AI Agent ROI: Real Numbers from Real Canadian Builds (2026)

If you have read enough vendor case studies promising 10x productivity gains and 90% cost reduction, you have probably also stopped believing any of them. Here are real numbers from six production AI agent builds we shipped to Canadian businesses in 2025-2026. Builds with names removed where clients prefer it, with actual costs, actual outcomes, and the parts that did not work.

Quick summary

Across our six recent builds, average year-1 ROI was 310%. Median payback period: 4.2 months. Best case: a contractor recovered build cost in 38 days. Worst case: a SaaS founder build broke even at month 9 because the original scope underestimated integration complexity. Range matters more than averages.

Build 1: Plumbing contractor, after-hours voice agent

Build cost: $11,500 Monthly running cost: $580 (Vapi + Claude + Twilio + monitoring) Time to ship: 4 weeks

The problem: Owner was missing 25-35% of after-hours calls. Estimated $8,000-$12,000 in lost monthly revenue, plus emergency calls going to competitors.

The solution: Vapi voice agent answering 24/7, triaging emergencies vs scheduled work, booking non-urgent appointments in Jobber, texting on-call tech for emergencies.

Results (year 1):

• Recovered calls: 187/month (previously voicemail).
• New booked appointments from those calls: 64/month.
• New revenue: ~$84,000/year.
• Net ROI: ($84,000 - $11,500 - $6,960) / ($11,500 + $6,960) = 350%
• Payback: 38 days.

What did not work the first time: The first version of the agent transferred too aggressively to humans when it should have handled the call itself. We tuned the confidence threshold in week 5.

Build 2: Dental clinic, intake + scheduling

Build cost: $16,000 (PIPEDA-compliant architecture, more expensive) Monthly running cost: $480 Time to ship: 6 weeks

The problem: Front desk burning out on phone calls. Intake errors. Booking inefficiencies.

The solution: Hybrid text and voice agent for new patient intake, appointment booking, and rescheduling. Self-hosted prompt and log layer for PIPEDA. Hands off to humans for clinical questions.

Results (year 1):

• Front desk phone time reduced by 14 hours/week.
• New patient bookings up 22% (faster response).
• Cancellation handling 4x faster.
• Annual labor cost recovered: $32,000.
• Net ROI: ($32,000 - $16,000 - $5,760) / ($16,000 + $5,760) = 47%
• Payback: 8.2 months.

What did not work: The voice agent struggled with patients who had heavy accents. We added a "press 0 for receptionist" fallback that solved 80% of edge cases.

Build 3: Law firm, intake screening

Build cost: $22,000 (self-hosted, privilege-aware) Monthly running cost: $650 Time to ship: 8 weeks

The problem: Founding partner spending 5-8 hours/week on initial intake calls, many of which were not the firm's practice area.

The solution: Custom agent doing initial screening, conflict checks against existing client database, practice-area routing, scheduling consults with the right lawyer. All on Canadian-residency infrastructure.

Results (year 1):

• Partner intake time: down from 6 hours/week to 1.
• Recovered billable hours: ~5/week × 48 weeks × $450/hr = $108,000.
• Bad-fit consults eliminated (saving another ~2 hours/week of associate time).
• Net ROI: ($108,000 - $22,000 - $7,800) / ($22,000 + $7,800) = 263%
• Payback: 2.7 months.

What did not work: Conflict check accuracy was 91% in month 1; we needed 99%+. Added a second-pass verification step which fixed it but added 3 seconds of latency per intake.

Build 4: E-commerce store, customer service agent

Build cost: $14,500 Monthly running cost: $320 Time to ship: 5 weeks

The problem: 400 support tickets/week, mostly "where is my order" / "how do I return this." Two support staff burning out.

The solution: Chat + email agent handling order status, returns, refunds (up to $100 auto-approve), and shipping changes. Escalates anything complex.

Results (year 1):

• Tickets auto-resolved: 67% (was projected at 75%, came in lower).
• Support team workload: down from 2 FTE to 1.
• Salary saved: $52,000/year.
• Net ROI: ($52,000 - $14,500 - $3,840) / ($14,500 + $3,840) = 184%
• Payback: 4.3 months.

