AEGIS · Module 4 · Deploy
Artifact 4.3
SOP Updates for AI-Assisted Work
How to rewrite standard operating procedures when AI joins the team — with overlay format, role changes, and audit-ready decision trails.
- Client
- [CLIENT NAME]
- Engagement
- [ENGAGEMENT ID]
- Version
- v1.0
- Issued
- 2026-05-18
Delivered by TechFides under the AEGIS Governance Operating Services engagement. This document is proprietary to the client named above. Redistribution beyond the engagement steering committee requires written consent.
Purpose
Intent — If SOPs don't change when AI enters the workflow, the workflow and the documentation diverge — and the audit, incident, and training problems that follow are all downstream of that gap.
Why SOPs are the forgotten artifact
Most organizations deploy AI into existing workflows without updating the SOP that describes the workflow. The first consequence is training — new hires learn a process that doesn’t match what actually happens. The second is audit — when an incident occurs, the documented process does not describe the failed system. The third is risk — nobody has consciously decided which steps the AI is allowed to change.
What this artifact provides
- A standard overlay format for updating any existing SOP with AI-assisted steps.
- Rules for what must be captured at each changed step.
- Two worked examples — contract review and support triage — that mirror workflows 4.1 WF-01 and WF-03.
- Role-change guidance: where the human role has changed, shifted, or disappeared.
Overlay Format
Intent — A single consistent format across every SOP update. Every step shows five columns: step number, legacy description, updated description, AI role, human role.
Required columns
- Step # — sequential, stable across the legacy and updated versions.
- Legacy step — the original, pre-AI step description. Never deleted, even when the step is fully replaced.
- Updated step — the step as it runs now. Must explicitly name the workflow (WF-NN) and prompt (PR-NNN) where applicable.
- AI role— precisely what the AI contributes at this step. “None” if unchanged.
- Human role — who remains accountable at this step and what changed, if anything, about their role.
Rules for the overlay
- Never delete a legacy step — if it is fully replaced, the cell reads “replaced by step N” with rationale in notes.
- Every AI role references a governed workflow (4.1) and a prompt (4.2). Unreferenced AI use is not sanctioned.
- If a step changes human accountability, it is flagged in the changelog and RACI (1.2) is updated in the same release.
- Every updated SOP has a version + effective date + named approver.
Example 1 — Contract Review SOP Overlay
Intent — Overlays the legacy contract review SOP with the AI-assisted workflow WF-01. Total cycle time dropped 60% while attorney accountability increased, not decreased.
| # | Legacy Step | Updated Step | AI Role | Human Role |
|---|---|---|---|---|
| 1 | Deal desk receives signed NDA and counterparty-drafted MSA. | Deal desk uploads MSA to CLM with tag `ai-review`. System confirms no P0 data present. | Pre-flight check by classifier — rejects upload if P0 / privileged markers detected. | Deal desk (unchanged responsibility, new tooling). |
| 2 | Attorney schedules 4–6 hour clause-by-clause first pass. | Workflow WF-01 runs prompt PR-001 against the attached playbook. