typescript
const leadMatch = text.match(/LEAD_CAPTURED:\s*(\{[^}]+\})/);
const leadData = leadMatch
  ? (() => { try { return JSON.parse(leadMatch[1]); } catch { return null; } })()
  : null;

// Strip token from visible reply
const reply = text.replace(/LEAD_CAPTURED:\s*\{[^}]+\}/, "").trim();

return NextResponse.json({ reply, leadCaptured: leadData ?? null });
The user sees a natural conversation. The system sees a structured event. No model fine-tuning required — just a well-designed prompt and a 4-line parser.

Same pattern works for BOOKING_CONFIRMED, ESCALATE, APPOINTMENT_REQUESTED. One pattern, infinite use cases.

Pattern 2: Multi-persona routing
DeployInfra.AI serves different business verticals — restaurants, real estate agents, clinics, e-commerce shops. Each vertical gets a different agent persona with vertical-specific knowledge and qualification flows.

OpenClaw handles conversation state. I handle persona selection at the API layer:


const PERSONA_PROMPTS: Record<string, string> = {
  "Property Assistant": `You are Alex, an AI property assistant...`,
  "Clinic Receptionist": `You are an AI receptionist for a medical clinic...`,
  "Restaurant Assistant": `You handle reservations, menu questions...`,
  "B2B Sales Agent": `You qualify inbound B2B leads using BANT...`,
};

const systemPrompt = PERSONA_PROMPTS[persona ?? ""]
  ?? PERSONA_PROMPTS["Property Assistant"];
One OpenClaw deployment. Multiple production agents. The persona is just a system prompt — OpenClaw handles everything else.

Pattern 3: Proactive nudges
A user opens the chat widget and goes silent. Maybe they're reading. Maybe they're hesitating. A static chatbot just sits there. A good employee follows up.


const NUDGE_MESSAGES = [
  "Still here? Tell me your business type — I'll show your AI ROI in 30 seconds 👇",
  "What kind of business do you run? I'll show what your AI would've done overnight →",
  "Your competitors are already using AI for this. Want to see what it looks like? →",
  "One question: what does a missed customer enquiry cost your business? →",
];

useEffect(() => {
  if (isOpen || nudgeDismissed) return;
  const timer = setTimeout(() => {
    setNudge(NUDGE_MESSAGES[Math.floor(Math.random() * NUDGE_MESSAGES.length)]);
  }, 18000);
  return () => clearTimeout(timer);
}, [isOpen, nudgeDismissed]);
18 seconds. One message. Auto-dismisses after 9 seconds. Dismisses immediately if they open the widget. Conversion rate on nudged sessions is meaningfully higher than cold sessions.

Pattern 4: The gap detection loop (the one I'm keeping close)
This is the pattern that separates an AI employee from a chatbot, and it's the part I'm not going to fully document here.

What I will tell you is the outcome:

Every agent on DeployInfra.AI tracks the quality of its own responses in real time. Not user ratings — the agent itself knows when it's confident and when it's guessing. Low-confidence responses don't disappear. They accumulate.

Once a week, every business owner gets a digest. Not a wall of logs. One clean list:

"Here are the 5 things your customers asked most that your agent couldn't answer this week."

One click to fix each one. Next week, the agent knows.

Week 1: agents answer around 60% of questions confidently.
Week 4: 80%+.
Week 12: 90%+.

The agent gets better because it's being used — not despite its gaps, but from them.

I'm deliberately not publishing the implementation. The confidence scoring mechanism, how gaps get ranked, how the digest gets generated — that's the moat. What I'll say is that it runs entirely on conversation data your agents are already collecting. No extra instrumentation from the business owner. No tagging. No manual review.

If you want to build something similar, the insight to start with is this: your AI already knows when it's guessing. You just have to ask it.

The rest is the hard part.

The meta that I didn't plan
Here's the part I didn't fully appreciate until it was live.

The AI agent in the bottom-right corner of deployinfra.ai is an OpenClaw agent running the "DeployInfra AI" persona. Its job is to qualify visitors, understand their business type, reflect their specific pain back to them, and capture leads.

It is selling the platform that deploys AI agents.

If you go to the site after reading this, the agent will ask what kind of business you run. It'll tell you what your AI would have done last night while you were asleep. It'll try to get your email.

That's not a demo. That's the product running in production, on itself, in real time.

The article you're reading is a proof of concept. The agent on the website is the same proof running live. You don't have to take my word for it — you can talk to it right now.

What I'd do differently
Rate limiting is harder than it looks. Anonymous demo traffic is unpredictable. I ended up building a simple in-memory rate limiter (5 requests/IP/hour for unauthenticated users) that resets on cold starts — good enough for demos, not production. If I were starting again I'd reach for Redis from day one.

Personas diverge fast. The six personas I started with have split into twelve variations based on what actually converts in each vertical. OpenClaw makes this easy — each variant is just a different system prompt — but managing prompt versions without a proper versioning system gets painful quickly. Build that early.

Low volume is the enemy of the learning loop. The gap detection system needs a minimum number of conversations per week before the signal is useful. Below that threshold, the agent collects data silently and the digest doesn't fire. Plan for a cold start period where you're seeding conversations manually.

What's next
WhatsApp channel: Same OpenClaw agents, same persona routing, same gap detection — over WhatsApp Business API. One weekly digest regardless of which channel customers used.
Knowledge grounding: Anchoring agents in business-specific content — menus, listings, FAQs, price lists. Eliminates the category of hallucinations that come from the agent not knowing the specifics of a particular business.
Multi-channel identity: One agent that recognises a customer across web, WhatsApp, and email. Picks up the conversation where it left off.