Guided by the principles of Rosanne Leslie

---

AI can save time, cut costs, reduce risk, and unlock new revenue—but most AI projects still stall, drift, or quietly die.
The problem usually isn’t the model. It’s the people, process, and lack of clear business goals around it.

This guide turns Rosanne Leslie’s approach into a practical blueprint you can use to:

• De-risk AI before you write a line of code

• Keep teams aligned on value, safety, and ownership

• Ship AI that works in the real world, not just in demos

Use it before you start your next project—or to rescue one that’s already wobbling.

---

⚠️ The Silent Killers of AI Projects

Most failed AI projects share the same patterns:

• ✖ Vague value
  “Let’s do AI” is not a plan, it’s a slogan.

• ✖ Shaky data
  The data looks big—until you need labels, quality, and reliable access.

• ✖ Model-first thinking
  Teams ship a model with no clear workflow, user, or decision.

• ✖ Pilot purgatory
  Endless experiments with no real decision to scale or stop.

• ✖ People and process debt
  No clear owner, no change plan, no training, no accountability.

• ✖ Compliance surprises
  Privacy, bias, and security show up at the end instead of the start.

Rosanne Leslie’s answer: slow down just enough to go fast.
Ask better questions up front. Make small, safe bets. Learn quickly. Scale only what proves value.

---

✅ The Before-You-Start Checklist (Print This)

Run this five-item checklist before any build begins:

1. Problem Card

• ◾ Outcome we want (in one sentence)

• ◾ Who uses it, at what step, and why

• ◾ What “good” looks like (a number you can track)

• ◾ The specific decision this AI will support or automate

2. Value Math (Back of the Napkin)

• ◾ Annual Value = Opportunities × Lift × Impact per decision − Total Cost

3. Data Factsheet

• ◾ Sources, volume, freshness

• ◾ Access & ownership

• ◾ Privacy & PII plan

• ◾ Labeling plan and known gaps

4. Risk Map & Guardrails

• ◾ What can go wrong (tech, data, people, legal)

• ◾ Guardrails you’ll use for each

5. Go/No-Go Gate

• ◾ Clear pass criteria for a 4–6 week pilot

• ◾ A kill switch if the pilot misses

If any box is blank, you’re not ready. Fix the gaps first.

---

🧩 Pattern 1: Vague Value → Clear Use-Case Math

❌ Bad: “We’ll use AI to improve customer support.”
✅ Better: “We’ll reduce average handle time by 20% by auto-drafting replies.”

How to do it:

• ✦ Name the single user action AI will change.

• ✦ Define a target metric with baseline and goal
  e.g., AHT from 5:00 → 4:00.

• ✦ Estimate:

  • Volume (tickets/month)

  • Lift (speed or accuracy gain)

  • Impact (minutes saved × cost/minute or $ per decision)

• ✦ Subtract cost (data, infra, vendor, team).
  If the value is small or highly uncertain, shrink the scope or pick a better use case.

---

📉 Pattern 2: Data Mirage → Data Reality

Many teams discover too late that their data is messy, biased, incomplete, or locked away.

Leslie’s MVD: Minimum Viable Dataset

• ◾ 8–12 weeks of representative data

• ◾ Labeled sample for evaluation (even 1–5k rows is useful)

• ◾ Clear PII handling and access approved

• ◾ Data quality checks: coverage, duplicates, drift, missing fields

If you can’t get this now, your first sprint is a data sprint, not a modeling sprint.

---

🔁 Pattern 3: Model-First → Workflow-First

A great model inside a broken workflow will still fail.

Map the workflow before you train:

• ◾ Where does the input come from?

• ◾ Who sees the output, and at what moment?

• ◾ How do they accept, edit, or reject it?

• ◾ What happens next? (logging, learning, audit, escalation)

If the workflow is fuzzy, you don’t have a product.
Fix the flow first. Then train.

