Written by: Jakub Pietroszek, Partnership Manager, Digital Colliers
Ask a parts distributor who can find any obscure bearing and you get three names. Ask what happens when those three retire and the room goes quiet. That silence is where most industrial AI projects should start, and almost none of them do.
The usual pattern is backwards. Someone reads about copilots, books a workshop, drafts a roadmap, and eight months later the pilot dies quietly. The reason it dies is rarely the model. It's that the pilot never mapped onto a real risk the business already knew it had.
The roadmap-first approach keeps failing in public
The numbers on this are not subtle. IDC with Lenovo found that 88% of AI proofs-of-concept never reach widescale deployment, and only four in every 33 make it to production. S&P Global's 2025 survey put abandonment at 42% of companies dropping most of their AI initiatives, up from 17% the year before. MIT's Project NANDA put a sharper edge on it: 95% of enterprise GenAI pilots deliver zero measurable P&L impact.
Those failures share a shape. The project chose a technology and then went hunting for a problem. When the problem it found was thin, the pilot had nothing to hold onto once the novelty wore off.
Key-person risk is the opposite kind of starting point. It's a problem the operations director can already name, already worries about, and already has no good answer for. That's the ground AI needs.
What key-person risk actually looks like on the floor
Walk any mid-sized industrial business and you'll find pockets of it. A few examples of the pattern:
- One estimator who knows which suppliers will actually hit a two-week lead time versus which ones quote it and miss.
- A service engineer who can diagnose a machine fault from the sound before the fault code even prints.
- A back-office clerk who knows which customer POs need a phone call before they're trusted.
- A distributor rep who remembers that the part number changed in 2011 and the old one is still in three warehouses.
None of this is written down. Some of it can't be, not cleanly. But most of it leaves a trail: emails, quotes, service tickets, chat logs, phone notes, the CRM comments field nobody reads. That trail is training data for a system that keeps the knowledge in the building after the person leaves.
Start from the interview, not the architecture
Operators who are shipping useful industrial AI in 2026 tend to run the same early move. Before anyone talks about models, they sit with the person who knows the thing and ask three questions.
- What do people ask you that nobody else in the building can answer.
- Where do you look, or who do you call, to answer it.
- What would go wrong tomorrow if you were off sick for a month.
The answers give you a shortlist of candidate use cases that are already load-bearing. From there the work is unglamorous. You find where the source material lives. You clean it. You decide what the assistant is allowed to say and what it has to escalate. You keep a human in the loop for the calls that matter.
This is also where most projects should stop and ask whether AI is even the right tool. Sometimes a checklist and a shared inbox does the job. The point isn't to build something clever. It's to make sure the answer survives the retirement party.
The compliance clock is already ticking
There's a second reason to start here rather than with a shiny roadmap. Anything that ends up touching decisions about people, safety, or critical infrastructure is going to sit under the EU AI Act's high-risk obligations from 2 December 2027, with the Article 50 transparency rules landing earlier on 2 August 2026. Fines for high-risk violations run up to 15 million euros or 3% of global turnover.
A system built around a named, documented knowledge domain is a lot easier to audit than a general-purpose assistant somebody bolted onto the intranet. You know what it's for. You know who signed off the source material. You know what it isn't allowed to do.
That's the quiet advantage of starting from key-person risk. You end up with an AI project that has a job description, a boss, and a reason to exist the morning after go-live. Most pilots can't say the same.