Picture a typical scene. A company spends six months running an AI pilot: a support chatbot, a model for scoring applications, product description generation. The demo gets applause, management is pleased, the budget is spent. And a quarter later it turns out that nobody counted the hours saved, the CRM integration ate just as much again, and the conversion increase coincided with a seasonal promotion — and it's unclear what AI had to do with it at all. The money is gone, the effect has dissolved. For a deeper dive into this topic, see AI for Professionals.
Below is a breakdown of typical failures that make AI implementation ROI negative, and what to do differently. This isn't about "which AI to choose" — it's about how to count the money before it's spent.
Failure #1: The pilot is launched without defining a metric
What it looks like: the project presentation says "improve support efficiency with AI." What that means in dollars is not stated.
Why it happens: AI sells itself. The demo is beautiful, the model's language is convincing, and the effect seems obvious. But "obvious" is not a metric.
How it ends: after the pilot, nobody can prove that things got better, and the project is quietly closed — or, worse, scaled up because "it's a shame to abandon it."
How to fix it: before you start, lock in one primary metric in money terms and one validation metric. For support, it's not "satisfaction" but, for example, the cost of handling one request. For sales, it's not "response speed" but the conversion of a lead into a deal and the average order value. For a deeper dive into this topic, see sales, it's.
Failure #2: AI's effect is confused with general growth
What it looks like: after launching a recommendation system, revenue grew by 12%. The team celebrates. But the same month had an advertising campaign, and before that there was a seasonal slump.
Why it happens: the temptation to claim all the growth is strong, and isolating one tool's contribution is work.
How it ends: the economics of AI projects are built on inflated numbers, the next implementation promises even more, and disappointment comes later and hurts more.
How to fix it: you need a control group. If possible, part of the traffic or leads is processed without AI — and you compare the two groups over the same period. If a control isn't possible, use a before/after comparison over a comparable period and subtract everything that changed in parallel: promotions, seasonality, price changes.
This is the same logic as in A/B tests and forecasts: without a baseline, any number is self-deception.
Failure #3: Hidden costs eat the calculation
What it looks like: you calculated the cost of the model and development and got a payback of 4 months. In reality, the project keeps draining money for another year.
Why it happens: the budget didn't include things that aren't visible in the demo.
What's usually forgotten:
| Cost item | What it means in practice |
|---|---|
| Data | Labeling, cleaning, standardizing — often more expensive than the model itself |
| Integrations | Connecting to CRM, ERP, internal systems, API support |
| Support | Quality monitoring, model updates, reacting to "gone off the rails" responses |
| People | Training employees, changing processes, resistance |
Artificial intelligence payback is calculated on total cost of ownership, not the pilot's price. The real payback period is often 2–3 times longer than the initial estimate — precisely because of these items.
Failure #4: No business case, only technical desire
What it looks like: "Let's try LLMs, everyone's trying them." Six months later — the question "why did we do this?"
Why it happens: the technology is interesting in itself, and that substitutes for the question about money.
How it ends: the budget is spent, and a business case for AI never appeared. The company concludes "AI doesn't work" — though it was the calculation that didn't work. For a deeper dive into this topic, see AI Business Automation.
How to fix it: before launching, answer five questions.
- What specific problem are we solving, and how much does it cost now?
- Which metric will show that things got better, and where do I see it?
- What will change in the process, and who will use it?
- What does owning the solution cost per year, including data and support?
- What do we do if after N months the metric hasn't moved?
If there's no answer to the fifth question, it's not a project but an experiment with no exit.
How to calculate ROI before you start
A simple formula worth keeping in mind:
ROI = (savings + additional profit − total costs) / total costs × 100%
Let's take an example. Automating request handling: 2,000 requests per month, manual handling takes 12 minutes, an operator's hour costs $8. AI removes 40% of the routine.
- Savings: 2,000 × 12 min × 40% = 160 hours ≈ $1,280 per month.
- Costs: development $5,000 + support and data $400/month.
- Development payback: about 5 months (a bit longer including support).
The numbers are illustrative, but the main point is clear: without savings in dollars and without accounting for support, ROI isn't calculated. And check that 40% of the routine is a realistic share, not a desired one.
Learning paths: what to choose
The topic of implementing AI in a company is structured so that theory without practice doesn't work, and practice without a framework is expensive.
- On your own through articles and case studies. Cheap, but there's no feedback: no one to point out that your calculation is built on an inflated metric. Mistakes are discovered after the budget is spent.
- Your own pilot "in the wild." The best school, but the price of a mistake is the company's real money. Suitable if you have a buffer and a mentor.
- A structured course. Provides a framework: how to calculate, which metrics to take, where the traps are. Costs time and money, but mistakes are worked out on training calculations, not on the budget.
The course AI Business Transformation covers exactly these calculations with examples — from maturity diagnostics to ROI and a roadmap — and gives feedback on your decisions so that the pilot is counted in money before the start, not after.
If you're going to eliminate pilots that don't pay off, you should start not with choosing a model but with one question: how many dollars will this project bring in or save, and when. Everything else is details that get resolved after an honest answer.