Every week someone comes to me with the same request: “We want to adopt AI. What do you recommend?” Almost always it’s a conversation starting in the wrong place. You don’t start with the technology — you start with a specific process, one that is eating hours and money today.
Where dead pilots come from
By various estimates, up to 80% of AI pilots never reach production. The cause is almost always the same: they started with the model, not the task. Bought access to an LLM, built a slick demo on synthetic data, showed it to the leadership — and got stuck. On real data the demo falls apart: the data is messy, the process was never written down, and nobody calculated the impact. And what AI actually isn’t — I sorted out separately, without the hype.
I’ve been deploying AI since 2017 — long before the hype. The main lesson from those years is simple: AI works where there is a measurable process, not where there is an enthusiast.
How I pick processes to automate
Before proposing any technology, I answer three questions:
- What does it cost today? Staff hours, errors, losses. If the process is cheap, AI won’t save it.
- What counts as success? KPIs are fixed before kickoff, in business units: hours, money, cycle time. Not “improve quality” but “cut the analytics cycle from weeks to hours”.
- What do we do if it doesn’t work? The stop criterion is also fixed upfront. Walking away from a pilot after a few weeks is normal — it’s better than burning a yearly budget.
Boring tech also works
A funny secret: after this kind of review, half the time AI turns out not to be needed at all. A plain script, a CRM rule or simply data discipline solves the problem cheaper and more reliably. An honest “you don’t need AI yet” costs less than a failed pilot — and that’s a valid consulting outcome too.
A function-level breakdown on the same principles: AI in accounting, the Russian way. And the scenario that most often passes this filter for my clients: a personal AI assistant for every employee — it has a price, a metric, and a separate conversation about risks. And the typical mistakes employees start making the moment they get access — from empty prompts to blind trust — are covered in ten mistakes employees make with AI.
In short
AI is not about models. It’s about processes that have a price, a metric and an owner.
If you want to review your processes — start with a diagnostic. Three biggest risks, a map of quick wins, and an honest answer on whether you need AI at all. That’s 3–5 days of work, not months.