Purpose of the Template
You're not deciding whether to use AI. You're deciding where it stops and where you start.
This template creates a written record of AI's role in specific decision contexts within your organization. It answers the question: "Who owns this decision when AI is in the room?" That clarity matters for three reasons. First, it preserves the judgment formation that Deborah Ancona and Katherine W. Isaacs identify as critical to leadership integrity. Second, it establishes accountability when outcomes need explanation or correction. Third, it gives your team permission to question AI outputs rather than defer to them.
The template works for any recurring decision where AI tools generate recommendations, drafts, or analysis: credit approvals, hiring decisions, content publication, risk assessments, resource allocation, customer communications, or strategic planning inputs.
Prerequisites
Before you fill this out, you need three things:
A specific decision type. Don't write "strategic decisions" or "operational choices." Write "monthly budget reallocation requests over $50,000" or "customer complaint escalation to legal review." The tighter the scope, the clearer the authority boundaries.
The actual AI tool or capability. Name the system: "GPT-4 via our internal API," "Salesforce Einstein recommendation engine," "our custom credit scoring model." Generic references like "AI assistance" create generic accountability.
The human role holder. Identify by title, not by name. "VP of Marketing," "Credit Risk Manager," "Regional Compliance Officer." If the role changes hands, the authority framework shouldn't need rewriting.
The Template
Copy this into your policy repository, decision log, or governance documentation:
AI Decision Authority Framework
Decision Type: [Specific recurring decision]
AI Tool/Capability: [Named system or model]
Human Authority: [Role title]
Effective Date: [Date]
Review Cycle: [Quarterly/Semi-annually/Annually]
AI's Role
AI will:
- [Specific task: "Generate three budget scenarios based on historical spend patterns"]
- [Specific task: "Flag applications with debt-to-income ratios above 43%"]
- [Specific task: "Draft initial response to standard complaint categories A, B, and C"]
AI will NOT:
- [Prohibited action: "Approve or reject any application"]
- [Prohibited action: "Send any communication without human review"]
- [Prohibited action: "Override risk tier assignments"]
Human's Role
The [Role Title] will:
- Review AI output for accuracy against [specific criteria]
- Verify that recommendations align with [policy/regulation/organizational value]
- Make the final decision and document reasoning when deviating from AI recommendation
- Communicate the decision to affected parties
- Own the outcome and any necessary corrections
The [Role Title] will NOT:
- Approve AI recommendations without independent verification
- Delegate final authority to AI output
- Accept AI reasoning as sufficient justification for external stakeholders
Judgment Checkpoints
Before accepting AI output, the decision-maker must confirm:
Contextual fit: Does the AI output account for [specific factor the AI cannot see]? Example: recent policy changes, customer relationship history, market conditions not in training data.
Voice authenticity: If this output will be attributed to me or my organization, does it reflect [our actual position/values/tone]?
Assumption validity: What assumptions did the AI make? Are they still true for this instance?
Boundary awareness: Did the AI stay within its defined scope, or did it offer recommendations beyond its competence?
Escalation Triggers
Escalate to [Next Level Role] when:
- AI recommendation conflicts with [specific policy or regulation]
- Decision involves [materiality threshold or sensitive category]
- Human judgment differs from AI output by [defined margin]
- Output quality degrades or shows signs of model drift
Documentation Requirements
For each decision, record:
- Date and decision-maker
- AI recommendation (verbatim or summary)
- Human decision
- Reasoning when human decision differs from AI output
- Any contextual factors AI could not evaluate
Customizing the Template
Adjust the judgment checkpoints to match what your decision-maker actually needs to verify. If you're reviewing AI-generated customer communications, add: "Does this response address the customer's underlying concern, not just their literal question?" If you're reviewing credit decisions, add: "Does this recommendation account for our relationship banking strategy?"
Set materiality thresholds that trigger escalation. For financial decisions, that might be dollar amounts. For content decisions, it might be audience size or regulatory sensitivity. For hiring, it might be role level or legal risk category.
Specify the documentation medium. Don't just say "document the decision." Say "log in the decision register," "note in the CRM record," or "add to the audit trail in [system name]." If documentation doesn't happen in the workflow, it won't happen at all.
Tailor the review cycle to model stability and regulatory requirements. If you're using a rapidly evolving foundation model, quarterly reviews make sense. If you're using a stable internal model under SR 11-7, align with your annual model validation cycle.
Validation Steps
After you deploy this framework, test it:
Shadow three decisions. Watch someone use this template in practice. Do they skip checkpoints? Do they struggle with any verification step? Revise based on friction points.
Check for delegation creep. One month in, audit whether decision-makers are actually reviewing AI output or just rubber-stamping it. If approval time drops to seconds, your checkpoints aren't working.
Review deviation documentation. When humans override AI, are they documenting reasoning? If not, your documentation requirement is too burdensome or too vague. Simplify or make it part of the workflow.
Measure judgment formation. Ask decision-makers: "Do you understand this domain better after using this framework than before?" If the answer is no, AI is eroding rather than supporting their expertise. That's the risk Ancona and Isaacs warn against.
The goal isn't to slow down decisions or eliminate AI value. It's to ensure that when something goes wrong, you can point to a human who understood the context, verified the output, and owned the choice. That's not a compliance checkbox. That's what makes leadership answerable.
