AI Employee ROI: The Payback Math for Small Business

Most AI sales pitches open with a percentage. Cut costs by 40 percent. Save 20 hours a week. Those numbers are unfalsifiable until you swap in your own, so here is the arithmetic we walk through with owners in Burnaby and New Westminster before anyone signs anything.

Payback period, in months, is:

build cost divided by ((monthly hours saved x loaded hourly cost) minus monthly running cost)

That's the whole model for AI employee ROI. Four inputs. If you can't fill in all four with numbers you'd defend to your accountant, you're not ready to buy yet.

The four inputs and where to get them

• Hours saved per month. Counted, not guessed. Log two weeks of the actual task: how many calls come in after hours, how many quote requests get retyped into the CRM, how many review replies someone writes by hand.
• Loaded hourly cost of whoever does that work now. Statistics Canada put average hourly wages among employees at $37.17 in July 2026, up 2.8 percent year over year (Labour Force Survey, July 2026). That's the wage line only. Your loaded cost adds CPP, EI, WorkSafeBC premiums, vacation, and the slice of a manager's time spent supervising.
• Monthly running cost. Model tokens, phone or SMS minutes, the software the agent plugs into, and monitoring.
• Build cost. The one time work: integrations, writing and testing prompts, building an evaluation set, staff training, and the inevitable two weeks of tuning after go live.

The running cost is usually smaller than owners expect

Token pricing is public, so you can compute this instead of trusting a quote.

Anthropic's pricing documentation carries a worked example: roughly 10,000 support ticket conversations, averaging about 3,700 tokens each, run on Claude Haiku 4.5 at $1 per million input tokens and $5 per million output tokens, come to approximately $37 in total (Claude platform pricing, figures current as of August 2026). A step up, Claude Sonnet 5, is published at $2 per million input tokens and $10 per million output tokens as of August 2026, and that same page states the launch introductory rate is now the standard price rather than rising to $3 and $15.

Two levers move the bill. Prompt caching bills a cache read at 0.1 times the base input price, which matters a lot when every call ships the same long system prompt, price list and FAQ. And the Batch API takes 50 percent off both input and output for anything that doesn't need an instant answer, like overnight review replies or a nightly pass over unpaid invoices.

There's a catch the marketing decks skip. The pricing page notes that newer models use a tokenizer producing roughly 30 percent more tokens for the same text, while Sonnet 4.6 and earlier use the previous one. Cheaper per token is not automatically cheaper per conversation. Price the work against your own transcripts.

A worked example, illustrative only

The numbers below are illustrative. They are not a result from any Autana client.

Say a four person shop loses 30 hours a month to after hours calls, callbacks, and copying job details into the calendar. Put the loaded cost of that time at $45 an hour, which sits above the national average wage once employer costs are added. That's $1,350 a month of recovered capacity. Suppose the AI employee costs $80 a month to run, all in, and $6,000 to build and integrate.

Payback = 6,000 / (1,350 - 80), or about 4.7 months.

Now stress it. If you actually save 12 hours a month instead of 30, the same build takes about 13 months to pay back. Drop the loaded rate to $32 and it stretches further. The input that decides your AI employee ROI is hours saved, and that's the one people invent. Measure it first.

For trades, clinics and salons, the bigger figure often sits on the revenue side rather than the cost side: jobs booked that would have gone to whoever answered first. That's real. It's only credible if you recorded your baseline missed call rate before switching anything on.

Where this doesn't apply

The evidence is mixed, and you should hear that part too.

A Statistics Canada study by Jiang Li and Huju Liu, released April 22, 2026, found AI adopting firms had 16.8 percent higher productivity than non adopters. Adjusting for the productivity those firms already had before adopting dropped it to 10.2 percent. Controlling for complementary capabilities such as R&D, cloud computing, data analytics and workforce training dropped it to 5.1 percent, and it was no longer statistically significant. The authors conclude there is no statistically significant direct association between AI adoption and productivity (The Role of Complementary Capabilities in AI Adoption and Productivity). The gains cluster in firms that were already competent at digital work. If your job data is scattered across three notebooks and a group chat, AI isn't the missing piece.

Plenty of businesses correctly opt out. In the second quarter of 2026, 19.2 percent of Canadian businesses reported using AI in the prior 12 months, up from 12.2 percent a year earlier, but 40.0 percent said AI simply isn't relevant to what they produce or deliver, 13.4 percent cited cybersecurity or privacy concerns, and 10.6 percent named cost (Statistics Canada, second quarter of 2026).

Then there's the part that quietly eats the savings. Federal, provincial and territorial privacy regulators published joint principles for generative AI in December 2023 stating that organizations remain accountable for decisions made with AI tools, must be transparent about the tool's role, and cannot rely solely on automated output without human review (OPC principles for generative AI). If your use case needs a person checking every response, your hours saved number is the review time you save, not the whole task. Budget for that honestly.

Small firms also underinvest in the surrounding work. Among Canadian businesses with 1 to 4 employees, only 24 percent provided AI related training, against 68.1 percent of businesses with 100 or more employees, per that same StatCan release. Untrained staff route around a tool they don't trust, and the payback never shows up.

Measure the before, not just the after

The NIST AI Risk Management Framework organizes this work into four functions: Govern, Map, Measure and Manage (NIST AI RMF). Measure is the one small businesses skip. Two weeks of baseline data costs you nothing and turns your payback calculation from a sales argument into a finance decision.

Pick one task. Time it. Run the arithmetic. If payback lands under a year on conservative inputs, build it. If it doesn't, don't.

Sources

• Statistics Canada. "Labour Force Survey, July 2026." The Daily, August 7, 2026. https://www150.statcan.gc.ca/n1/daily-quotidien/260807/dq260807a-eng.htm
• Anthropic. "Pricing." Claude Platform Docs, retrieved August 2026. https://platform.claude.com/docs/en/about-claude/pricing
• Statistics Canada. "Analysis on artificial intelligence use by businesses in Canada, second quarter of 2026." Catalogue 11-621-M, June 11, 2026. https://www150.statcan.gc.ca/n1/pub/11-621-m/11-621-m2026010-eng.htm
• Li, Jiang, and Huju Liu. "The Role of Complementary Capabilities in AI Adoption and Productivity: Firm-Level Evidence from Canada." Canadian Public Policy (forthcoming), released by Statistics Canada April 22, 2026. https://www150.statcan.gc.ca/n1/pub/36-28-0001/2026004/article/00002-eng.htm
• Office of the Privacy Commissioner of Canada and provincial and territorial privacy commissioners. "Principles for responsible, trustworthy and privacy-protective generative AI technologies." December 2023. https://www.priv.gc.ca/en/privacy-topics/technology/artificial-intelligence/gd_principles_ai/
• National Institute of Standards and Technology. "AI Risk Management Framework (AI RMF 1.0)," NIST AI 100-1, January 2023. https://www.nist.gov/itl/ai-risk-management-framework

If you want to run these numbers against a real task in your Metro Vancouver business, we'll do it with you. Autana Solutions builds AI employees for small and mid sized companies around Burnaby and New Westminster, and a free call is enough to work out whether the payback is actually there. If it isn't, we'll say so.