Generative AI pilots stall, with little or no measurable impact on the business.
Agentic AI · Why AI tools aren't enough
A tool that acts still needs an owner
An agent is a tool that can take a step on its own. The step still needs a person who is allowed to check it, stop it or repair it.
The tools teams already use
-
AI coding agents
- AIDEN
- GitHub Copilot
- Cursor
-
Code and delivery
- GitHub
- GitLab
-
Work tracking
- Jira
- Azure Boards
-
Team methods
- Scrum
- RACI
-
Training
- generic AI literacy e-learning
-
Governance frameworks
- NIST AI RMF
- ISO/IEC 42001
- EU AI Act
-
Monitoring
- MLOps and model monitoring
Each one is useful. None of them was built to answer who owns the outcome for the person on the other end?
What each tool covers, and what it leaves open
The gap: ownership, rehearsal, the right to stop
covers partly not covered
Scroll →
| The gap | The gap | ||||||
|---|---|---|---|---|---|---|---|
| Tool | Speeds up the work (individual or team) | Review before release | Owner of the outcome for the person affected | Role-specific learning | Team rehearses the handoffs | Frontline can stop or override | Frontline feedback flows back |
| AI coding agents (AIDEN, Copilot, Cursor) | covers | partly covers. Code and spec review, not review of the outcome. | not covered | not covered | not covered | not covered | not covered |
| Code platforms (GitHub, GitLab) | partly covers | covers. Pull-request review of code changes. | not covered | not covered | not covered | not covered | partly covers |
| Work tracking (Jira, Azure Boards) | partly covers | partly covers | partly covers. Assigns tasks, not the outcome for citizens. | not covered | not covered | not covered | partly covers |
| Team methods (Scrum, RACI) | partly covers | partly covers | partly covers. RACI names roles on paper. | not covered | partly covers. Retrospectives look back; they don't rehearse failures. | not covered | partly covers |
| Generic AI training | not covered | not covered | not covered | partly covers. The same course for every role. | not covered | not covered | not covered |
| Governance frameworks (NIST AI RMF, ISO/IEC 42001, EU AI Act) | not covered | partly covers | partly covers. They require it, not how each role does it. | partly covers. They require it, not how each role does it. | not covered | partly covers. They require it, not how each role does it. | partly covers |
| Monitoring (MLOps) | not covered | partly covers | not covered | not covered | not covered | not covered | partly covers. Technical signals, not citizen experience. |
| Many Hands Proposed · to be tested in the pilot | partly covers. Indirectly: less rework, faster handoffs and fixes. To be tested in the pilot. | covers | covers | covers | covers | covers | covers |
AI tools speed up each person. Many Hands speeds up the team, and makes sure someone owns what happens next.
My assessment, based on each tool's public description and typical use. Many Hands is a proposal and will be tested in the pilot.
Read the chart as a list
- AI coding agents cover speeding up the work, partly cover review before release, and leave ownership, role-specific learning, rehearsal, a frontline stop, and frontline feedback open.
- Code platforms partly cover speeding up the work and frontline feedback, cover review before release, and leave ownership, role-specific learning, rehearsal, and a frontline stop open.
- Work tracking partly covers speed, review, ownership of the outcome, and frontline feedback, and leaves role-specific learning, rehearsal, and a frontline stop open.
- Team methods partly cover speed, review, ownership, rehearsal, and frontline feedback, and leave role-specific learning and a frontline stop open.
- Generic AI training partly covers role-specific learning and leaves the other six needs open.
- Governance frameworks partly cover review, ownership, role-specific learning, a frontline stop, and frontline feedback, and leave speeding up the work and rehearsing handoffs open.
- Monitoring partly covers review before release and frontline feedback, and leaves speed, ownership, role-specific learning, rehearsal, and a frontline stop open.
- Many Hands, proposed and still to be tested, partly covers speeding up the work and covers review, ownership, role-specific learning, rehearsal, a frontline stop, and frontline feedback.
What the tool does not settle
-
Who may act
Which steps the agent may take alone, and which wait for a named person.
-
Who sees it first
Whether a designer, a developer or a caseworker sees the output before a citizen does.
-
Who repairs it
The route back when the step was wrong, including a person the citizen can reach.
The numbers
Employees who feel adequately trained to use AI.
Employees who rely on AI output without evaluating its accuracy.
Frontline staff whose leaders give enough guidance on AI.
Germany’s place among 47 countries for AI literacy.
The EU AI Act’s AI literacy duty has applied to providers and deployers since this date.
The same gap, one step further
Teams already have coding agents, chat windows and workflow tools. The work still breaks in the handoff: a draft nobody reviewed, a status nobody owns, a citizen letter nobody can explain.
An agent makes that handoff easier to miss, because the next action can happen before the team meets. A log of what it did is not the same as a decision the team agreed.