Our research · November 2026
Even the most powerful AI
is only as good as the
architecture behind it.
We tested KYP against five of the most common ways of using AI, on real renewable energy projects, and let government regulators decide what counted as a problem.
1 · The test
Real projects. Real regulatory findings.
We compared what each AI approach found in real renewable permitting documents with the issues government regulators flagged.
The question was simple: how much of what the regulators found could each AI approach uncover?
- 20 real projects
The developer’s own documents: impact studies, scoping reports, permit applications.
- 6 ways to use AI
KYP and five common alternatives, all on the same frontier AI model.
KYP · Frontier AI · RAG · GraphRAG · Advanced RAG · Agentic RAG - The regulator’s list
979 problems the authorities actually raised, written down before any AI ran.
- Who found the most?
Each approach scored on the share of the regulator’s problems it found.
- 4countries: Italy, UK, US, Spain
- 4+ GWof solar, wind and storage
- 5,111pages in three languages
- 13cases KYP had never seen, used for every result
2 · The results
KYP found about twice as much, in every case.
- Arrow Valley UK72%23%
- Deanmoor UK92%48%
- Lombardore IT80%45%
- Byers Gill UK77%43%
- Algiedi ES62%30%
- Guglionesi IT67%39%
- Valdemoro ES54%27%
- Potentia-Viridi US73%46%
- Perkins US47%23%
- Soda Mountain US92%69%
- Ascoli Mezzaluna IT83%61%
- Darden US62%42%
- Marta IT47%31%
3 · Why it matters
What you miss before signing, you pay for after.
Every problem in this test was raised by a regulator on a real project. Each one missed is a delay, a renegotiation or a write-down later.
| Problem | Frontier AI | KYP |
|---|---|---|
| SpainA turbine 170 m from a stream, with no erosion plan The regulator approved the wind farm only if erosion near water is controlled. Frontier AI saw the stream, but not the risk. KYP flagged it. If missed: erosion work nobody budgeted for, typically ~3% of construction cost1 | Missedmissed | Found · p. 44found |
| UKUnexploded WWII bombs not ruled out The developer wanted to skip a site-specific bomb check. The regulator said no. Frontier AI never mentioned it. KYP saw the check used only general maps. If missed: an extra bomb-risk study, typically £36k–49k for 40 engineer days2 | Missedmissed | Found · p. 225found |
| USThe battery’s fire risk left out The hazard list left out the lithium-ion batteries, and the regulator asked for them. Frontier AI saw the fire design was unfinished, but missed the gap. KYP flagged it. If missed: fire data the regulator is waiting for, typically $40k–100k+ for battery fire testing3 | Missedmissed | Found · p. 197found |
Typical costs are public figures, not from these projects.
¹ Stormwater magazine survey of 27 US states
² Natural England tender, Contracts Finder 2025
³ SunLith Energy, UL 9540A test guide
- You can’t tell what AI missed
A short, confident summary looks complete. The gaps only show up later, when the regulator or the market finds them.
- Every finding has its page
KYP links each finding to the exact sentence it rests on, so your team can check it in seconds.
- You decide before closing
More of the real problems, found early, with the evidence attached: the basis for price, conditions or walking away.