Most teams don't math out AI agent ROI before buying. Most then quietly cancel within 6 months. Here's the structured math that surfaces whether a tool will earn its keep.
The five inputs
For any agent at $X/mo subscription, calculate:
- Hours saved per week (real, measured during trial — not demo claim)
- Cleanup hours per week (when agent fails, you fix)
- Maintenance hours per month (re-tuning, vendor changes)
- Hourly cost (loaded — salary × 1.5 for fully-loaded)
- Tool cost (subscription + API overage + integration amortized)
Net monthly value = (saved - cleanup) × 4 × hourly - tool cost - (maintenance × hourly)
The 5x rule
For a buy decision, require:
- Net monthly value ≥ 5x tool cost
Why 5x:
- Demos overstate value 2-3x
- Cleanup time is always underestimated
- Vendor changes break things periodically
- The agent might not be as needed in 6 months
If math only works at 1-2x ROI on demo claims, production will be net-negative.
Worked example — Cursor Pro for an engineer
Inputs (from honest trial measurement):
- Hours saved/week: 6 (mostly Cmd+K inline edits, multi-file refactors)
- Cleanup hours/week: 1 (occasional generated code that needs rewrite)
- Maintenance/month: 1 hour
- Hourly cost: $80 (loaded for $120k/year engineer)
- Tool cost: $20/mo
Net monthly value:
- (6 - 1) × 4 × $80 = $1,600/mo saved
- Maintenance: 1 × $80 = $80/mo
- Tool: $20/mo
- Net: $1,500/mo. ROI multiplier: 75x.
This is a clear buy. The math has so much margin that even if my measurements are 50% off, ROI is still 30x+.
Worked example — $500/mo Devin
Inputs (from honest trial):
- Hours saved/week: 8 (well-scoped tickets, dependency upgrades)
- Cleanup hours/week: 2 (PRs that need rework)
- Maintenance/month: 4 hours (tuning, monitoring runs)
- Hourly cost: $100 (loaded for senior eng)
- Tool cost: $500/mo
Net monthly value:
- (8 - 2) × 4 × $100 = $2,400/mo
- Maintenance: 4 × $100 = $400/mo
- Tool: $500/mo
- Net: $1,500/mo. ROI multiplier: 3x.
This is borderline. Below the 5x threshold I'd recommend. The decision depends on whether the 8 hours saved actually materializes in your workflow OR you have more delegatable work than my estimate.
Worked example — generic "$50/mo AI writing assistant"
Inputs:
- Hours saved/week: 1 (you already have ChatGPT)
- Cleanup hours/week: 0
- Maintenance: minimal
- Hourly cost: $60
- Tool cost: $50/mo
Net monthly value:
- 1 × 4 × $60 = $240/mo
- Tool: $50/mo
- Net: $190/mo. ROI multiplier: 4.8x.
Just under the 5x threshold. Probably not worth it because you're paying $50 for marginal value over a tool you already have. Better: invest the $50 in a more differentiated tool.
The categories where ROI math usually breaks
- AI "for X" wrappers that mostly use the same model you already pay for — marginal value, full price
- Tools you'd use 1-2x/week but feel guilty canceling — sunk-cost fallacy
- Tools sold on emotional value ("less stress!", "feel productive!") without quantifiable savings
- Tools where time saved is 5-10 min/task instead of 30+ — overhead of remembering to use it exceeds value
If a tool's pitch is mostly emotional, ROI is probably weak.
The hidden ROI: avoided hires
The biggest ROI category that's hard to math: not hiring someone you'd have hired without the agent.
If Lindy at $49/mo replaces a $30k/year virtual assistant, that's not "$49/mo cost" — it's "$30k saved per year minus $588 spent". The math shifts dramatically when an agent prevents a hire.
For most teams in 2026:
- Voice CS agents → 1-2 CS hires avoided per ~50k-100k tickets/month
- AI SDRs → 1 SDR hire avoided per 200-400 prospects/week
- Inbox automation → 0.5-1 EA hire avoided
These are decisive ROI cases. The $500-2000/mo tool replaces $50-150k/year in salary + benefits.
The spreadsheet
The actual calculator is one row per tool with these columns:
- Tool name
- Monthly subscription
- Monthly API overage (estimate)
- Hours saved/week
- Cleanup hours/week
- Monthly maintenance hours
- Hourly rate (loaded)
- Computed: Net monthly value
- Computed: ROI multiplier
- Decision (buy/skip/audit-in-3-months)
Run this every time you consider a new tool. Re-run quarterly on existing tools. Cancel anything that drops below 3x ROI in actual usage.
For more frameworks see how to pick an AI agent 2026, how to budget for AI tools 2026, and our TCO calculator.