How to Measure AI ROI: A Practical Framework
Last updated: July 20, 2026
Quick answer
Measure AI per workflow, against a baseline, with three numbers: hours recovered (priced at fully loaded labour cost), error rate vs the human baseline, and cycle time. ROI = (annualized value of those gains − total AI cost) ÷ total AI cost. Well-chosen SME workflows typically pay back in 2–4 months. If you cannot name the baseline, you are not measuring ROI — you are collecting vanity metrics.
Why Do Most AI ROI Numbers Fail?
The Canadian adoption data is blunt: ~93% of executives say their organizations use AI, ~2% can show clear ROI. The gap is a measurement problem before it is a technology problem — AI deployed “everywhere” helps everyone a little in ways no ledger can isolate, and nobody recorded what the workflow cost before. The fix is structural: deploy narrowly, baseline first, count three things.
The Three Numbers That Matter
- Hours recovered. Measure the workflow's human hours for two weeks before the pilot, then after. Price at fully loaded cost (salary + benefits + overhead ≈ salary × 1.3–1.4). This is usually 70-90% of the total value.
- Error rate vs the human baseline. Not “is the AI perfect” but “is it better than the tired human on Friday afternoon.” Count corrections per hundred items, both eras. Include the cost of catching errors (review time) in the AI's cost.
- Cycle time. Quote sent in 20 minutes instead of 2 days changes win rates, not just costs. Track request-to-response time; where it moves revenue (quotes, follow-ups), estimate the conversion effect conservatively or report it separately.
The Formula, With a Worked Example
ROI = (annual value − annual cost) ÷ annual cost. A real-shaped example — quote drafting at a 20-person contractor:
- Baseline: 6 hrs/week of estimator time at $65/hr loaded → $390/week → $17,900/yr (46 working weeks)
- After AI (drafts + human review): 1.5 hrs/week → hours value recovered ≈ $13,450/yr
- Costs: $3,500 build + $150/mo run/API → $5,300 first year
- First-year ROI = ($13,450 − $5,300) ÷ $5,300 ≈ 154%, payback in ~4 months — before counting the faster-quote conversion effect
Run your own numbers in the free ROI calculator — it does this per workflow. For comparing model and tool choices once you are running, the intelligence-per-dollar lens is the sharper instrument than benchmark scores.
Which Metrics Should You Ignore?
Seats activated, prompts sent, survey-reported “time saved,” and adoption rate as a goal. All four can rise while nothing improves. The test for a real metric: it has a pre-AI baseline, it is denominated in hours, errors, or dollars, and the P&L or a customer would notice if it reversed. Measurement discipline is also a phase gate in our Canadian AI Strategy Framework — nothing scales without passing its numbers.
Frequently Asked Questions
How do you measure ROI on AI?
Measure three things per workflow, before and after: hours of human work per week, error rate against the human baseline, and cycle time (how long the customer waits). Convert hours to dollars at fully loaded labour cost, subtract the AI's total cost (build + subscriptions + API usage + review time), and divide by that cost. Anything above zero within the first year is working; most well-chosen SME workflows pay back in 2-4 months.
What is a good ROI for an AI project?
For focused workflow automation in a small or mid-sized business, a payback period of 2-4 months (roughly 300-600% first-year ROI) is a normal, achievable outcome — because the costs are small, not because the gains are huge. Broad platform rollouts measure worse: Canadian data shows about 93% of executives claim AI use while only ~2% report clear ROI, and the difference is almost always scope. Narrow workflows measure; platforms diffuse.
Why is AI ROI so hard to measure?
Three reasons: no baseline (nobody measured the workflow before AI arrived), diffuse deployment (a copilot "helping everyone a bit" cannot be isolated), and vanity metrics (prompts sent, seats active, "time saved" surveys). All three are avoidable: measure the workflow for two weeks before the pilot, deploy narrowly, and count only hours, errors, and cycle time.
What metrics should I ignore when measuring AI?
Seats activated, messages/prompts sent, self-reported time savings from surveys, and "adoption rate" as an end in itself. All four rise while nothing improves. The tell for a real metric: it is measured against a pre-AI baseline, it is denominated in hours, errors, or dollars, and someone would notice in the P&L or the customer experience if it reversed.
How do I estimate AI ROI before building anything?
Take the workflow's weekly hours × fully loaded hourly cost × 46 weeks for the annual cost of the status quo. Assume AI removes 50-70% of those hours (the typical range for pattern-heavy work with a human checkpoint), then subtract realistic costs: $1,500-$10,000 one-time build plus $50-$300/month running. Our free ROI calculator does this arithmetic per workflow, and the intelligence-per-dollar lens helps compare model choices once you are running.
Related Guides & Tools
Get AI insights for Canadian businesses
Weekly tips on ChatGPT, automation, and cost-saving strategies. No spam, unsubscribe anytime.
Related Articles
The Canadian AI Strategy Framework: 4 Phases + a Governance Spine
Intelligence Per Dollar: A Smarter AI ROI Metric
AI Adoption Gap Canada: 93% Say They Use AI, ~2% See ROI
AI consultants with 100+ custom GPT builds and automation projects for 50+ Canadian businesses across 20+ industries. Based in Markham, Ontario. PIPEDA-compliant solutions.