Overview

Wednesday, 9:00 AM. You're updating your operational runbooks, the documents that describe how processes actually work. For the vendor selection process, you realize: the documented process (written two years ago) doesn't mention AI. But AI is now used in the analysis phase. The runbook is outdated.

More importantly, if someone new joins your operations team and reads the runbook, they won't know that AI is involved in vendor selection. They won't know what AI outputs to expect, what they're supposed to verify, or what happens if the AI output is rejected.

This is the documentation problem: as you integrate AI into operations, your processes change. Your documentation must reflect these changes. At Level 3, you're not just documenting individual decisions. You're documenting how AI is integrated into the process itself.

Process-Level Documentation for AI

When AI is integrated into a process, update your process documentation to include:

1. Where AI is Used**

At which step(s) in the process does AI participate? Be specific.

Example: "In the vendor selection process, AI is used in Phase 1 (Analysis) to analyze supplier data and identify concentration risks. AI is NOT used in Phase 2 (Review) or Phase 3 (Approval)."

2. What AI Does (Inputs and Outputs)**

What data does AI work with? What does AI produce?

Example: "Input: Supplier dataset (pricing, delivery, quality history, financial stability). AI Output: Concentration risk analysis and recommended vendors for dual-sourcing strategy."

3. What Humans Do (Review and Override)**

What's the human role? What are they checking? What can they override?

Example: "Human Role: Procurement manager reviews AI analysis, verifies vendor capacity, ensures analysis doesn't violate policy constraints, approves or modifies recommendation before escalation."

4. What Happens if AI Output is Rejected**

If the human reviewer doesn't accept the AI output, what's the next step?

Example: "If vendor capacity can't be verified or if recommendation violates policy, the AI is provided feedback ('these vendors don't have capacity for volume requested') and re-runs analysis with corrected constraints. Output is re-reviewed before escalation."

5. Decision Authority and Escalation**

Who can approve the AI-informed recommendation? When is escalation required?

Example: "Procurement manager can approve recommendations for routine vendors under $100K. Recommendations for strategic vendors or above $100K are escalated to VP Procurement."

6. Metrics and Monitoring**

What metrics are tracked to ensure the AI-integrated process is working well?

Example: "Track: AI recommendation approval rate (should be 60-80% without revision), average review time, cost savings achieved vs. AI estimate, vendor performance vs. AI predictions."

Tip: Update your runbooks as you integrate AI. Don't wait until six months later when knowledge about the AI-integrated process exists only in people's heads. Document it while the integration is fresh and while the team can help refine the documentation.

Impact Assessment Documentation

When AI is integrated into a process, create an impact assessment documenting: How does this change affect the organization?

Area 1: Decision Quality**

Does AI improve or degrade decision quality? Measure against: accuracy (are decisions aligned with desired outcomes?), consistency (are similar situations handled similarly?), speed (are decisions faster?).

Example: "Vendor selection with AI shows: (1) Accuracy: 8/10 AI recommendations are approved without modification (80% approval rate); (2) Consistency: vendor selection now uses consistent criteria regardless of procurement manager; (3) Speed: analysis time reduced from 6 hours to 30 minutes."

Area 2: Efficiency and Costs**

Does AI integration improve efficiency? What are the costs?

Example: "Efficiency gains: AI analysis reduces procurement manager time from 6 hours to 1 hour per decision. Annual time savings at 40 decisions/year: 200 hours = ~$25K. Costs: AI subscription $5K/year, training 20 hours. Net benefit: ~$20K annually plus 200 hours freed for higher-value work."

Area 3: Risk Profile**

Does AI integration introduce new risks or reduce existing ones?

Example: "Risk changes: (1) Concentration risk better identified and managed (risk reduction). (2) Vendor relationship risk: AI may recommend dropping long-term vendors if not strategic (need escalation process). (3) Bias risk: AI may systematically favor certain vendor types (need bias monitoring). Mitigation: escalation protocols, quarterly bias audits."

Area 4: Capability and Training**

Does AI change what skills the team needs?

