Confusing 'consequential decision' with 'high-risk'
All high-risk AI systems make consequential decisions, but a 'consequential decision' is specifically about education enrolment, employment, financial/lending, essential gov service, healthcare, housing, insurance, or legal services (§1701(3)). HR / lending / insurance + healthcare + housing are the four most common Colorado scope traps.
Assuming the small-deployer carve-out fully exempts you
<50 FTE + no own-data training + intended use = exempt from §1703(2) risk management program AND §1703(3) impact assessment AND §1703(5) public statement — but NOT from consumer notice duties (§1703(4)) or AG disclosure duties (§1703(7)). Most startups are in this 'middle' zone and don't realise it.
Using a third-party HR or credit AI without provider documentation
Deployers must perform an impact assessment that depends on developer-supplied information (training data, foreseeable misuse, performance). If your vendor refuses to provide the §1702(2) statement, you cannot complete §1703(3) — and you become liable for the gap. Procurement agreements need an AI-Act-style flow-down.
Substantial modification triggers full re-assessment
Fine-tuning, prompt engineering that materially changes outputs, or changing the input data significantly = 'intentional and substantial modification' under §1701(7). Re-assessment within 90 days. Many SaaS companies layer prompts over GPT/Claude weekly without realising each material change resets the clock.