Five people answer
A product lead, an engineering lead, a platform lead, a risk and compliance owner and a sponsor. About fifteen questions each, about how teams work today.
02 · AI for IT · Assessment
Where your software delivery stands with AI, step by step.
Most large companies use AI in some steps of software delivery and not in others. The assessment gives each of the eight steps a level, shows the tools and rules around it, and sets out what to do next, in the right order.
8Steps assessed
50Questions
5Roles answer
The Eight Steps
Each step gets its own level.
01Intake and prioritisation
How work arrives, is sorted and is routed.
02Specification and planning
From a need to clear rules, acceptance criteria and tasks.
03Design
Architecture, interfaces and user experience.
04Implementation
Code, tests and documentation.
05Testing and checks
Automated tests, security analysis and the evidence they leave.
06Review and acceptance
Who reviews a change, and who decides to accept it.
07Release and change control
Change approval, deployment and rollback.
08Operations and maintenance
Monitoring, incidents, fixes and upgrades.
Security, the audit trail and the knowledge agents can use are assessed inside every step, not as separate ones.
The Levels
Six levels, from manual to self-improving.
LevelNameWhat it looks likeWho decides
L0Manual
No AI in this step.
People
L1Assisted
People use approved AI assistants. No whole task goes to an agent.
People
L2Delegated
Agents carry out whole tasks. A person checks every output.
A person, every time
L3Orchestrated
Agents work in a connected pipeline with automated checks.
People, at set points
L4Governed autonomy
In categories your rules allow, agents complete the work and evidence accepts each change.
A named owner per category; people handle exceptions
L5Self-improving
The system measures its own results and proposes improvements to itself.
A person approves every change to the system
The top level is not software delivery without people. At every level, a named person owns each rule and each approval, and no agent approves its own work. For critical systems, review and release stay at L3 unless your internal audit agrees otherwise.
How It Works
Answers, evidence, then a session with your teams.
Evidence, not opinion
Each answer describes a practice and names what shows it: an approval log, a pipeline, a policy. We record what is declared, what is shown and what is checked.
Levels set together
We go through the answers with your teams in one working session. A step’s level is set by its weakest practice, so a strong tool cannot hide a gap.
The next step, in order
You set a target per step. We check what has to come first, split large jumps into single levels and write what to build for the first one.
What You Hold at the End
A level for each step, and a plan your teams can start on.
- A map of the eight steps: today’s level, the teams ahead of the rest, and your target
- The tools behind each step, and what they can already do
- A plan in the right order: what has to come first, and why
- A written plan for the first step: what changes, what to build, what stays with people, how to measure it
- An evidence register your auditors can read
Built for Regulated Companies
Your rules decide how far agents go.
- Change control mapped to DORA: every change recorded, tested, approved by someone independent, and verified
- EU AI Act and GDPR reviewed with your compliance team
- Teams are assessed, never individuals: no answer rates a named person
- A risk and compliance owner answers alongside engineering
Where It Runs
Your models, your perimeter, your choice.
Your cloud
The AI models you already use under your cloud agreements, in your own tenant.
Private
Dedicated model endpoints in your private cloud, with outbound traffic restricted.
Sovereign
European or open-weight models on your servers or a SecNumCloud-qualified cloud, with no fallback to other models.
We are not tied to any model provider. Which model does which step is decided by how each one performs on your own past work, within your rules.
Start with thirty minutes.
Tell us how your teams build software today. We will tell you plainly whether the assessment is the right first step, and who should answer.