Example: "Capability changes: (1) Less need for spreadsheet analysis skills. (2) More need for critical evaluation of AI outputs. (3) New need for understanding AI limitations and bias risk. Training required: 4 hours on 'how to evaluate AI vendor analysis.'"

Area 5: Stakeholder Impact**

How does AI integration affect other teams, customers, or vendors?

Example: "Stakeholder impact: (1) Vendors: faster decision timelines (improvement). (2) Internal teams: approval timelines shorter, more transparent criteria (improvement). (3) Finance: cost savings support budget targets (improvement). (4) HR: no headcount reduction, but skill mix changes (neutral/manageable)."

Real Schema: Process Documentation with AI Integration

```json
{
"process_documentation": {
"process_id": "PROC-VENDOR-SEL-001",
"process_name": "Vendor Selection - Raw Materials Category",
"version": "2.0 (AI-Integrated)",
"version_date": "2026-04-09",
"ai_integration": {
"ai_involved": true,
"ai_location": "Phase 1: Supplier Analysis",
"ai_inputs": {
"data_source": "Procurement database",
"data_elements": ["pricing", "delivery_history", "quality_metrics", "financial_stability", "capacity_information"]
},
"ai_outputs": {
"analysis": "Concentration risk assessment, vendor recommendations",
"format": "JSON with recommendation and supporting analysis"
}
},
"process_steps": [
{
"phase": "Phase 1: Analysis",
"step_description": "AI analyzes supplier data and provides concentration risk assessment",
"responsible_party": "AI system",
"human_role": "Trigger analysis, monitor for errors",
"inputs": "Supplier dataset from procurement database",
"outputs": "Risk analysis and recommended vendors",
"ai_involved": true,
"review_required": true
},
{
"phase": "Phase 2: Human Review",
"step_description": "Procurement manager reviews AI analysis",
"responsible_party": "Procurement Manager",
"checks": ["Verify vendor capacity", "Confirm analysis aligns with policy", "Check for missing factors"],
"decision_options": ["Approve and escalate", "Request AI revision with feedback", "Escalate with notes"],
"ai_involved": false,
"review_required": true
},
{
"phase": "Phase 3: Escalation",
"step_description": "VP Procurement reviews for strategic/high-value decisions",
"responsible_party": "VP Procurement",
"escalation_triggers": ["Strategic vendor", "Contract > $500K", "Values conflict", "Compliance gray area"],
"decision_options": ["Approve", "Modify and approve", "Reject"],
"ai_involved": false,
"review_required": true
}
],
"if_ai_output_rejected": {
"process": "Feedback provided to AI with specific reasons (e.g., 'vendor lacks capacity for 100 units/month'). AI re-analyzes with corrected constraints. Output is re-reviewed."
},
"metrics_tracked": [
"ai_approval_rate",
"ai_recommendation_value_realized",
"procurement_time_saved",
"vendor_performance_vs_prediction"
]
}
}
```

Change Management for AI Integration

When you integrate AI into an operational process, you're implementing a change. Treat it like any significant operational change: plan the transition, manage stakeholders, measure impact.

Change Management Component 1: Stakeholder Communication**

Who needs to know about the AI integration? Internal teams (procurement, finance), external stakeholders (vendors who will interact with new process). Communicate: What's changing? Why? What does it mean for them?

Change Management Component 2: Training**

What training does the team need? How to evaluate AI outputs, what to watch for, how to escalate if something doesn't look right.

Change Management Component 3: Phased Rollout**

Don't flip the switch on all decisions at once. Start with lower-risk decisions (commodity suppliers, <$100K contracts). Validate that the process works. Then expand to higher-risk decisions.

Change Management Component 4: Success Metrics**

Define what success looks like: time savings, cost savings, quality improvements, stakeholder satisfaction. Measure before and after integration.

Change Management Component 5: Rollback Plan**

If the AI integration isn't working, what's the plan to revert to the previous process? This safety net helps teams trust the integration.

