The Boring AI Wins. The Exciting AI Gets Cancelled.

It scared a lot of owners off spending anything. Which is a shame, because the timing was odd. A capable AI employee-equivalent now costs about $21,500 a year (roughly $1,792 a month). A person doing the same repetitive work costs about $80,000 once taxes, benefits, training and turnover are counted. That gap doesn't close because a headline was scary.

So I went back and read the report. The footnotes tell a different story.

The figure comes from MIT NANDA's The GenAI Divide: State of AI in Business 2025, published in July 2025. It's built on interviews with people at 52 organizations, survey answers from 153 senior leaders and a review of more than 300 public AI projects. The executive summary does say "95% of organizations are getting zero return." The authors also say, fairly, that their numbers are directional, drawn from interviews rather than audited company results.

Keep reading and two findings jump out.

First, how you get the tool matters. Projects done with an outside partner made it into daily use about twice as often as the ones companies built themselves. The report is careful to say that's a correlation, not proof. Still, if you're a 15-person firm, it's a strong hint that your first AI project shouldn't be a homegrown science experiment.

Second, where you point it matters. Roughly half of generative AI budgets went to sales and marketing, the visible, demo-friendly stuff. MIT found back-office automation often delivered the better return. The money was sitting in invoicing, document handling and admin, and most budgets walked right past it.

So the lesson isn't "AI doesn't work." It's "most companies bought the wrong thing for the wrong job, or tried to build it themselves."

Adoption isn't really the question anymore. In Google Cloud's ROI of AI study (September 2025, 3,466 senior leaders in 24 countries), 52% of executives said their organization was actively using AI agents, and 39% had launched more than ten. 74% said they'd seen a return within the first year.

Closer to home, the numbers are smaller but they point the same way on both sides of the border. Statistics Canada found 19.2% of Canadian businesses used AI to produce goods or deliver services in the year to mid-2026, up from 6.1% in 2024. In the U.S., the Census Bureau put AI use at 19.8% of businesses in May 2026, counting any business function.

But look at what happens when the AI is asked to run on its own. Gartner surveyed 360 IT application leaders in mid-2025 about fully autonomous agents, the kind that pursue a goal without a person checking. The answers were cautious, and for good reason.

Only 15% were even trying fully autonomous agents. Only 13% strongly agreed they had the governance to manage them. That's the exciting AI: the agent that's supposed to run a whole department. It demos beautifully, nobody can audit it, and it gets quietly shelved in the next budget review.

The businesses getting real AI cost reduction have one unglamorous habit. They give AI the dull, repetitive work first, they make it reliable there, and they keep a person on anything that needs judgment.

Picture a 6-person bookkeeping office. The exciting pitch is an AI that "manages client relationships." The boring version reads incoming receipts, matches them to bank lines, and flags the twelve that don't match for a person to review. Only one of those will still be switched on in a year.

1. 01

   Find the boring, costly job

   Repetitive, rules-based work that eats hours: data entry, invoicing, lead follow-up, inbox triage. Not the moonshot.

2. 02

   Buy and deploy before you build

   MIT's data says partner-built tools reach daily use about twice as often. Unless AI is your product, building from scratch is the slow, expensive road.

3. 03

   Write the No-List

   Decide up front what AI never does without a person: refunds, legal wording, the sensitive client. One confident mistake on the wrong task wipes out a lot of quiet wins.

4. 04

   Ask for reliability, not novelty

   If you can't see what the tool did and why, it's the wrong tool. Dependable beats clever.

Do this in order and the math gets less scary and more obvious. You're not replacing your team. You're taking the drag off them so their hours go to the work that actually needs a person.

That's what we build at QLIA: the unexciting kind of AI employee that does one job well, logs everything, and hands the hard calls to your people. It's not much to film. It is the kind that pays for itself.

Before you buy the tool that looks most impressive, find the job that's quietly costing you the most. Start there.