AI in Your Company: How to Cut OPEX Measurably, Not Experimentally

According to EY's third-edition AI in Business report (October–November 2025, ~499 mid-to-large firms), 49% of companies reported disappointment with AI results — and 17% said that knowing the outcome in advance, they would not have pursued implementation at all.

The pattern is consistent: AI initiatives get launched, pilots run, results get presented to the board — and then the system never reaches production. The costs are real. The savings are on a slide.

At Mindgram — a mental health platform operating in 80 countries and 37 languages — I implemented AI across three operational areas: customer support automation, translation workflows and content management.

Result: approximately 25 FTE OPEX reduction per year, with hundreds of thousands of EUR in annual savings. The system is in production. It's not a pilot.

Start with the cost, not the technology. The question is never "how can we use AI?" It's "which operational cost can be reduced by 60–80% with acceptable quality risk?"

Prioritise high-volume, repetitive, low-stakes tasks first. Customer support routing, content translation, first-draft generation. These have clear quality thresholds, measurable output and recoverable errors.

Build for handoff, not dependency. Every AI-assisted workflow needs a human escalation path that the team actually uses. Systems that can't be overridden get abandoned.

Measure against a baseline, not a feeling. Before: how many FTE hours per week does this task consume? After: what's the delta? If you can't measure it, you can't claim it.

Ship to production in week 6, not month 6. If the implementation plan has more than two months before any production deployment, the project will not survive contact with a board review.

This framework works for startups and scale-ups with clear operational cost lines: support, content, QA, data processing. It doesn't work for companies that haven't yet identified which processes they want to automate — the starting point is a cost map, not an AI strategy.