Before AI vs. With AI: Documentation Evolution

Before AI:** Process documentation described how humans made decisions. It was relatively stable. Training a new team member meant reading the runbook and observing someone experienced.

With AI:** Process documentation must cover both AI and human roles. It's more complex but also clearer: "Here's what AI does. Here's what you check. Here's when you escalate." New team members can learn the AI-integrated process faster because AI behavior is explicit.

Failure Scenarios: When Documentation Breaks Down

Scenario 1: Documentation Lag**

AI is integrated into a process, but the runbook isn't updated for six months. New team members read the old runbook and don't expect AI output. They miss AI-generated insights or make errors by not understanding the AI step.

Mitigation: Update documentation before or immediately after AI integration. Make documentation a required part of the integration project, not an afterthought.

Scenario 2: Incomplete Impact Assessment**

You integrate AI into a process but don't document the impact on other teams. Later, you discover that the faster decision timelines create workflow bottlenecks downstream (e.g., faster approvals mean more contracts to execute simultaneously). The downstream impact wasn't anticipated.

Mitigation: Include downstream stakeholders in impact assessments. Ask: How will this change affect your team? What bottlenecks might it create?

Scenario 3: Undocumented Escalation Decisions**

Escalation protocols exist, but escalation decisions aren't documented in process records. Later, someone asks: "When did we decide to escalate this type of decision?" You can't answer because it's not in the process documentation.

Mitigation: Include escalation decisions as part of process documentation. When you decide a certain decision type requires escalation, document it in the runbook.

Building the AI Integration Record

Over time, as you integrate AI into multiple processes, build a master record of AI integration across your operations.

This record should include:

• All processes where AI is used
  - What AI does in each process
  - Impact of each integration (time saved, cost changed, quality impact)
  - Escalation triggers for each process
  - Metrics being tracked for each integration
  - Dates of integration and any major updates

This master record becomes your proof of thoughtful AI integration. Regulators, auditors, or internal reviewers can see: where is AI used, what's the impact, and how is it governed?

Monday Morning to Takeaways

Monday Morning Scenario:** You're onboarding a new procurement manager. Instead of learning the vendor selection process through six weeks of observation, they read the updated runbook which explains: "Here's where AI is used (Phase 1 analysis). Here's what AI outputs. Here's what you verify. Here's when you escalate. Here's what happens if you reject AI output." With clear process documentation, the new manager is productive in days, not weeks. The runbook also serves as your record for auditors: you can show exactly how AI is integrated, what checks exist, and what metrics you're tracking.

Key Takeaways:**

• Update process documentation to reflect AI integration: where AI is used, what AI does, what humans do, when escalation happens.
  - Impact assessments should cover: decision quality, efficiency, costs, risk profile, capability needs, stakeholder impact.
  - Change management for AI integration includes: stakeholder communication, training, phased rollout, success metrics, rollback plan.
  - Build a master AI integration record showing all processes where AI is used, impact, and governance mechanisms.
  - Documentation should be current, not historical. Update it as you integrate AI and as processes evolve.

Maintaining Living Documentation

Process documentation for AI isn't a one-time deliverable. It's a living document that evolves as you learn how the AI-integrated process actually works.

Documentation Review Cycle:** Schedule quarterly reviews of AI process documentation. Ask: Is this still accurate? Have we discovered anything new about how AI is being used? Have we changed the process? What should the next iteration include?

Version Control:** Use version control (date + version number) for process documentation. "V1 dated 2026-04-09" vs "V2 dated 2026-07-15 (added escalation trigger documentation based on 3 months operational experience)." This shows evolution and allows someone to understand what the process looked like at any point in time.

Change Log:** When you update process documentation, include a brief change log: "Updated AI output format description based on feedback from reviewers; clarified when escalation is required; added three new examples." This helps people understand what changed and why.

Input from Frontline Users:** Ask your team what's in the documentation that's unclear or outdated. Frontline users often know where documentation diverges from reality. "The runbook says we escalate strategic vendors to the VP, but we've actually been escalating to the COO." Update documentation to match reality, then clarify authority if the reality is unintended.

