Click any text to edit inline
AI Productivity Initiative

The Future of Knowledge Work

An internal skills ecosystem for AI-augmented productivity at Booking.
Not another tool. Infrastructure that makes us uniquely capable.

See the Vision ↓
The Problem

Knowledge work is broken

And it's not a tools problem - it's a structure problem

72% Enterprises now use AI agents Zapier 2026 Survey
12 Avg AI agents per company Salesforce 2026 Report
45+ Min autonomous agent sessions Anthropic Feb 2026
0 Booking skills ecosystems Current state

Context Fragmentation

Knowledge scattered across Slack, Gmail, Docs, Jira, Confluence, Zoom, Glean. Each tool captures a slice. None connects them.

Reconstruction Overhead

Every decision, every meeting, every handoff requires reassembling context from scratch. Hours lost every week.

AI That Doesn't Persist

ChatGPT answers questions. It doesn't do work that lasts. Conversations vanish. Context is lost.

Siloed Craft Knowledge

Each PM reinvents workflows. Best practices don't compound. New PMs start from zero.

Industry Reality

What's Happening Outside

The agent revolution is here. The question is: are we building or buying?

CUTTING-EDGE AGENT PATTERNS (2026) Based on published research from Anthropic, OpenAI, Vercel Claude Code Architecture OpenAI Codex Vercel AI SDK +-----------------------+ +--------------------+ +------------------------+ | Extended Autonomy | | App Server | | Agent Patterns | | - 45+ min sessions | | - Client protocol | | - Parallel Processing | | - Persistent context | | - IDE integration | | - Orchestrator-Worker | | - Tool orchestration | | - Enterprise APIs | | - Evaluator-Optimizer | +-----------------------+ +--------------------+ +------------------------+ | | | v v v +-----------------------------------------------------------------------------------+ | COMMON PATTERN: Skills/Capabilities as Composable Units | | | | Skill = Prompt Template + Tool Access + Context Rules + Output Format | | | | Claude: skills.sh OpenAI: Codex abilities Vercel: Agent Skills | +-----------------------------------------------------------------------------------+ | v THE GAP: None of these are enterprise-ready for Booking - No Glean integration - No internal data access (BDX, Research, CS) - No compliance/security review - Generic context, not Booking context

Claude Code

AI agents now run 45+ minutes autonomously, up from 25 min three months ago. Proves AI can do sustained, durable work in structured environments.

Skills framework already production-proven
Source: Anthropic Feb 2026, TechRadar

OpenAI Codex

GPT-5.2-Codex launched Jan 2026. App Server provides stable protocol for agent integration across IDEs and enterprise systems.

Enterprise integration patterns established
Source: OpenAI Jan 2026

Vercel AI SDK

Published patterns: Parallel Processing, Orchestrator-Worker, Evaluator-Optimizer. Authenticated Workflows for enterprise security.

Multi-agent orchestration is mature
Source: Vercel AI SDK Docs 2026
Why enterprise-level is non-negotiable: External tools cannot access internal data (BDX, Research, CS). They lack Booking context. They haven't passed security review. Building internal infrastructure isn't about NIH - it's about capabilities external tools cannot provide.
The Vision

An Internal Skills Ecosystem

Building infrastructure that makes Booking uniquely capable

📌

PM Spine

Common structure for meetings, docs, tasks, analysis. Every PM starts with the same foundation.

Skills on GitLab

Curated AI capabilities hosted internally. Install what you need. Governed and secure.

🔌

Enterprise Skills

Ask Research, Ask CS, Ask BDX. Booking context that generic AI can never provide.

📈

Layer Model

Individual -> Team -> Leaders. Start with personal productivity, scale to team intelligence.

