Secure AI Adoption Maturity Roadmap
Where does your organization actually stand?
Five levels, scored separately across five control domains, so you get a profile rather than a grade. Each level is described by what it lets you safely do next, not by how locked down you are.
The levels
Control buys autonomy
Each level adds controls, and each set of controls lets you safely do more. Find where you are today, then find where you need to be.
Everyone configures their own setup. No policy, no visibility.
Settings enforced, destructive actions denied, secrets out of reach, agent identity read only.
Trust boundary written down, connections allowlisted, changes approved, policy and runbook in place.
Agent activity visible and attributable, evidence maintained, controls reviewed, drills run.
Autonomy tiered by risk, control testing automated, red teaming scheduled.
Level 4 is not more controls. It adds five practices on top of all thirty two, such as automated control testing and scheduled red teaming, and it is only worth reaching where agents run infrastructure or pipelines.
The five domains
Scored separately, because nobody is uniformly weak
Scoring
Scored on evidence, not opinion
A control counts when the evidence exists. "We think we do that" scores zero. We report two numbers: the average across the domains, and the floor, which is your lowest domain. Risk follows the floor, so that is the number to act on.
Our model maps to the Cloud Security Alliance's agentic governance work and to OWASP's agentic risk list. You can score yourself against those for free. What we add is evidence, implementation, and someone to keep it current.
Targets
Level 3 is not the answer for everyone
| Your situation | Twelve month target |
|---|---|
| Under 50 engineers, no regulated data | Level 2 across the board. Assurance machinery costs more than it returns at your size. |
| 50 to 300 engineers | Level 2, with level 3 on observability and identity, so you can answer customer security reviews. |
| Regulated, or holding SOC 2 or ISO 27001 | Level 3 across the board. Your auditor will ask, and the evidence pack is the answer. |
| Agents touching infrastructure or pipelines | Level 3, plus level 4 in identity and permissions. |
Two rules of thumb.
- 1
Never leave a domain at 0 once Claude has production access.
- 2
Move one level at a time, because a two level jump in a quarter usually means the controls went in without the adoption.
Questions
Questions about the maturity model
It is a way to measure how safely your organization uses AI agents. Ours has five levels, from 0 to 4, scored separately across five control domains. Each level is described by what it lets you safely do next, not by how locked down you are.
Level 1 needs 8 of the 32 controls in the DIGU AI Security Baseline. Level 2 needs 20. Level 3 needs all 32. Level 4 keeps all 32 and adds five practices, such as automated control testing and scheduled red teaming.
Most software companies should target level 2. Regulated organizations, and those holding SOC 2 or ISO 27001, should target level 3 across all five domains. Level 4 is only worth reaching where agents provision infrastructure or run in your pipelines.
On evidence. A control counts only when its evidence exists. We report two numbers: the average across the five domains, and the floor, which is your lowest domain. Risk follows the floor, so the floor is the number to act on.
One level at a time. A two level jump in a single quarter usually means the controls went in without the adoption. Never leave a domain at level 0 once an AI agent has production access.