Cross-Functional Documentation Coordination

When AI changes an operations process, it often affects multiple teams. Coordinate documentation updates across functions.

Example: Invoice Processing with AI**

If you implement AI to extract data from invoices, documentation updates are needed in multiple places:

• Accounts Payable Process: Invoice receipt, extraction by AI, reviewer verification, payment approval
  - Finance System: How extracted data flows into the GL; reconciliation procedures if extraction is incomplete
  - Vendor Documentation: How vendors submit invoices; what format works best with your AI; what to do if extraction fails
  - Finance Team Training: How to handle exceptions when AI extraction is uncertain
  - IT Documentation: System integration points, data flow, security considerations

If these aren't coordinated, someone gets outdated documentation. Vendors submit invoices in the old format. Finance team doesn't know how to handle extraction exceptions. IT doesn't understand the data flow.

Assign one owner for "AI invoice processing documentation" who coordinates updates across AP, Finance, Vendors, Training, and IT.

Documentation for Different Audiences

One process, multiple audiences, multiple documentation needs.

For Operational Teams:** "Here's the step-by-step process. Here's where AI participates. Here's what you check. Here's when you escalate. Here's what the AI outputs look like and what to do if they don't look right."

For Auditors: "Here's the control architecture. Here's how human review provides control. Here's what escalation triggers mean. Here's how exceptions are handled. Here's the audit trail we maintain. Here's how we test that controls are working."

For Executives: "Here's what changed. Here's the benefit. Here's the risk and how we're managing it. Here's the governance structure ensuring this is working properly."

For Vendors/External Stakeholders:** "Here's how we've improved our process. Here's what this means for you. Here's what didn't change. Here's who to contact with questions."

Each audience sees different documentation from the same underlying process. Create audience-specific documentation rather than expecting everyone to digest comprehensive technical documentation.

Red Flags in Undocumented AI Processes

Watch for these signs that your AI integration documentation has gaps:

Red Flag 1: "Only John knows how this works"** If AI decisions are made through informal process that exists only in John's head, that's a documentation gap. When John leaves or is unavailable, the process breaks. Document it explicitly.

Red Flag 2: Inconsistent Application** Two procurement managers use the same AI tool but in different ways. One escalates certain decisions; the other doesn't. This suggests undocumented process variation. Document the right way, then standardize.

Red Flag 3: Auditor Surprises** An auditor asks about escalation triggers and your team doesn't have a clear answer. This is a documentation gap. Auditors expect to see escalation rules documented, not explained verbally.

Red Flag 4: Training Confusion** When you train new team members on the AI process, do you find yourself saying "well, we actually don't do it exactly like the runbook says..."? That's a gap between documented and actual process. Fix the documentation.

Red Flag 5: Metrics Nobody Tracks** Your process documentation says you track "AI recommendation approval rate" but nobody actually measures it. Either remove it from documentation or add it to your metrics dashboard. Documentation should describe what actually happens, not what should happen.

The Role of Impact Assessment in Continuous Improvement

Impact assessments aren't just for decision-making at the start. Use them to measure whether your AI integration is delivering expected benefits.

Baseline Assessment (Pre-AI):** Before you integrate AI, document the baseline: current process quality, efficiency, cost, risk profile, team capability needs.

3-Month Assessment:** After three months of operation, remeasure. Is quality improving as expected? Is efficiency gain matching projections? What's worse than expected? What's better?

6-Month Assessment:** Did the initial trends continue or reverse? What have you learned about AI performance in your context? What adjustments are you making?

12-Month Assessment:** Full year of data. Actual ROI vs. projected. What worked? What didn't? Should you expand to other processes or dial back this integration?

Build impact measurement into your governance. If you're not measuring impact, you can't know whether the AI integration is actually working.

Building Your Documentation Portfolio

Over time, as you integrate AI into multiple processes, build a portfolio of documented integrations.