Key insight: This is not about using external tools. It's about building internal infrastructure. Skills hosted on GitLab. Deeply integrated with Glean, Research, CS, BDX. Designed for Booking.
Foundation

The PM Spine Structure

A minimal knowledge base structure optimized for AI retrieval

PM-KB/ +-- 1-Projects/ # One folder per project | +-- Project-Name/ # PRDs, briefs, context | +-- Meetings/ # Chronological log | +-- YYYY-MM/ # NOT by project | +-- People/ # Stakeholder files | +-- FirstName-LastName.md | +-- 0-Inbox/ # Quick capture | +-- Claude-Questions.md | +-- skills/ # Bundled skills +-- _Templates/ # PM artifacts +-- _My-Work-Board.md # Weekly Kanban +-- CLAUDE.md # AI instructions

Why This Structure Works

  • Retrieval-optimized - AI can quickly find context without searching everything
  • Meetings are chronological - Not buried in project folders. Time-based retrieval.
  • People are central - Stakeholder context always one query away
  • Minimal but extensible - Start simple. Add folders/skills as needed.
  • Tested in daily use - Developed through iteration, not theory
Honest assessment: This structure emerged from one PM's daily workflow. It's a starting point, not a proven enterprise solution. Needs validation with diverse PM workflows.
Unique Pattern

Process Comments Workflow

Inline instructions that execute automatically

THE >ANSHUL: COMMENT SYNTAX Embed actionable instructions directly in markdown files EXAMPLE: Meeting Notes File +------------------------------------------------------------------+ | # 2026-02-28 Product Review | | | | ## Discussion Points | | - Reviewed Q1 roadmap progress | | - Discussed timeline for PE flow | | | | >Anshul: Create action item for Christina re: API review | | >Anshul: Add this decision to Dynamics/Decisions.md | | >Anshul: Schedule follow-up for next week | | | | ## Next Steps | | ... | +------------------------------------------------------------------+ EXECUTION FLOW: +-----------------+ +-------------------+ +------------------+ | /process-comments|---->| Parse all files |---->| Execute each | | command | | for >Anshul: | | instruction | +-----------------+ +-------------------+ +------------------+ | +-------------------------------+ v +------------------+ | Remove processed | | comments from | | source files | +------------------+ WHY THIS MATTERS: - Capture intent at moment of thought - Batch process asynchronously - No context switching during meetings - Instructions persist until processed - Works with any markdown file
This is a novel pattern - not found in standard AI workflows. It bridges async capture with automated execution. One of several workflows that emerged from daily use.
See It In Action

Before & After

Concrete examples from the working prototype

/process-meeting Process any meeting transcript

Drop a Zoom transcript, say the command, get structured notes with decisions and actions extracted.

Before (Manual)

  • Take notes during meeting
  • Type up summary afterward
  • Manually identify decisions
  • Create action items in Jira
  • Search Glean for context
  • Update project files
~30-45 minutes per meeting

After (With Skill)

  • Drop transcript, say "/process-meeting"
  • AI creates structured notes
  • Decisions extracted automatically
  • Action items with owners
  • Glean context auto-enriched
  • Work Board updated
~2-3 minutes per meeting
# 2026-02-28 BHFS Alignment Sync ## Attendees - [[Christina Mueller]] - Engineering Lead - [[Sarah Chen]] - Compliance - [[Anshul Goyal]] - PM ## Decisions Made 1. **PE flow will be modular** - can complete across sessions 2. **Direct/Indirect gating question** confirmed for all partners ## Action Items - [ ] @Anshul: Draft PE matching logic by Friday - [ ] @Christina: Review API contract for PE creation ## Context (from Glean) - Related Jira: BHFS-1234, BHFS-1256 - Prior discussion: Email thread from Feb 25
/morning-summary Start your day with context

One command pulls calendar, tasks, recent context. You know exactly what matters today.