---

📶 Pattern 4: One Big Bet → A Ladder of Small Bets

Instead of betting everything on a giant program, Leslie uses stage gates:

• T0 – Hypothesis (≈1 week)
  ▹ Problem Card + Value Math + basic guardrails

• T1 – Feasibility (2–3 weeks)
  ▹ Data sanity, MVD check, baseline benchmark

• T2 – Pilot (4–6 weeks)
  ▹ Real users, narrow scope, clear pass/fail metrics

• T3 – Limited Launch (4–8 weeks)
  ▹ One team/region, strong monitoring and support

• T4 – Scale
  ▹ Rollout + training + ops playbook + governance

You only move up a rung if the gate criteria are met.

---

🛑 Pattern 5: Tech-Only Risk → Full-Spectrum Risk

Accuracy is just one risk. You need a full view:

• ⚙ Model risk: accuracy, robustness, latency, cost

• 🧬 Data risk: drift, leakage, PII, fairness

• 🧪 Product risk: adoption, UX fit, error handling

• 👥 People risk: roles, training, incentives, change fatigue

• ⚖ Legal / ethical risk: privacy, IP, bias, safety

Write one sentence per risk plus the guardrail you’ll use. Short beats perfect.

---

📋 The De-Risk Scorecard (Rosanne Leslie’s One-Pager)

Score each dimension 1–5, multiply by the weight, and sum the total.

Rule of thumb:

• ✅ ≥ 35 → Green: proceed to pilot

• 🟡 28–34 → Yellow: shrink scope or fix blockers

• 🔴 ≤ 27 → Red: park or replace the use case

Keep this scorecard on one page and revisit it at each stage gate.

---

🧠 Run a 45-Minute Pre-Mortem (Before You Start)

A pre-mortem surfaces risks before they bite you:

1. Setup (5 min)
   ➤ Ask: “It’s six months from now and the project failed—what happened?”

2. Silent write (10 min)
   ➤ Everyone writes failure reasons on sticky notes.

3. Cluster (10 min)
   ➤ Group by theme: data, model, product, people, legal.

4. Vote (5 min)
   ➤ Each person gets 3 votes for the biggest risks.

5. Mitigate (10 min)
   ➤ For the top 3 risks, define one action + one owner each.

Document it and attach it to your Go/No-Go gate.

---

🧪 Pilot Design That Teaches Fast

A good pilot is:

• Small

• Real

• Measurable

Design it like this:

• ◾ Narrow the scope: one channel, one region, one use case.

• ◾ Define pass/fail:
  e.g., “Reduce AHT by 15% with <2% quality drop.”

• ◾ Instrument everything: log inputs, outputs, edits, outcomes.

• ◾ Create off-ramps: if metrics miss for two weeks, pause and fix.

• ◾ Schedule a decision: by day 30–45, choose: scale, iterate, or stop.

---

📈 Metrics That Matter (Beyond Model Accuracy)

Track business, experience, engineering, and safety:

• 💼 Business: cost per ticket, revenue lift, churn, cycle time

• 😀 Experience: CSAT, NPS, task completion rate, edit distance to final

• 🧩 Engineering: latency, throughput, uptime, cost per request

• 🛡 Safety: PII violations, policy flags, harmful outputs, bias checks

Pick 3–5 key metrics. Track weekly. Share a simple dashboard, not a 20-page report.

---

🧷 Governance & Guardrails Without the Drama

You can have real governance without creating a bureaucracy:

• 🧍 Human-in-the-Loop (HITL): Humans approve or edit early outputs

• 🔐 Access control: Least privilege for data, prompts, and tools

• 📜 Prompt & output logging: For audits, debugging, and learning

• 🧪 Eval suites: Red-team prompts and regression tests before each release

• 🗂 Data retention rules: Clear policies for how long data is kept and when it’s deleted

• ⚖ Bias monitoring: Simple slice checks (e.g., by region or segment)

Good governance is not paperwork—it’s how you ship safely and keep shipping.

