministic rules, rubrics and private evals. In the Control Tower I built, about 170 rules produce the numbers, and AI text never overrides them.
-
04 Agents
Set the level of AI, then scale it
An assistant, a human-agent team, or human-led and agent-operated, chosen by risk and data readiness. Agents cite their sources, an evaluator gate blocks anything that fails a fact check, and one engine scales through business areas and brand packs.
-
05 People
Keep people at the center
Honest, hands-on enablement (AI Days, Copilot coaching, champions and workshops), a person approving every decision that matters, and security and data boundaries built in from day one.
Levels of AI: L1 · assistant every person has an AI helper · L2 · human-agent team agents take tasks under human direction · L3 · human-led, agent-operated agents run the workflow and people steer. Every project below is tagged with its recipe, level and objective. The vocabulary follows Microsoft's Frontier Firm playbook (second edition).
AI-first engineering
Built with Claude Code, reusable skills and automated checks
As champion of the Claude Code & Design Pilot, I codify reusable prompt, skill, and agent patterns into a shared library of governed delivery practices. These are built on the corporate design system and include reusable M365 skills and brand kits to ensure repeatable, high-impact outcomes.
-
Design to code
Prototype, build, standardize
I prototype in Figma Make, build with Claude Code, and standardize on the corporate design system. The Control Tower, Retina Evidence Hub and EURETINA map all began as Figma Make prototypes, and CLAUDE.md project memory carries context from one session to the next.
-
Skills library
Repeatable work becomes a skill
Skills cover gap assessment, Fast Start workshops, consulting-plan validation, strategy workshops, congress abstracts and sponsorship requests, in Claude and in Microsoft Copilot for PowerPoint, Word and Excel. I test them with evaluations, including side-by-side runs against a no-skill baseline.
-
Verification
Checked by default
Over 1,000 automated checks safeguard the Control Tower; their efficacy was validated through adversarial testing by deliberately introducing errors to ensure detection. Scheduled jobs validate data before publishing, and independent subagents review outputs for unsupported claims.
| tnt-gap-skill/ | Gap assessment engine with brand packs, shipped as Copilot skills for PowerPoint |
|---|---|
| mhi-fast-start/ | Fast Start workshop and consulting-plan kit for any therapeutic area |
| eyecare-fast-start/ | Eye Care edition of the workshop kit |
| ccg-plan-validator/ | 40 deterministic rule tests, then an 8-dimension advisory judge |
| necsv-sap-skill/ | Strategic Imperative and Scientific Objective validation with readiness scoring |
| congress-abstract-intelligence/ | Congress program to abstract index to session tiering |
| mhi-abstract/ | Builds and checks branded congress abstract decks |
Automations in the flow of work
-
Self-validating dashboard
A scheduled automation rebuilt an allocation dashboard from the latest planning workbook daily. The process incorporated guardrails to halt execution on unexpected row counts or totals, and passed syntax and render checks before deploying the live update.
-
Audited reconciliation
AI-driven invoice processing was reconciled against timesheets and the finance ledger, with a secondary PDF engine validating every invoice. The resulting accuracy reframed strategic budget discussions.
-
Skills that refuse bad input
A sponsorship-request workflow was converted into an installable skill that prevents submissions with missing fields. The system was rigorously stress-tested by independent subagents prior to release.
-
Multi-format workshop packs
Workshop decks, pre-reads, interactive planning tools, and posters are maintained under strict version control, ensuring a consistent narrative across all formats.