How to Build an AI ROI Business Case (Template Included)

The difference between an AI pilot that gets approved and one that gets shelved is a credible business case. This article provides a complete template for building a board-grade AI ROI proposal—with all six sections you need, worked examples, and the exact numbers to include. The template works for any AI deployment: customer support, claims processing, back office. A CFO can follow it in 45 minutes and present it to the board in 15 minutes.

The Six-Section Business Case Template

Every board-ready AI ROI business case has six sections. Use this structure:

Section 1: Problem Statement (1 page)

Define the business problem you're solving and quantify the baseline cost.

What to include:

• What's the current manual process?
• What does it cost today (in dollars)?
• What's the volume (daily/monthly/annual)?
• Why are we looking at AI? (Capacity constraint? Cost pressure? Quality issue?)

Worked example:

Current customer support operation:

Our customer support team currently handles 2,400 support tickets per month across web, email, and chat. The team consists of 12 CSRs, each fully loaded at $52,000 per year.

Current cost structure:

• Total annual cost: 12 × $52,000 = $624,000
• Average tickets per person per month: 200
• Baseline cost per resolved ticket: $21.67

Why this matters: At this cost, a 10% improvement in efficiency saves $62,400/year. Scaling from 2,400 to 4,000 tickets per month (40% growth) would require hiring 5 more people at a cost of $260,000/year. AI offers a way to handle that growth without headcount.

Section 2: Current Cost Baseline (1 page)

Walk through the math that gets you to your baseline cost per outcome. A CFO will challenge this, so make it bulletproof.

What to include:

• Total annual cost for the function (payroll, benefits, overhead)
• Productive hours per employee per year
• Average work items handled per employee
• Cost per outcome calculation

Worked example:

Cost per ticket calculation:

| Item | Value | |---|---| | Base salary per CSR | $42,000 | | Benefits (health, 401k, etc.) | $6,000 | | Overhead (facility, tools, training, management) | $4,000 | | Fully loaded cost per CSR | $52,000 | | Productive hours per year (2,000 hrs - 100 hrs vacation/training) | 1,900 | | Productive minutes per year | 114,000 | | Average handle time per ticket | 6.5 minutes | | Tickets handled per CSR per year | 17,538 | | Cost per ticket (CSR cost only) | $2.97 | | Blended cost (accounting for management, QA, rework) | $21.67 |

Note: The 7x difference ($21.67 vs. $2.97) comes from: 30% of tickets are customer callbacks (rework), 25% overhead for management/QA, 15% training/onboarding, and 10% slack time.

Section 3: Proposed AI Solution (1-2 pages)

Describe the AI agent, its scope, and how it works. Make it concrete enough that a CFO understands what they're funding.

What to include:

• What AI model will you use? (Claude, GPT-4, etc.)
• What will the agent do? (First-touch resolution on routine tickets, triage, escalation to human)
• What's the escalation rate? (What percent of cases still go to human?)
• What integrations are required? (CRM, ticketing system, knowledge base)

Worked example:

Proposed AI customer support agent:

Agent scope: The AI agent will handle first-contact resolution on routine customer support tickets: account issues, password resets, subscription questions, billing inquiries. It will escalate complex issues (refund disputes, account security, custom use cases) to human support.

Agent architecture:

• Model: Anthropic Claude 3.5 Sonnet
• Input: Customer message + ticket history + company knowledge base
• Output: Proposed resolution or escalation with reason
• Integrations: Zendesk (ticketing), Stripe (payment lookup), Salesforce (customer data), internal knowledge base

Accuracy and escalation expectations (based on pilot):

• First-touch resolution rate: 76% (agent resolves without human)
• Escalation rate: 24% (requires human review or handoff)
• Accuracy of escalated decisions: 94% (human agrees with escalation reasoning)

Human review workflow: Escalated tickets are routed to a human CSR, who either approves the agent's recommendation (2 min), modifies it (5 min), or handles from scratch (8 min). Average review time: 4 minutes.

Section 4: Year-1, 2, and 3 P&L Impact (2 pages)

This is the heart of the business case. Show the financial impact clearly.

What to include:

• Year-by-year cost savings
• Year-by-year cost of running the agent
• Net impact (savings minus cost)
• Payback period

Worked example:

Year 1 P&L impact:

| Item | Monthly | Annual | |---|---|---| | Current cost structure | | | | Current tickets per month | 2,400 | 28,800 | | Current cost per ticket | $21.67 | | | Current monthly cost | $51,984 | $623,808 | | | | | | With AI agent (Year 1) | | | | Tickets handled by AI (76% of 2,400) | 1,824 | 21,888 | | Tickets handled by human (24% of 2,400) | 576 | 6,912 | | | | | | AI cost breakdown | | | | API inference cost | $0.12/ticket | $2,626/month | | Retries and overhead | 0.05/ticket | $1,092/month | | Infrastructure and observability | 0.18/ticket | $3,276/month | | Subtotal: AI cost per ticket | $0.35/ticket | $6,994/month | | | | | | Human review cost | | | | Escalated tickets: 576/month | | | | Average review time per ticket: 4 min | | | | Monthly human review hours: 38.4 | | | | Cost per hour (loaded): $27.50 | | | | Subtotal: Human review cost | $1,056/month | $12,672/month | | | | | | Total AI + human cost | $8,050/month | $96,600/month | | | | | | Savings vs. baseline | $43,934/month | $527,208/month | | Savings vs. baseline (%) | 85% | 85% | | | | | | One-time investment | | $75,000 | | (Build, test, integrate, train) | | | | | | | | Year-1 net impact | | $452,208 | | (Annual savings - one-time cost) | | |

