The number deserves a pause: in MIT's "State of AI in Business 2025" study, 95 percent of enterprise GenAI pilots produced no measurable profit impact. The figure, reported by Fortune in August 2025, corroborates what Gartner had predicted: a large share of generative AI projects are abandoned after proof of concept.
And yet Swedish companies keep investing. Statistics Sweden (SCB) reports that 35 percent of Swedish enterprises used AI in 2025 — top of the EU. The problem, then, is not engagement. It is the transition from pilot to production. Here are the four patterns we see separating the winners from the rest.
Pattern 1: Winners start in the process, not the model
The losing pilot starts with "what can we do with this technology?". The winning pilot starts with "which process costs us the most — and why?". A GenAI solution saving three minutes per case can be brilliant engineering and worthless business: three minutes on a thousand cases is fifty hours — but only if the workflow around the tool is designed to capture them.
Pattern 2: Evaluation is built before scaling — not after
"Feels good" pilots die on first contact with reality, often during scrutiny of the first wrong answer. Pilots that reach production have an evaluation harness from the start: representative test cases built from your real questions, metrics for correctness and groundedness, and recurring runs that catch regressions — for instance when models are swapped.
The harness is also what makes price and effect visible. Without it, you don't know whether the solution got better or more expensive since last month.
Pattern 3: The data and permission model is taken seriously
A solution that answers from internal documents must also know who is allowed to see what. Pilots that ignore the permission model become either dangerous (answers leak privileges) or useless (everyone answers timidly to be safe). Retrofitting permission filtering is far more expensive than designing for it from the start.
Pattern 4: Someone owns the business outcome — not just the project
The most common death notice we see: the pilot ran out, and nobody had the mandate to carry it forward. Scaling requires an owner accountable for the metric that defined success, an operations plan, and a decision point where you deliberately choose to scale, adjust or stop. Stopping is not failure — hesitating for two years is.
The Sweden paradox
Here is the most interesting number right now: IT & Telecom Företagen reports that the share of companies describing themselves as AI leaders fell from 22 to 5 percent in one year, while usage grows. Skills are the top barrier for those who haven't started.
Read it this way: getting started is easy, being good is hard. The advantage therefore shifts to those who build the right foundations early — not those who experiment the most.
How to avoid the patterns
- Choose the pilot on process economics, not technical curiosity.
- Build the evaluation harness before the scaling decision.
- Design for permissions, logging and cost control from day one.
- Appoint an owner with a mandate — and set a scale-or-stop date.
It doesn't sound glamorous. That's why it works.
Sources
- MIT Project NANDA — "The GenAI Divide: State of AI in Business 2025", reported by Fortune, 18 August 2025
- Gartner — press release 29 July 2024: 30% of GenAI projects abandoned after PoC
- Statistics Sweden (SCB) — AI in enterprises 2025, published 11 December 2025 (scb.se)
- IT & Telecom Företagen — Sweden losing ground in AI, 2026 (itbranschen.com)