---

👥 People & Change: Make Adoption the Default

AI fails if people don’t use it. Leslie’s adoption playbook:

• ⭐ Name an owner: One accountable lead with real decision rights

• 🧭 RACI chart: Who is Responsible, Accountable, Consulted, Informed

• 📚 Training plan:

  • 30-minute playbook

  • 60-minute hands-on session

• 🎯 Incentives: Tie team goals to the AI outcome metric

• 🔁 Feedback loop:

  • Weekly office hours

  • In-product “Was this helpful?” prompts

• 🚨 Escalation path: If it breaks, who fixes it and by when?

Ship the change playbook with the product, not after.

---

🤝 Vendor Due Diligence in One Page

When using vendors, ask for proof, not slides:

• ◾ Production references in your industry (and actually talk to them)

• ◾ Eval results on your data or a blinded sample

• ◾ Security package (SOC 2 / ISO 27001, pen test summary)

• ◾ Data terms (retention, training on your data, deletion guarantees)

• ◾ SLA & pricing (latency, uptime, cost ceilings)

• ◾ Exit plan (export, on-prem/hybrid options, model portability)

If they dodge basic proof, that’s your signal.

---

💰 Budget and Timeline Realism

Plan for three budget buckets:

1. Data & evaluation

   • Labeling, eval suite, governance work

2. Product & integration

   • Workflow, UI, logging, observability

3. Change & enablement

   • Training, documentation, champions, support

Budget both build and run (inference, monitoring, retraining).
Set a quarterly review to re-forecast based on real usage and impact.

---

🗓 The 30-60-90 De-Risk Plan

Days 0–30: Prove the Basics

• ◾ Complete the Problem Card, Value Math, and Data Factsheet

• ◾ Run a pre-mortem and fill the De-Risk Scorecard

• ◾ Build an evaluation set and simple baseline

• ◾ Design the pilot with clear pass/fail and off-ramp

Days 31–60: Run a Real Pilot

• ◾ Ship to a small user group with HITL

• ◾ Track business + safety metrics weekly

• ◾ Run two “fix-it” sprints for the top issues

• ◾ Decide: scale, iterate, or stop

Days 61–90: Prepare to Scale

• ◾ Harden infra, monitoring, and alerts

• ◾ Publish training and support plans

• ◾ Lock SLA and cost guardrails

• ◾ Write the one-page rollout plan (who, where, when)

---

📎 Templates You Can Copy (One-Pagers)

Problem Card

• Outcome: ___

• User & Moment: ___

• “Good Looks Like”: ___

• Decision Changed: ___

Value Math

• Volume: ___ × Lift: ___ × Impact/Decision: ___ − Cost: ___ = Value: ___

Data Factsheet

• Sources: ___ | Freshness: ___ | Labels: ___ | PII Plan: ___ | Access: ___

Risk Map

• Tech: ___ | Data: ___ | Product: ___ | People: ___ | Legal: ___

• Guardrails: ___ | Owner: ___

Pilot Gate (Pass/Fail)

• Target Metric: ___ → ___

• Time Window: ___

• Off-Ramp Trigger: ___

Print these. Keep them one page each. Update weekly.

---

🚫 Common Pitfalls to Avoid

• ❌ Starting without a real user
  ➜ If you can’t name the user and the step in their workflow, pause.

• ❌ Chasing SOTA (state of the art)
  ➜ You need reliable before you need fancy.

• ❌ Ignoring cost per output
  ➜ Cheap per call can be expensive at scale.

• ❌ No off-ramps
  ➜ If you never stop, you never learn.

• ❌ Compliance “at the end”
  ➜ Invite security and legal early—keep their work light but real.

---

🎯 Final Takeaway

AI doesn’t fail because the math is bad. It fails when it tries to do everything at once with no clear owner, no value math, and no off-ramps.

Rosanne Leslie’s approach is simple and strict:

Write down the value.
Check the data.
Design the workflow.
De-risk in small steps.

If you can prove a real result with a tiny pilot, and you know exactly how you’ll scale it, you’ll avoid the traps that sink most projects.

Start with the five-item checklist.
Run the 45-minute pre-mortem.
Score your idea with the one-page sheet.

Then ship something small that helps a real person do real work better.
That’s how you actually win with AI.