# Turning a website chatbot into a qualified-lead engine

> How one of Singapore's largest private education institutions replaced a menu-driven bot that generated noise — and roughly doubled the qualified leads reaching its admissions team.

_2026-07-22_

---

---

## Results

Einstein Bot → Agentforce · Oct–Nov vs Jan–Feb.

| Metric                    | Before | After   | Change      |
| ------------------------- | ------ | ------- | ----------- |
| Leads per month           | 176    | 411     | +134%       |
| Converted leads per month | 4      | 8.5     | +113%       |
| Lead conversion rate      | 2.3%   | 2.1%    | held steady |
| Human escalation rate     | 10%    | 6%      | −4 pts      |
| CSAT score                | —      | 4.3 / 5 | 150 ratings |
| Avg. session time         | —      | 9.8 min | new signal  |

Conversion held near 2% even as monthly volume more than doubled, so the near-2× rise in converted leads comes from a wider top of funnel, not a softer bar — the added leads converted at the same rate the old, smaller funnel did, so the extra volume kept its quality.

---

## The Problem

The institution already had a website chatbot — a menu-driven Einstein Bot. That was the problem. Prospects had to navigate fixed options instead of simply asking what they wanted to know, so many dropped off before reaching an answer. The leads that did come through were often low quality — incomplete or unqualified — leaving the admissions team to triage rather than advise. And with answers spread across a large knowledge base and a separate programme catalogue, the bot was slow to respond, compounding the drop-off. They didn't need a chatbot; they had one. They needed it to actually generate pipeline.

---

## Why It Mattered

For an institution recruiting at this scale, the website is the top of the funnel — the first and often only touch a prospect has before deciding where to enquire. Every drop-off is a prospective student who went to ask a competitor instead; every unqualified lead is advisor time spent triaging rather than converting the ones who are ready. A bot that can't answer plainly isn't a neutral cost — it quietly caps growth at the widest part of the funnel.

---

## Our Approach

The gains didn't come from a bigger model — they came from measuring first and rebuilding the flow around what the data showed.

- **Measured before we optimised.** We turned on agent analytics and traced every response end to end, and found that ~80% of prospect queries could be answered from the knowledge base alone — the insight that reshaped the whole flow.
- **Made the knowledge base the source of truth.** Instead of querying knowledge and website in sequence every time, the agent checks knowledge first and only falls back to a live web crawl for programme and fee data — cutting redundant calls and sharpening response time.
- **Added friction where it pays.** A short pre-chat qualification form filters out spam and half-formed enquiries up front, so the leads that reach admissions are real and ready to advise.
- **Put the numbers in the business's hands.** Every metric is reconciled against agreed definitions and published as native Salesforce reports, so stakeholders track the agent's impact themselves rather than relying on a slide.

---

## What We Built

**Conversational agent.** Natural-language Q&A across 1,000+ published knowledge articles — prospects ask in plain words and get an answer, not a menu.

**Automated lead capture.** A pre-chat form with email validation and consent, writing qualified leads straight into Salesforce with their interests attached.

**Knowledge-first retrieval.** The agent answers ~80% of queries directly from the knowledge base, with a web-crawl fallback for live programme and fee details.

**Business-ready reporting.** Salesforce-native reports on leads, conversion, escalation and CSAT — so marketing and admissions see impact in their own dashboards.

---

## What We Learned

**Instrument before you optimise.** The 80%-answerable finding reshaped the whole design. Without analytics on first, we'd have tuned the wrong things.

**Let prospects ask in their own words.** A conversational agent self-qualifies where a menu just filters — the same traffic yields more, and better, leads.

**Friction in the right place raises quality.** A short qualification step up front cut junk without shrinking the funnel — volume more than doubled and conversion held.

---

## Partnership

The institution owned what only it could: the knowledge base content across 1,000+ articles, the live programme and fee data, and the definitions behind every metric that mattered. That last part is where most "impact" claims fall apart — so we reconciled every number against definitions the team agreed to, and published them as native Salesforce reports the institution runs itself. They own the knowledge and they own the scoreboard; we built the agent that sits on top.