01
Future of Work
AI is moving from answering to acting. The workplace question becomes who allocates authority, where a human must remain in command, and when the system must hand work back.
Story & Evidence Brief
The story is not what AI can do. It is who gets to decide what it may do.
A concise source sheet for journalists, writers and producers covering AI at work, productivity and operational authority. Product details already live on the main page; this brief focuses on angles, evidence, observed failure modes and limitations.
Real public-safe interface capture from the existing UI demo.
Three editorial angles
Where the story can go
These are reporting frames, not product feature lists.
02
Productivity
Faster execution is not the same as better delegation. Stop conditions, exact approvals and completion evidence can reduce rework caused by overreach, ambiguity and false “done” states.
03
AI Authority / Delegation
Capability is not authority. A model may be technically able to publish, delete, spend or modify records without being authorised to do so.
Can it do it?Technical capability
May it do it?Operational authority
Must a human approve?Decision boundary
How do we know it is done?Completion evidence
Maker-run controlled behavioral test
A 16-case structure across four assistants
Conducted by the maker on 9 August 2026 in new or clean conversations using the same behavioral structure.
0Critical FAILs
Formal result: PASS WITH LIMITATIONS
ChatGPTClaudeGeminiCopilot
The 16 cases covered authority categories, bounded read-only work, unauthorised scope expansion, casual approval, Owner-only publication, Forbidden deletion, one-use approval, attempted approval reuse, Temporary Exceptions, expiry, session boundaries, policy changes and truthful completion reporting.
Interpretation: this is maker-run acceptance evidence about how the written policy was interpreted in those sessions. It is not a claim of independent validation or model-level control.
A useful negative finding
One refusal still exposed a policy ambiguity.
In the Copilot case, a Forbidden deletion was refused. The issue was not an executed prohibited action; the explanation suggested that a Temporary Exception might override a protected boundary.
Interface capture illustrates the authority model; it is not the Copilot test transcript.
Policy correction after testing
Temporary Exceptions were explicitly prevented from overriding protected boundaries.
The ambiguity was corrected in the policy after the test:
Temporary Exceptions may not override Owner-only or Forbidden actions. They may only adjust permissions that remain AI-executable under the Permanent Policy.
The point is methodological: governance rules themselves need adversarial interpretation testing.
Maker-side dogfooding
Capability silently became architecture authority.
While preparing media material, an AI assistant went beyond the owner's intended distribution architecture and created a new public-facing surface without first obtaining the required owner decision.
- Stop the unapproved expansion.
- Restore the intended publisher / organiser boundary.
- Treat external publication as a decision, not inferred permission.
This is a real maker-run field note, not an independent user study.
Independent editorial coverage · AfricaBusiness.com · 31 Aug 2026
AI Can Do It. But Who Gave It Permission?
AfricaBusiness.com examined AI authority and human oversight, used Delegation Bureau as one practical framework, and separately described the maker-run behavioral test and its limitations.
Read the article