Your portfolio should include:

• Process-level documentation for each AI integration (what AI does, what humans do, where AI participates, escalation rules)
  - Impact assessment summary for each integration (baseline, current state, ROI, lessons learned)
  - Change management documentation (who was trained, what training they received, adoption metrics)
  - Risk assessment and mitigation for each integration (what could go wrong, how we're managing risk)
  - Escalation trigger documentation (when this type of decision goes to whom)
  - Master AI integration record (all processes where AI is involved, when each was implemented, current status)

This portfolio becomes your proof of systematic, thoughtful AI integration. It shows regulators, auditors, and new stakeholders that you've thought through the implications and are managing governance carefully.

What to Do Monday Morning

If you don't have documented AI integration for your processes:

• Pick one process where you're using AI (vendor selection, invoice processing, demand forecasting, whatever you're doing).
  - Create a 1-page process map showing: where AI is used, what AI does (inputs/outputs), what humans do, where escalation happens, who decides.
  - Share the draft with 3-5 people involved in that process. "Is this accurate? What's missing? What's wrong?"
  - Refine based on feedback.
  - Create a simple impact assessment: what was true before AI? What's true now? What changed? Any surprises?
  - Put the process documentation and impact assessment in your team wiki or document management system where people can find it.
  - Schedule a quarterly review of the documentation to keep it current.
  - When you integrate AI into your next process, use the same documentation template. Build consistency across your function.

Key Takeaways

• Update process documentation when AI is integrated: where AI is used, what AI does, what humans verify, when escalation happens, what metrics are tracked.
  - Impact assessments should measure: decision quality, efficiency/costs, risk profile, capability needs, stakeholder impact. Measure before and after to know if integration is working.
  - Change management for AI integration includes: stakeholder communication, training, phased rollout, success metrics, and rollback plan if needed.
  - Create different documentation for different audiences: teams need process details; auditors need control architecture; executives need benefits and risks; external stakeholders need transparency about what changed.
  - Build a master AI integration record showing all processes where AI is used, when they were implemented, what impact they're having, and how they're governed.
  - Maintain living documentation. Schedule quarterly reviews to keep process docs current, accurate, and useful for frontline users.
  - Coordinate documentation updates across functions. If AI changes accounts payable, vendors, finance, and IT are all affected. Update all relevant documentation together.
  - Track impact metrics continuously. If you're not measuring whether the AI integration is delivering expected benefits, you can't know whether it's working.

Frequently Asked Questions

Q: How detailed should process documentation be for AI?**

A: Detailed enough that someone new to the role can understand where AI participates and what their role is. You don't need to document the AI's internal logic, but you do need to document inputs, outputs, human review steps, and escalation triggers. If a new team member can read the documentation and do the job, you have enough detail.

Q: What if the process changes frequently because the AI is still learning?**

A: That's okay. Document the current process, but note that it's under active evaluation. Include a review date (e.g., "Process documented 2026-04-09; scheduled for review 2026-07-09 after 3 months of operation"). Update the documentation after each review cycle. Rapid iteration is fine; just document the current iteration clearly.

Q: Do I need to share AI integration documentation with vendors?**

A: Only the parts that affect them. Vendors don't need to know your internal review steps or escalation architecture. But they should know they're participating in an AI-assisted process (transparency is important for trust). You might share: "Your proposal will be analyzed using AI tools that focus on [specific criteria]. A procurement specialist will review the AI analysis. A procurement manager will make the final decision." This tells vendors the process is human-centered while being transparent about AI involvement.

Q: What's the right level of detail for impact assessment?**

A: Specific and measurable. "Efficiency improved" is too vague. "Procurement manager time per decision reduced from 6 hours to 1 hour per decision, freeing approximately 200 hours annually for vendor relationship work" is concrete and trackable. Your impact assessment should have numbers you can measure and track over time so you know whether the integration is actually delivering value.

Q: What if our AI process is different than our documented process?**

A: Update the documentation to match reality. Documented process should describe what actually happens, not what theoretically happens. If there's a gap between documented and actual, it means either: (1) the documentation is outdated (update it), or (2) people aren't following the documented process (retrain them or adjust the process). Don't let the gap persist.