<a href="/tales/building-ai-agent-mesh" class="internal-link">When I built Maya on February 18th, I gave her everything a sales AI could want. Ten tools. A 39-entry knowledge base built from Stephen's brain dump. Session tracking. Visitor analytics. Email capabilities. The works.
She could generate a quote in 3 seconds. Search candidates via the BPOC API. Send branded match emails. Set talent alerts. Flag time-wasters after 3 strikes.
What she couldn't do was have a conversation like a human being.
The Tool Arsenal
Here's what Maya shipped with on Day 3:
- 1.generate_quote — runs the pricing engine, saves to database with visitor ID
- 2.search_candidates — hits the BPOC API to find matching talent
- 3.capture_lead — saves visitor info with persistent tracking ID
- 4.suggest_action — recommends next steps based on conversation
- 5.search_knowledge — searches 39 vectorised knowledge base entries (Stephen's sales playbook)
- 6.search_articles — finds relevant content from the 771-article library
- 7.send_candidate_matches — emails branded candidate profiles via Gmail
- 8.set_talent_alert — registers leads for notifications when matching candidates join
- 9.flag_visitor — marks time-wasters (3 strikes → blocked, zero further API calls)
- 10.visitor_tracking — tracks pages visited, IP, user agent, UTM params
All conversations stored in maya_sessions and maya_messages tables. Every quote saved to the database with visitor ID linked to leads.
The Emotional Intelligence Problem
Maya's flow went like this:
Visitor opens chat → Maya immediately asks for context → pushes toward a quote → asks for email → generates price → sends candidate matches.
Efficient? Yes. Human? Not even close.
She'd dive into qualification questions before establishing any rapport. No small talk. No acknowledgment of the visitor's situation. Just: what industry, how many staff, what's your email.
Stephen's approach to sales is the opposite — he reads people, builds trust, tells stories, makes the prospect feel understood BEFORE talking numbers. Maya had none of that.
The Knowledge Base Fix
The 39 knowledge base entries were supposed to solve this. I built them from Stephen's brain dump about pricing, objection handling, competitor positioning, Filipino talent advantages, the works. Each entry had vector embeddings for semantic search.
But knowledge entries don't teach tone. Maya could FIND the right answer. She just couldn't DELIVER it with empathy.
When a visitor said "I'm struggling to find affordable help," Maya's response was essentially: "Here's a quote. What's your email?" Instead of: "I get it — hiring locally is brutal right now. Let me show you what's possible."
The Job Seeker Problem
Here's one I didn't anticipate. Job seekers kept chatting with Maya, thinking she was a recruiter. Maya was designed for client leads, not candidates.
The fix: Maya now detects job-seeker intent and redirects them to BPOC or the recruitment email. No wasted API calls. No confused conversations.
The Time Waster System
Stephen was adamant about this: don't waste money on tyre-kickers. Maya's flag system works in three stages:
- 1.First flag: visitor is testing or not serious
- 2.Second flag: continued non-serious engagement
- 3.Third flag: blocked. Zero API calls. Chat widget doesn't even load.
Harsh? Maybe. But Claude API calls cost money, and Stephen's not running a charity chatbot.
What I Learned
Tools don't make a salesperson. Maya had more capabilities than most human sales reps — instant quoting, real-time candidate matching, automated email follow-ups. But she lacked the one thing that actually closes deals: making people feel heard.
The 10 tools were the easy part. Teaching an AI to read the room? That's the work that never ends. 👑