"It'll be transformative" is not a ROI calculation. Neither is "AI is the future" or "our competitors are doing it."
If you're making a case to a board, a CFO, or your own judgment about whether a custom AI build is worth $20K–$80K, you need numbers. This post gives you the framework to produce them.
The four value levers
Custom AI automation creates financial value through four mechanisms. Not every project touches all four — most have one or two dominant levers.
1. Time savings
The most common lever. A process that took humans 15 minutes per instance now takes 30 seconds. Multiply by volume, multiply by loaded hourly cost, get annual savings.
The formula: (time per task saved) × (tasks per period) × (fully-loaded hourly cost) = period savings
Key variable: fully-loaded cost, not salary. Include employer taxes, benefits, office overhead — typically 1.25–1.5× base salary for full-time employees. For the conservative case in a board presentation, use 1.3×.
2. Error reduction
Automated processes make different errors than humans — usually fewer errors on repetitive, structured tasks. Errors have costs: rework time, downstream corrections, customer service interactions, chargeback processing, compliance penalties.
The formula: (current error rate) × (cost per error) × (volume per period) × (error reduction %) = period savings
This lever is often underestimated because teams track process time but not error-related downstream costs.
3. Capacity multiplier
AI automation doesn't always reduce headcount — often it allows the same team to handle more volume. This is the right frame when: demand is exceeding supply, you'd otherwise need to hire, or growth is constrained by team bandwidth.
The formula: (headcount that would need to be added) × (fully-loaded cost per head) = avoided hiring cost
Or alternatively: (additional revenue enabled by capacity freed up) × (contribution margin) = additional profit
4. Cost avoidance
Preventing costs that would otherwise be incurred. This includes: avoiding software license costs by automating a process that was routed through expensive software, reducing external vendor costs by bringing a process in-house, or preventing error-related costs before they materialize.
Worked example 1: invoice processing
Situation: a finance team processes 200 invoices per month manually. Each invoice takes an average of 15 minutes to process (extract line items, match to PO, route for approval).
Current state cost:
- 200 invoices × 15 minutes = 50 hours/month
- Finance coordinator fully-loaded cost: $65,000/year = $31.25/hour
- Monthly cost: 50 hours × $31.25 = $1,562.50
- Annual cost: $18,750
AI system: automated extraction and matching, with human review only for exceptions (estimated 15% exception rate).
Future state cost:
- AI handles 85% automatically: 170 invoices × 1 minute each = 2.8 hours
- Human handles 15% exceptions: 30 invoices × 10 minutes each = 5 hours
- Monthly cost: 7.8 hours × $31.25 = $243.75
- Annual cost: $2,925
Annual savings: $15,825
One-time build cost: $18,000–$25,000
Payback period: 14–19 months
This is a conservative case. Error reduction (mismatched POs caught automatically vs caught in audit) adds additional value not captured here.
Worked example 2: lead qualification
Situation: an SDR team receives 500 inbound leads per month. Currently, SDRs spend 12 minutes per lead doing initial qualification research (company size, industry, tech stack, buying signals). Only about 30% pass qualification to a discovery call.
Current state cost:
- 500 leads × 12 minutes = 100 hours/month qualification research
- SDR fully-loaded cost: $90,000/year = $43.27/hour
- Monthly research cost: 100 hours × $43.27 = $4,327
- Annual research cost: $51,924
AI qualification system: automatically researches each lead and produces a qualification score with key data points. SDRs review the AI output (2 minutes each) rather than doing research from scratch. Manually review only borderline cases (estimated 20% of leads).
Future state cost:
- SDR reviews AI output: 500 leads × 2 minutes = 16.7 hours/month
- SDR manually qualifies borderline cases: 100 leads × 8 minutes = 13.3 hours/month
- Monthly cost: 30 hours × $43.27 = $1,298
- Annual cost: $15,571
Annual savings: $36,353
Secondary benefit: SDRs spend the freed 70 hours/month on outbound or discovery calls rather than administrative research. Even a 10% improvement in meetings booked from that time has significant revenue impact.
Build cost: $22,000–$35,000. Payback period: 7–12 months.
Worked example 3: support ticket deflection
Situation: a SaaS company handles 2,000 support tickets per month. Average resolution time is 18 minutes. Support staff fully-loaded cost is $75,000/year. Approximately 40% of tickets are repetitive questions that could be answered by documentation.
AI-powered support chatbot: handles the 40% of tickets that are documentation-answerable (800 tickets/month). Escalates the rest to humans with a pre-populated context summary.
Current state cost of those 800 deflectable tickets:
- 800 tickets × 18 minutes = 240 hours/month
- Cost: 240 × $36.06 = $8,654/month
- Annual: $103,848
Future state (80% deflection rate on eligible tickets):
- AI deflects 640 tickets fully (no human involvement)
- 160 tickets escalated to human with AI summary (average handling time drops to 12 minutes due to pre-populated context)
- Cost: (160 × 12 minutes = 32 hours × $36.06) = $1,154/month
- Annual: $13,842
Annual savings: $90,006
Secondary benefits: faster resolution on escalated tickets (12 min vs 18 min = 33% reduction, improving CSAT). No hiring needed as ticket volume grows proportionally with customer base.
Build cost: $30,000–$45,000. Payback period: 4–6 months.
What makes a ROI case credible to a board
Board-level presentations on AI investment often get rejected not because the ROI is bad but because the assumptions are indefensible. A few things that strengthen the case:
Use conservative assumptions. If the case works at 50% of projected savings, it's a credible investment. If it requires hitting 95% of the projection, it's a bet. Boards fund the former; they question the latter.
Document your assumptions explicitly. Every number should trace back to a data source: current time measured (not estimated), fully-loaded cost from HR, volume from ops data. "We estimate" is weaker than "ops recorded 200 invoices/month over the past 12 months."
Account for implementation costs honestly. Don't just count the build cost. Add data preparation time, internal team time for stakeholder involvement, testing and validation, and a 20% contingency for unexpected work.
Include ongoing costs. LLM API costs (typically $200–$1,500/month depending on volume), monitoring infrastructure, and annual maintenance (15–20% of build cost). A project with $90K annual savings and $5K annual operating costs is still a strong investment; hiding the operating costs undermines credibility.
Identify the risk factors. What has to be true for this ROI to materialize? If data quality is a known risk, say so and explain the mitigation. If the volume assumption depends on business growth, say so. Boards trust people who identify risks over people who pretend risks don't exist.
For a specific use case, contact us and we can help scope the build cost side of this equation. The ROI framework is yours to run with your actual numbers.