What did not work: Hallucinated return policy details in 0.3% of cases. We had to add a hard rule that the agent must quote policy verbatim from the knowledge base, not paraphrase.

Build 5: SaaS founder, lead qualification + nurture

Build cost: $19,000 Monthly running cost: $410 Time to ship: 7 weeks

The problem: 200+ inbound leads/month, no time to qualify them. Founder doing everything himself.

The solution: Multi-step agent reading inbound forms, asking 3-5 qualifying questions via email, scoring leads, scheduling demos with founder for hot ones, nurturing warm ones, declining cold ones.

Results (year 1):

• Lead qualification time: down from 12 hours/week to 1.
• Founder demo close rate: up 18% (only seeing qualified leads).
• New ARR attributed to faster response: ~$140,000.
• Net ROI: ($140,000 - $19,000 - $4,920) / ($19,000 + $4,920) = 486%
• Payback: 1.8 months.

The honest part: Original scope underestimated the CRM integration. Took 9 weeks instead of 5. Lost some early velocity. Once shipped, ROI was the strongest of any build we did this year.

Build 6: HVAC contractor, full stack (the underperformer)

Build cost: $26,000 Monthly running cost: $890 Time to ship: 9 weeks

The problem: Voice agent + lead qualifier + quote follow-up + review collection, bundled.

The solution: All four workflows, integrated with ServiceTitan.

Results (year 1):

• Recovered revenue: ~$72,000 (vs $100,000+ in our forecast).
• Net ROI: ($72,000 - $26,000 - $10,680) / ($26,000 + $10,680) = 96%
• Payback: 8.9 months.

What did not work: ServiceTitan integration was buggier than we planned. Review collection workflow underperformed because the contractor's after-job communication was unclear, so review SMS replies came through confused. Quote follow-up agent generated leads but the contractor's quote-to-close rate was the real bottleneck, which AI did not fix.

Honest lesson: AI agents amplify a working business process. They do not create one. This client's pipeline had upstream issues no agent could solve.

What these numbers mean for you

The median build paid back in 4.2 months. A 5-month time horizon is the right starting point for budgeting.

The biggest variance was scope creep. Builds that stuck to a narrow first scope shipped on time and hit ROI projections. Builds that expanded mid-flight took longer and underperformed.

The voice agent for contractors was the easiest win. If you are a Canadian contractor missing calls, this is the build with the highest probability of strong ROI. Five out of six contractor voice builds we have shipped paid back in under 90 days.

The clinic, law firm, and accountant builds were slower payback (6-12 months) but higher long-term value because they recovered partner/owner time that has no replacement cost.

The e-commerce build came in below forecast but still strong. The most accurate model for ticket-deflection ROI is "what would you pay a junior support hire to do?"

How to estimate your own ROI

Three quick math problems:

1. Time recovery: Hours/week saved × $/hour value × 50 weeks. For owner-operator time, use what the hours are worth in revenue, not your salary equivalent.
2. Revenue recovery: New booked work × close rate × average ticket size × 12 months.
3. Cost reduction: FTE displaced × annual cost. Apply a haircut (typical real-world deflection rate is 60-75% of projected).

Add them together. Subtract build cost + 12 months of running cost. Divide by total cost. That is your year-1 ROI.

If the number is under 100%, the build is probably not worth it. Over 200%, you should be moving on it.

Frequently asked questions

Why is your "real" data still pretty optimistic? Honest answer: agencies tend to ship for clients who are good fits. Selection bias is real. The numbers are accurate for what we shipped; we have also turned away builds where we did not think the ROI was there.

Can I do this in-house? You can. Expect 2-3x longer time to ship and lower production reliability for the first version. We have shipped builds in 4-8 weeks that have taken in-house teams 6-9 months to match.

What if my industry is not on this list? The pattern holds for most service businesses: contractors, clinics, salons, gyms, accountants, law firms, real estate, restaurants. The further you get from "service business with bookings or quotes," the less the pattern transfers.

How do I avoid being the underperformer build? Three things: (1) narrow scope on the first build, (2) make sure your upstream business process is working before you automate it, (3) commit to 30-60 days of tuning post-launch.

Want a real ROI projection for your business?

Free 30-minute call. We will walk through your numbers and tell you honestly whether a build is likely to pay off. Book one.