The Honest ROI Conversation

Every week, a small-business owner sends me an article claiming that AI saves forty percent of working time, or that AI pays back its cost in three months, or that companies adopting AI see thirty percent productivity gains. The articles usually have a chart. The chart usually has confident gradient bars.

I want to be careful here. I am not saying AI doesn't deliver returns. It does, often significant ones. But the numbers being passed around are mostly fiction, and the conversation they push small-business owners into is the wrong one. Let me tell you what an honest ROI conversation looks like, because the alternative — believing the chart — leads to decisions you'll regret.

The dominant ROI claim is: AI saves X percent of working time on tasks Y. By extrapolation, an organisation that adopts AI broadly should expect a productivity uplift of X percent across knowledge work, which translates to enormous bottom-line gains.

Three things are wrong with this story, and they're worth understanding individually.

The time-savings number is measured wrong. Most of the studies behind the forty percent numbers measure something narrow — the time to complete a specific task, in a controlled setting, when the AI works. They do not measure the time to verify the output, the time to fix the bad ones, the time to onboard new users, the time to maintain the system. The headline number is the time saved on the task; it is not the time saved by the business.

The extrapolation is invalid. Time saved on one task does not roll up into business productivity by simple addition. A team that saves twenty minutes a day on email drafting does not produce twenty more minutes of value-creating work. They produce a meeting that's twenty minutes longer, or a coffee break, or twenty minutes of higher-quality but still-bounded output. The relationship between micro-time-savings and macro-productivity is much weaker than the spreadsheet suggests.

The cost side is undercounted. AI adoption costs more than the subscription. It costs change management, training, review time, tooling sprawl, mistakes-in-production, customer-trust hits when the AI gets it wrong publicly, and the opportunity cost of attention that could have gone to non-AI work. Almost none of this shows up in the AI ROI calculation.

The framing I'd encourage instead is closer to how you would think about any other capability investment, with one important addition for AI.

Instead of asking what's the ROI of AI?, ask three specific questions about each candidate deployment.

What does this workflow currently cost, in time and quality? Not the average. The actual cost, measured. Time per task, frequency, error rate, downstream consequences of errors, opportunity cost of the people doing it. You need this number to know whether any improvement matters.

What is the realistic cost of deploying AI here? Not the subscription. The total. Build time, review time, training time, ongoing maintenance, the cost of mistakes during the climb up the trust ladder. Be generous with this estimate — it is always larger than the team initially thinks.

What is the realistic improvement, after the trust ladder has been climbed? Not the demo number. The number after the system has been running for six months, accounting for edge cases, errors caught and not caught, drift, and the actual rather than imagined behaviour of users.

If the difference between the cost-before and the cost-after, minus the deployment cost, is positive over a horizon you actually care about, the deployment is worth doing. If you cannot answer all three questions, the deployment is speculative, and you should run it as a deliberate learning bet rather than a confident ROI play.

An honest ROI assessment, done this way, looks much less impressive than the gradient charts. You will discover, often, that the deployment you were excited about saves your business a real but unspectacular amount of money. You will also discover, sometimes, that a deployment you weren't excited about saves an enormous amount because the baseline cost was higher than anyone had measured.

Both of these are healthy outcomes. The first protects you from spending money on AI deployments that won't pay back. The second points you at the workflows where AI is genuinely transformational for your specific business. Neither comes from believing the industry-wide headline number.

The deepest ROI of AI, in my experience, is usually not in the tasks people talk about. It is in workflows that nobody had thought about, that were costing the business orders of magnitude more than anyone realised, that AI happens to handle well. You find those by doing the measurement honestly. You miss them by chasing the headlines.