# Session 2 Handout: Choosing and Justifying Use Cases

<p class="meta">FAST-09 AI for Managers and Leaders · 7formations · Instructor: Wadood Chaudhary · All examples fictional · Training, not legal advice</p>

## Why this session matters

Surveys keep finding the same gap: many people report personal productivity gains, but far fewer organizations can show financial results. McKinsey's August 2026 survey reported that about 37% of respondents saw any EBIT impact from AI. The difference is usually choosing the right first use cases and measuring them honestly. You can't do them all; pick well and show the math.

## Key concepts

### 1. The use-case matrix

Score each candidate 1 to 5:

| Factor | High score means |
| --- | --- |
| **Value** | Saves real time, improves quality or service, happens often |
| **Risk** | Errors would reach customers, money, safety, or decisions about people |
| **Data sensitivity** | Needs personal information, PHI, employee data, or confidential business data |
| **Effort** | New tools, integration, a lot of training or process change |

Start with **high value, low risk**. Park high-value, high-risk ideas until controls exist. Drop low-value, high-risk ideas.

Always high risk: decisions about hiring, firing, pay, credit, insurance claims, or anything customers will rely on without a person checking. Massachusetts regulators expect accountability for these (AG advisory, April 2024; DOI Bulletin 2024-10), and states such as Colorado (SB26-189, applying from January 1, 2027) are adding rules for AI in consequential decisions.

### 2. ROI math

```text
Hours saved per week = old hours − (new hours + checking hours)
Gross annual saving  = hours saved × loaded hourly cost × working weeks
Net first year       = gross − tool cost − training time − maintenance time
```

- **Loaded cost** = wage plus benefits and employer taxes. If you lack payroll data, BLS reported $46.89 per hour worked for private industry (June 2026).
- **Include checking time.** Leaving it out is the most common error.
- **State every assumption** and its source: logs, pilot, payroll, or "estimate, to be tested."
- Simple math beats impressive guesses.

### 3. Buy, build, or wait

| Choice | When |
| --- | --- |
| **Buy** (often: use what you already pay for) | AI features already in your suite; vendor terms and data handling checked |
| **Build** | Only simple, low-risk helpers within policy (a prompt template, a shared instruction) |
| **Wait** | Value unclear, risk high, or rules and controls missing |

Waiting is a valid decision.

### 4. Reading vendor terms: five lines to find

Price and minimums · initial term · auto-renewal and cancellation notice · price increases at renewal · whether your content is used to improve the vendor's service (and how to opt out).

## Worked example (fictional, Pioneer Valley Mutual)

Four adjusters spend 3 h/week each on claim summaries. With AI, drafting takes 1 h plus 0.5 h checking. Loaded cost $48/h, 48 weeks.
Gross: 4 × 1.5 × $48 × 48 = **$13,824**. Tool: 4 × $30 × 12 = $1,440. Training: 4 × 3 h × $48 = $576. **Net: $11,808.**
Weakest assumption: checking stays at 0.5 h when harder files arrive. Re-measure in days 31 to 60.

## Checklist

- [ ] 8 to 10 candidates scored on all four factors.
- [ ] Two pilots chosen with value and risk reasons.
- [ ] High-risk ideas parked, with what would unpark them.
- [ ] ROI includes checking, training, and tool costs.
- [ ] Every assumption has a source; the weakest is named.

## Do and don't

| Do | Don't |
| --- | --- |
| Start with tools you already own | Buy the flashiest product first |
| Measure a baseline from logs | Guess the "before" time |
| Keep people deciding about people | Pilot AI that rejects applicants or denies claims |
| Say "wait" when it's right | Start a pilot just because a trial is free |