# Friday, February 28, 2026 ## Today's Focus 🔥 **BHFS 5.3 deadline in 2 weeks** - PE matching logic critical path ## Calendar (from Glean) - 10:00 - BHFS Standup - 14:00 - Stakeholder Review with [[Sarah Chen]] - 16:00 - 1:1 with [[Manager Name]] ## Pending from Yesterday - [ ] Review Christina's API proposal - [ ] Respond to compliance questions ## This Week's Progress ✓ PE flow design approved ✓ Presentation drafted ⌛ Matching logic (in progress)
Honest Assessment

What's Missing Today

Critical gaps that need enterprise solutions

CURRENT ACCESS LANDSCAPE WHAT WORKS TODAY | WHAT'S BLOCKED +--------------------------------+ | +--------------------------------+ | Glean (via MCP) | | | Calendar (no native access) | | - Document search ✓ | | | - Only via Glean search | | - Comment retrieval ✓ | | | - Can't create/modify events | +--------------------------------+ | +--------------------------------+ | Zoom Transcripts | | | Gmail (no native access) | | - Local file access ✓ | | | - Only via Glean search | | - Auto-processing ✓ | | | - Can't send/draft emails | +--------------------------------+ | +--------------------------------+ | Local Files | | | Slack (no native access) | | - Full read/write ✓ | | | - Only via Glean search | | - Markdown native ✓ | | | - Can't post messages | +--------------------------------+ | +--------------------------------+ | Jira (via Glean) | | | Markdown Viewer | | - Ticket search ✓ | | | - No approved tool like | | - Status lookup ✓ | | | Obsidian, Notion, etc. | +--------------------------------+ | +--------------------------------+ GLEAN MCP = PARTIAL SOLUTION - Read access: YES (search docs, tickets, calendar events) - Write access: NO (can't create, modify, or send) - Real-time: NO (search-based, not live connection)
Blocked

Markdown Viewer

No approved tool for viewing/editing KB files. Obsidian ideal but not on approved list. VS Code works but not user-friendly for non-technical users.

Need: IT approval for Obsidian or equivalent

Blocked

Calendar Native Access

Can search calendar via Glean but can't create events, send invites, or modify existing meetings programmatically.

Need: Calendar API access (Google/Outlook)

Partial

Gmail Access

Can search emails via Glean MCP. Cannot draft, send, or manage emails directly. Manual copy-paste required.

Need: Gmail API (at least draft creation)

Partial

Slack Access

Can search Slack via Glean. Cannot post messages, react, or manage channels. Context retrieval only.

Need: Slack API (at least read channels)

These gaps are solvable: They require IT/Security approval, not technical invention. The patterns exist. The integrations are standard. We need organizational support to unlock them.
Skills Library

Installable Capabilities

Curated skills hosted on internal GitLab

Skill Category What It Does
/morning-summary Base Pull calendar, tasks, recent context -> daily priorities
/end-of-day Base Summarize accomplishments, queue tomorrow
/process-meeting Base Transcript -> notes + decisions + actions
/process-comments Base Execute >Anshul: instructions across all files
/ask-research Enterprise Query Booking Research findings across the company
/ask-cs Enterprise Understand customer service insights for your domain
/ask-bdx Enterprise Pull data insights from Snowflake/BDX layer
/draft-prd Craft Structured PRD generation with Booking templates
/stakeholder-update Craft Weekly update formatted for leadership
Scale Path

The Layer Model

Focus on Layer 0 now. Design for team and leader layers.

0

Individual PM

PM spine + base skills + enterprise skills. Personal productivity transformation. Context always current. AI contributions persist.

Focus Now
1

Team

Published team context (goals, roadmaps, services). PM writes brief -> agent checks other teams. Cross-team dependency intelligence.

Vision
2+

Leaders

Project dashboards pulling from Jira, Docs, Gmail. Configurable tracking: progress, conflicts, risks. Teams push updates; leaders consume.