Year 2 P&L (assume 10% ticket volume growth):

| Item | Annual | |---|---| | Tickets per year | 31,680 | | AI + human cost (same per-unit rate) | $106,260 | | Baseline cost (would require hiring) | $686,188 | | Net savings | $579,928 | | Recurring annual savings | $579,928 |

Year 3 P&L (assume 5% growth):

| Item | Annual | |---|---| | Tickets per year | 33,264 | | AI + human cost | $111,566 | | Baseline cost (would require hiring) | $720,981 | | Net savings | $609,415 | | Recurring annual savings | $609,415 |

Section 5: Risk and Sensitivity Analysis (1 page)

Show what happens if your forecast is wrong. This is where you build credibility with the CFO.

What to include:

• Best case, realistic case, worst case
• Key assumptions you're uncertain about
• Sensitivity to key variables

Worked example:

Sensitivity analysis:

What happens if key assumptions change?

| Scenario | Assumption Change | Year-1 Net Impact | Sensitivity | |---|---|---|---| | Realistic (base case) | — | $452,208 | — | | Conservative | Escalation rate +5% (24% → 29%) | $398,544 | -12% | | Conservative | AI accuracy -10% | $385,922 | -15% | | Conservative | Both above | $332,258 | -27% | | Optimistic | Escalation rate -5% (24% → 19%) | $505,872 | +12% | | Optimistic | AI accuracy +10% (higher first-touch) | $539,494 | +19% | | | | | | | Payback period (base case) | $75,000 investment ÷ $43,934/month | 1.7 months | | | Payback period (conservative) | $75,000 investment ÷ $33,188/month | 2.3 months | | | Payback period (optimistic) | $75,000 investment ÷ $54,656/month | 1.4 months | |

Key risks and mitigation:

1. Escalation rate is higher than 24% in production. Mitigation: Run a 6-week pilot at production scale (500+ tickets) before full deployment. Measure actual escalation rate.
2. AI accuracy degrades with domain drift. Mitigation: Monthly accuracy review and prompt retraining. Set a 90% minimum accuracy threshold.
3. Human review overhead is higher than forecast. Mitigation: Log review time for every escalated ticket. Alert if it exceeds 5 minutes average.
4. Integration with Zendesk takes longer than expected. Mitigation: Allocate 6 weeks for integration. Budget $20,000 for technical resources.

Section 6: Decision Recommendation (1 page)

End with a clear ask and a decision framework.

What to include:

• What are you asking the board to approve? (Budget? Headcount reallocation? Vendor contract?)
• What's the decision framework? (If savings >$300k, approve?)
• What are the next steps?
• What are the go/no-go milestones?

Worked example:

Recommendation:

We recommend approving the $75,000 investment to deploy an AI customer support agent. The agent will reduce cost per ticket by 85% and generate $452,208 in net savings in Year 1 alone, with payback in less than 2 months.

Decision framework:

• Approve if: Savings forecast >$300k and payback <6 months. ✓ Both met.
• Approve with caution if: Savings $200-300k or payback 6-12 months.
• Reject if: Savings <$200k or payback >12 months.

Next steps (30-60 days):

1. Conduct a 4-week production pilot: Run agent on 20% of tickets, measure accuracy and escalation rate.
2. Secure buy-in from operations: Brief support team on agent workflow and escalation process.
3. Complete Zendesk integration: Work with IT to connect agent to ticketing system.
4. Finalize vendor contract: Negotiate API pricing and SLA with Anthropic.

Go/no-go milestone (end of pilot):

• If pilot achieves 70%+ first-touch resolution and <25% escalation, proceed to full deployment.
• If either metric misses, run a second 4-week pilot with prompt refinements.
• If second pilot still misses, the business case is no longer viable.

Common Mistakes to Avoid

1. Forgetting hidden cost. Your AI cost is not just API bills. Include retries, infrastructure, observability, human review, integration overhead. If you forget one layer, your ROI is wrong.

2. Assuming perfect accuracy. Don't forecast 95% first-touch resolution based on lab testing. Production will be 70-85%. If you're optimistic and reality delivers pessimistic, the board will blame you.

3. Burying assumptions. The CFO will challenge your assumptions. State them clearly (escalation rate is 24%, human review time is 4 minutes). Make them easy to audit.

4. Ignoring headcount reallocation. If you save 4 FTE in cost, what happens to those 4 people? Are they redeployed? Laid off? This is a political question the CFO cares about.

5. Missing the one-time cost. Many teams underestimate integration, training, and pilot costs. Budget $50,000-150,000 depending on complexity.

6. Measuring ROI only in Year 1. Show Year 2 and 3. The real story is Year 2 and 3 recurring savings.

Using This Template

Download the template (available as a Google Sheets calculator and a PowerPoint deck) and fill in your numbers. The template auto-calculates sensitivity analysis, payback period, and charts.

For a complete ROI methodology, see How to Actually Measure AI ROI (With Numbers). For cost benchmarks in your vertical, see the cost-per-outcome benchmarks articles for customer support, claims, and back office.

Most CFOs will approve an AI investment if the payback period is <6 months and the Year-1 net impact is >$300,000. Build your business case to those standards, and you'll get approval.