AI Cost Per Task: The 2026 SMB Number That Matters

Most Ottawa owners pick AI tools the way they pick a phone plan.

They look at the monthly fee, squint at the savings, and trust whatever the salesperson says. That worked in 2024. In 2026 it's the wrong question.

The number that decides whether AI works for your business isn't the subscription. It's the AI cost per task. And the math is moving fast. Well-scoped SMB workflows are now landing at $0.02 to $0.20 per task — that's down roughly 80% in the last 18 months. Voice agents are around $0.20 a minute. FAQ chats are pennies. The owners who track this number are quietly compounding. The ones still staring at the monthly invoice are about to be out-priced by their own competitors. This post is the playbook — what CPST actually means, the three numbers to track, and where the math works first.

What "AI Cost Per Task" Actually Means

Cost per Successful Task — CPST — is your total AI bill divided by the number of outcomes the AI actually delivered. Not attempts. Not tokens. Outcomes.

An outcome is a booked appointment, a qualified lead, a closed FAQ ticket, a recovered missed call. Anything you can point to and say "that would've taken a human 5 minutes."

The math gets clear fast. A $200/month chat agent that closes 10,000 customer questions costs you $0.02 per question. A voice receptionist at $0.20 per minute averages about $1.50 per 8-minute call. A scheduling agent at $50/month booking 500 appointments costs $0.10 per booking. Compare that to a $20/hour admin handling 6 inquiries an hour — that's $3.33 per inquiry, before benefits and overhead.

So-what: when your CPST drops below your human labour cost on the same task, you've stopped debating whether to automate. You're now debating how fast.

The Three Numbers You Actually Need to Track

You don't need a dashboard. You need three numbers in a spreadsheet, updated monthly.

1. Cost per Successful Task. Total AI bill ÷ number of successful outcomes. Ignore attempts, tokens, sessions, "engagements." Outcomes only. If you can't define what counts as a success in one sentence, you're not ready to deploy.

2. Deflection Rate. The percent of attempts the AI completed without handing off to a human. Most Ottawa SMBs land at 30 to 50% in the first 60 days, then climb to 70 to 80% as the prompts get tuned. Below 30% and you're paying for software that creates more work than it removes.

3. Payback Months. When does (monthly hours saved × your hourly cost) cover the AI bill? Typical answer in 2026: 3 to 9 months. Anything past 12 is a scope problem, not an AI problem. This is the same pattern you see in the first 90 days of AI deployment — month 1 is messy, month 3 is where the math turns.

So-what: pretty dashboards don't matter. These three do. If your vendor can't show you all three after 30 days, you bought a toy.

Where the Math Works First for Ottawa SMBs

Start where the gap between AI and human cost per task is widest. Not where the AI sounds coolest.

Missed-call recovery. AI voice receptionist at ~$1.50 per call vs an admin who chases the missed call later at ~$25 in loaded labour. That's a 94% reduction. Best for HVAC, plumbing, dental, and any service business where after-hours calls go to voicemail. The same math drives $0.20-per-minute voice AI — and it's why this is usually the first place to deploy.

Inbound lead reply. AI at $0.05 per reply vs $4 to $8 in staff time. Speed matters more than volume here — calling within 60 seconds of a lead beats calling at 60 minutes by a lot. If you're sitting at a 47-hour reply time, the CPST gap looks small but the conversion gap is enormous.

FAQ chat and scheduling. AI at $0.02 to $0.10 per resolved question vs $1.50 in staff reply time. Hours, services, pricing, parking, where to send the deposit — none of it needs a human. Layered with AI scheduling, the same agent can hand off booked appointments straight to your calendar.

So-what: don't lead with AI marketing. Lead with AI operations. The 4× ROI gap between operations and marketing AI is real and the CPST math is why.

The Honest 2026 Payback Window

Here's what actual deployments are running across Canada in 2026.

3 months if you keep scope tight: one workflow, one tool, one team member owning it. Voice receptionist for an Ottawa HVAC company replacing a $400/month answering service is the cleanest case.

6 months for two-channel deployments with measurement turned on. Voice plus SMS plus a basic FAQ widget for a clinic or a multi-location dental office.

9 months for full-stack agent deployments — the AI handles inbound calls, replies, scheduling, follow-up, and review requests across operations.

12+ months usually means scope creep, the wrong tool, or a vendor charging too much for too little. It almost never means "AI doesn't work for my business."

Recent 2026 data is consistent: 68% of SMBs that deployed agents reported around $84,000 a year in operational savings on average. The other 32% didn't measure. Don't be the 32%.

So-what: pick one workflow, lock down the three numbers, and run it for a quarter. The owners who do this are looking at sub-$0.10 task costs by Q3 — and a labour line item that's stopped growing for the first time in years.