Vision
Expansion Path

Beyond Product Management

The skills ecosystem pattern applies to other knowledge work crafts

CRAFT EXPANSION MODEL Same infrastructure, different skills and spines SHARED INFRASTRUCTURE +--------------------------------+ | GitLab Skills Repository | | Glean MCP Integration | | Enterprise Data Connectors | +--------------------------------+ | +---------------------------+---------------------------+ | | | | v v v v +-------------+ +-------------+ +-------------+ +-------------+ | PM Spine | | Eng Spine | | Design Spine| | Ops Spine | | /process- | | /code-review| | /critique | | /incident | | meeting | | /arch-doc | | /user-flow | | /runbook | | /draft-prd | | /tech-spec | | /design-doc | | /postmortem | +-------------+ +-------------+ +-------------+ +-------------+ | v START HERE
📋

Product Management

Current focus. Spine defined. Skills in development.

💻

Engineering

Code review, architecture docs, tech specs, debugging.

🎨

Design

Design critiques, user flows, accessibility checks.

📊

Data/Analytics

Query generation, dashboard briefs, metric definitions.

👥

People Ops

Performance summaries, 1:1 prep, team health tracking.

📚

Legal/Compliance

Contract review, policy lookup, compliance checks.

💰

Finance

Budget tracking, forecast summaries, variance analysis.

🎙

Marketing

Campaign briefs, content calendars, performance reports.

The infrastructure is craft-agnostic. Once we prove the PM use case, the same GitLab repo, MCP integrations, and skill patterns can serve any knowledge work discipline at Booking.
Resourcing

What Would It Take?

Industry benchmarks and realistic team sizing

INDUSTRY BENCHMARKS: AI STRATEGY TEAM SIZES Source: Wharton/GBK Collective AI Adoption Report 2025 Team Size Distribution (Enterprises with AI Strategy Teams): 1-4 members ████████████████████████████████████ 36% 5-9 members ██████████████████████████████████ 34% 10-24 members ████████████████████████ 24% 25-49 members ██████ 6% 50+ members (emerging in purchasing, ops, finance) RECOMMENDATION FOR BOOKING SKILLS ECOSYSTEM PHASE 1: Validate (Q1-Q2) PHASE 2: Scale (Q3-Q4) +---------------------------+ +---------------------------+ | 1 PM Lead (could be | | 1 PM Lead | | existing allocation) | | 1 Eng (MCP/integrations) | | 1 Part-time Eng support | | 1 Content/Training | | 5-10 Pilot PMs | | 20-50 Active Users | +---------------------------+ +---------------------------+ | | v v Prove Value Build Platform Refine Skills Enterprise Skills Document Patterns Team Onboarding MINIMAL VIABLE TEAM: 1 dedicated PM + part-time eng support This is a platform play, not a product team - infrastructure investment
Key insight: This doesn't require a large team initially. It requires dedicated time and organizational support. The core work is skill design and adoption, not engineering. Enterprise connectors need eng support but are bounded scope.
Timeline

Phased Approach

Prove individual value first. Then scale.

Phase 1: Validate Layer 0

Q1-Q2 2026
  • Pilot with 5-10 PMs
  • Refine base skills from real usage
  • Document adoption path
  • Identify "magic moments"
  • Establish GitLab skill repo
  • Request Obsidian approval

Phase 2: Enterprise Skills

Q2-Q3 2026
  • Build Ask Research connector
  • Build Ask CS connector
  • Build Ask BDX connector
  • Partner with data teams
  • Expand skill library to 15+
  • Calendar/Gmail API access

Phase 3: Design Layer 1

Q3-Q4 2026
  • Define team context format
  • Prototype dependency checking
  • Engineering partnership
  • One team pilot
  • Evaluate other craft expansion
What We Need

The Ask

Help us make this real

1

Collaborators (2-3 people)

Help refine the vision. Challenge assumptions. Shape what this becomes. PMs, EMs, or anyone passionate about AI productivity.

2

Pilot Team (5-10 PMs)

Adopt Layer 0. Test the spine structure. Install skills. Give feedback. Help us find what works and what doesn't.

3

IT/Security Support

Obsidian approval. Calendar API access. Gmail draft creation. These are the critical unblocks for enterprise adoption.

4

Sponsorship

Support to pursue this as a formal AI productivity initiative. Dedicated time allocation. Visibility with leadership.

Sources & References