AI Agent is Our New CRM

Who am I?

• Ben Cheng, CEO @ Oursky / SkyMakers Digital
• Building software products and services since 2008

Software Today

• Clear things you can point at on the screen

AI Agent Interface

• Chat with verbs: "do this", "show me that", "update the deal"

Problem: Discoverability

• "What can this thing even do?" - classic bad UI problem
• But: people are getting comfortable (chat box is everywhere now)

What My CRM Agent Can Do

/daily          Full briefing: overdue leads, tasks, meetings, key contacts

/company        Add ACME Corp
                Look up companies

/person         Add Jane Doe from ACME
                Manage contacts

/deal           New deal with ACME
                Manage deals across 3 pipelines

/meeting        Log a discovery meeting with Jane

/note           ACME is evaluating FormX

/task           Send proposal to ACME by Friday

Problem: It's Slow

• Agent thinking + tool calls + response = seconds to minutes
• Especially bad with thinking/reasoning mode enabled
• Visual indication of progress is critical
• For non-tech users: stream token output so they see something happening
• User adaptation: learn to multi-task while agent works

Best of Both Worlds - UI Inside Chat

• Generate UI components inline within conversation
• Claude Artifacts: Generate full HTML/React inside chat
• Google A2A / A2UI: Agent-to-UI protocol
• Vercel AI SDK: JSON schema to rendered components

Practical Reality - Web UI for Reports

• I still have a web dashboard for common pipeline reports
• Sales team uses it for quick scanning (pipeline.html)
• Agent for operations, Web UI for visualization

[ Quick demo of pipeline report ]

Why Build This Way?

Not just because it's cool. It enables:

• Auto-draft email replies and sales follow-up emails (CRM to Gmail)
• Auto-generate weekly sales pipeline reports
• Meeting transcript → Notes → Internal Deals Memo, automatically
• Any sales can compose workflows across tools without custom code

WTF is MCP?

• Model Context Protocol - open standard by Anthropic
• Lets AI agents call external tools and access data
• Think of it as "USB-C for AI agents" - one protocol, many services
• Your app exposes capabilities as MCP tools
• Any MCP-compatible agent can use them

Why Remote MCP?

Remote MCP Setup

Remote MCP - OAuth 2.1

• Login via OAuth - each user has their own identity and permissions
• Scopes limit what the agent is authorized to do
• No more shared admin API keys with god-mode access
• Optionally: ClientID Metadata Document / Dynamic Client Registration
• We use Authgear for the OAuth layer

Design for Human Usage

• "RESTful" MCP tools are not enough
• CRUD is for developers, not users
• Think: what would your user actually ask?

The N+1 Problem

• If you mirror REST APIs: agent calls search_leads (N results), then N x get_lead_details
• AI is smart - it tries to fetch everything and filter locally
• Result: expensive (tokens), slow (round trips)

Trick: set MAX_MCP_OUTPUT_TOKENS low in Claude Code to force efficient tool design

Design for Tokens

• Denormalization modes: concise vs detailed
• Add more filters so agents fetch less data
• Always limit to first N results
• Include clear message: "Showing 10 of 47. Ask agent to refine search."

Unknown Parameters

• AI hallucinates new parameters: list_leads(name_contains="ACME")
• Your tool doesn't support name_contains - it silently ignores it
• The call "works" but returns wrong results

It's for AI Agents!

• Write clear, specific tool descriptions - agents read them
• Don't make tools too similar (agent gets confused)
• Sometimes execute_sql just works - AI is already great at SQL
• Think about tool composition: how will agents chain your tools?

MCP = Enforcement

• Guided by code
• Authorization, permissions, data validation
• Audit trail on every mutation
• Financial calculations, compliance rules
• The server enforces what users can and cannot do

Skills = Guidance

• Guided by team knowledge and domain expertise
• Best practices, workflow advice, templates
• Easy to change - updated daily by the team
• "How should I qualify this lead?" "What's the deal health checklist?"
• The skill guides; the human decides

The Tricky Question

"We don't want duplicate companies/persons in the CRM"

Should the duplicate check be in Skills or MCP tools?

How I Built It

• Started with a plan document, then coded for 3 days for v1
• 10 Agent Skills layered on top
• Run code review agents for every non-trivial commit

Unit Testing / TDD

• Test-driven development with AI coding agents works really well
• Tests are the specification - agent reads tests, writes implementation
• Catches regressions when agent refactors code
• Chrome DevTools MCP - very useful for testing agent-generated UI in real-time

First-Class Artifacts

• Execution Plans - structured plans before coding starts
• Business Rules / Design Specs - the agent's requirements doc
• CLAUDE.md - project context that persists across sessions

CLAUDE.md

# CRM Project

## MCP Server
- Base URL: https://mcp.example.com
- Auth: OAuth 2.1 (Authgear)
- Stack: TypeScript, Hono, SQLite, Fly.io

## Pipelines
- Services (Oursky): Lead → Discovery → Proposal → Closing → Won/Lost
- FormX.ai: Trial → Onboarding → Active → Expansion
- Authgear: similar to FormX

## Skills Available
- /daily, /company, /person, /deal, /meeting, /note, /task

## Conventions
- All amounts in USD unless noted
- Lead = unqualified, Deal = qualified opportunity

The agent reads this every session - it's the source of truth

Takeaways

• AI Agents are a new form of software, not a chat layer on existing apps
• MCP-first: design tools for agents, add UI on top (not the other way around)
• Remote MCP + OAuth: the only path for non-developer users — For Now
• Skills + MCP: guidance vs enforcement, both needed
• The next Salesforce might not have a traditional UI at all