Capstone: The AI Chatbot Launch

Overview

The Scenario

You are the product manager for Fieldwork, a B2B SaaS platform used by 2,400 mid-market operations teams to manage field service scheduling, dispatch, and reporting. Average contract value is $18,000/year. Current NPS is 42. Your support team handles approximately 1,800 tickets per month.

The CEO returns from a conference and sends you this message on a Monday morning:

Your job: figure out whether this is worth doing, how to do it right, and what to deliver at each stage.

Stage 1: Signal Capture — Is This Worth Building?

Do not open a project in Jira. Do not schedule an engineering scoping session. The first work is validation.

Signal Capture Questions

Is there evidence of real user need?

Pull your last 90 days of support ticket data. Categorize tickets by:

• Type (how-to / bug report / billing / data question / other)
• Resolution path (self-serve possible / required human judgment / required engineering)
• Volume by category

What you might find for Fieldwork:

Analysis: Approximately 1,060 tickets per month (59%) are candidates for AI deflection. If the AI deflects 40% of those, that is approximately 424 fewer tickets per month, which is meaningful for a team managing 1,800/month.

The "unlimited humans" test: If you had 10 more support reps, what would they do? They would answer the same how-to and configuration questions faster. An AI doing this is not transforming anything. It is automating a cost center. That is legitimate value, but frame it correctly.

Is 90 days realistic?

Run the Integration Complexity Estimation (Chapter 10):

Total: 13 — Medium complexity. 90 days is aggressive but feasible for an MVP. Not feasible for a polished, production-grade chatbot.

Signal Capture recommendation: Proceed, but reframe the CEO's request from "ship in 90 days" to "run a controlled pilot in 90 days with defined success criteria."

Signal Capture Deliverable

A one-page recommendation memo to the CEO:

Stage 2: Value Hypothesis and Kill Criteria

Value Hypothesis

We believe that adding an AI-powered support chatbot to Fieldwork

Will help operations managers who encounter setup or how-to questions during their workflow

By giving them instant, accurate answers without waiting for a support ticket response

Resulting in a 30%+ reduction in Tier 1 support ticket volume, with user satisfaction scores on chatbot interactions above 3.5/5.0

We will know this is true when the chatbot resolves 30%+ of tickets without human escalation, and users who used the chatbot have equivalent or better support NPS compared to ticket-based support.

Kill Criteria

Define these before you start, write them down, and share them with the CEO.

If any kill criterion is met at the 45-day pilot midpoint review, the pilot is paused and the team convenes a kill/pivot decision. No exceptions. The CEO agreed to these criteria at the start; hold the line.

Stage 3: Grounded Delivery — Phased Plan

Goal: Get the infrastructure right before showing anything to users.

Work:

• Export and structure all support documentation (KB articles, FAQs, common responses) for RAG ingestion
• Set up shadow mode: chatbot processes all new support tickets in the background; team reviews outputs but nothing is shown to users
• Build evaluation dataset: 200 representative support queries with human-written ideal answers
• Define chatbot persona and escalation rules

Deliverable: Shadow mode live; evaluation baseline established.

Shadow mode targets: Achieve >70% match rate between chatbot responses and human-ideal responses before proceeding to Phase 1.

Goal: Validate with zero customer risk.

Work:

• Enable chatbot for internal Fieldwork employees only
• Collect thumbs-up/down, corrections, and escalations
• Manual review of 100% of responses during this phase
• Iterate on prompt engineering and knowledge base gaps

Deliverable: Internal quality report; decision to proceed to customer pilot.

Gate criteria: Internal user satisfaction >4.0/5.0; error rate <5%.

Goal: Validate with real users in a controlled environment.

Work:

• Enable chatbot for 10–15% of customer base (opt-in during onboarding or via prominent in-app prompt — NOT default-on at this stage)
• Measure against all kill criteria weekly
• Human review of escalated conversations daily
• 45-day midpoint review with go/kill/pivot decision

Deliverable: Pilot results report with recommendation.

Who to include in pilot: Power users who are comfortable with beta features; accounts with dedicated CS managers who can provide qualitative feedback; avoid your 10 largest accounts until Phase 3.

Goal: Expand safely with quality maintained.

Work:

• 25% → 50% → 100% rollout over three weeks
• Default-on for new users; opt-in for existing users for first 30 days
• Continue monitoring all quality metrics
• Update support team on volume trends; retrain on escalation patterns

Deliverable: Full production deployment; operational handoff to support team.

Stage 4: Evaluation Framework

Evaluation Dimensions

Tier 1: Accuracy

Run your 200-query evaluation set through the chatbot every time you update the knowledge base or prompt. Track:

• Match rate against human-ideal answers (automated semantic similarity)
• Factual error rate (human review of random sample)
• Hallucination rate (answers that assert things not in the knowledge base)

Tier 2: User Satisfaction

After each chatbot interaction, show a brief survey:

• "Did this help you solve your issue?" (Yes / Partially / No)
• "How satisfied are you with this response?" (1–5)
• [If "No" or <3]: "What was missing or wrong?" (free text)

Track these by query category, user segment, and over time.

Tier 3: Business Impact

• Weekly ticket volume trend in pilot vs. control group
• Average time to resolution (chatbot vs. ticket)
• Support team capacity freed (tickets saved × avg. handle time)
• Support NPS trend in pilot vs. control group

Tier 4: Cost

• Cost per chatbot interaction (all-in: tokens, retrieval, infrastructure)
• Cost per deflected ticket
• Monthly AI cost vs. monthly support cost savings

Evaluation Cadence

Stage 5: Rollout and Monitoring Plan

Rollout Communication

Internal: Brief the support team before the pilot launches. Frame the chatbot as a tool that handles routine volume so support can focus on complex, relationship-driven issues. Provide training on the escalation workflow. Establish a direct feedback channel for support reps to flag chatbot errors.

Customer-facing: "We're piloting an AI support assistant to help you get faster answers. If it doesn't solve your issue, it will connect you to our team in one click. [Enable for my account / Keep standard support]."

Do not call it "AI-powered" in customer-facing language without first confirming your customer base reacts positively to AI branding. Some B2B operations buyers are skeptical. "Intelligent assistant" or "automated help" may be preferable framing depending on research.

Monitoring Setup

Before Phase 2 launches, the following must be live:

Template Deliverables at Each Stage

Common Failure Modes to Avoid

Shipping without shadow mode: Teams skip shadow mode because 90 days is short. Do not. Shadow mode in weeks 1-3 is what makes the rest of the plan safe. If shadow mode reveals that the chatbot is wrong 40% of the time, you have saved yourself from a customer trust crisis.

Letting the CEO's timeline override the kill criteria: The most important decision in this capstone is maintaining discipline at the 45-day review. If the metrics do not support proceeding, say so. A failed chatbot in production is more damaging to the CEO's goals than a delayed launch.

Ignoring the support team: The support team is a critical stakeholder and a crucial source of quality feedback. Loop them in early, treat them as partners, and make sure their workflow is respected during the transition.

Optimizing for deflection rate instead of resolution quality: A chatbot that deflects 50% of tickets by giving vague non-answers that frustrate users has made your product worse, not better. The success metric is resolved issues, not deflected tickets. Define the distinction clearly and track both.

Never closing the feedback loop: If users flag errors and nothing changes, they stop flagging errors. Demonstrate that feedback leads to improvements. Announce when you have fixed something a user reported. This converts skeptics into advocates.