Tier AI systems by asking four questions: how sensitive is the data, how much does the output affect people, how much can the system act on its own, and how widely is it used. Score each, set tier thresholds, and attach review requirements to each tier.
Why tier at all
NIST's framework says not all AI risks are the same, and that policies and resources should be prioritized based on assessed risk level and potential impact. Trying to eliminate risk entirely can waste scarce resources. Tiers let a small team spend effort where it counts.
NIST's prioritization cues
The framework notes that higher initial prioritization may be called for where a system is trained on large datasets of sensitive or protected data, such as personal information, or where outputs have direct or indirect impact on humans. Systems that only interact with computational systems and use non-sensitive data may call for lower initial prioritization, though regular reassessment still matters because non-human-facing systems can have downstream effects. It also notes that the degree of customization by the deployer can matter.
Four scoring questions
Score each 1 to 3.
- Data. 1 for public or non-sensitive, 2 for internal business data, 3 for personal, regulated or confidential customer data.
- Impact on people. 1 for convenience features, 2 for outputs shaping customer-facing work, 3 for decisions about access, money, employment, health or safety.
- Autonomy. 1 for suggestions a human reads, 2 for actions after approval, 3 for actions without approval.
- Reach. 1 for a handful of users, 2 for a department, 3 for customers or the public.
Thresholds
- Total 4 to 6: low tier.
- 7 to 9: medium tier.
- 10 to 12: high tier.
- Any 3 on impact on people: at least medium, regardless of total.
Tune the cutoffs after you score your first ten systems.
Requirements by tier
- Low: inventory entry, named owner, acceptable use policy applies.
- Medium: add risk assessment, documented testing, vendor review, monitoring.
- High: add impact assessment, independent review, executive risk acceptance, rollback plan, incident drill.
Re-tier when things change
New data source, new capability such as tool access, wider rollout, or an incident should trigger a re-score. Record the score and date in the inventory.
An example scoring
An internal meeting summarizer using company data, read by attendees, with no actions, used by one department: data 2, impact 1, autonomy 1, reach 2, total 6, low tier. A customer-facing assistant with access to account records that can issue credits without approval: data 3, impact 2 or 3, autonomy 3, reach 3, total 11 or more, high tier. The scores make the argument visible and easy to challenge.
Quick check
Should every chat assistant be low tier?
No. A chat assistant with access to customer records and the ability to send messages scores high on data and autonomy.
Who assigns the tier?
The system owner proposes it and the governance group confirms medium and high.
Sources
Last reviewed: 19 October 2026.
