A proposed framework earns its credibility through what it explains, what it misses and whether practitioners can put it to work. I am inviting a small number of practitioners and researchers to challenge this working model, not to endorse it.
Published by Alper Benli, 31 October 2026. No external review or independent validation is claimed at this stage.
What is being proposed?
The four-contract model asks organizations to specify the outcome, decision rights, access permissions and assurance arrangements for each operational AI agent. A companion 12-check worksheet encourages teams to document evidence before granting wider operational autonomy.
The motivating question is not simply whether an agent is technically capable, but whether a business can demonstrate who owns the result of the work and what it costs to achieve a successful outcome once people, agents and exception handling are accounted for.
What already exists?
Delegation, oversight, agent identity and risk management have extensive prior work. This framework does not claim otherwise. These sources are especially relevant:
- Kasirzadeh and Gabriel (2026), Agentic profiles for effective AI governance, Nature distinguishes agent classes across autonomy, efficacy, goal complexity and generality. It helps identify how governance requirements vary with an agent's capabilities.
- WEF/Capgemini, AI Agents in Action (2026) sets out organizational authorization and scaling considerations.
- NIST AI Risk Management Framework establishes broad risk management functions and a lifecycle orientation.
The proposed additional emphasis is an explicit link from delegated work to accountable business owners and a financially meaningful denominator: cost per successful, quality-adjusted end-to-end outcome. Whether this adds practical value beyond existing governance approaches is precisely what needs testing.
Five questions for reviewers
- Originality: Does this framework add a useful operational connection to existing guidance, or simply rename familiar controls? What prior work is missing?
- Measurement: Is cost per successful end-to-end outcome the right primary economic unit? In which contexts would it distort decisions?
- Accountability: How should ownership work when multiple agents, teams or suppliers contribute to one business outcome?
- Safety and access: Which failure cases, human-approval thresholds or control obligations are absent from the 12-check worksheet?
- Adoption: Could a business owner and risk lead use this tool in a real deployment review? What would need to change?
How feedback will be handled
Specific criticism is welcome, including examples where the model fails or duplicates established practices. Feedback will be used to improve the wording, scope and evidence requirements. Contributors will be acknowledged or quoted only with their explicit permission. No endorsement will be inferred from providing feedback.
Any future update will identify material revisions and their date. A later field trial would need a documented method, appropriate permissions and clear limitations before being described as validation.
Share feedback
If you are a researcher, enterprise AI leader, business owner, risk professional or practitioner with a useful counterexample, please send a note. A 15–20 minute critical conversation or concise written comments would both be valuable.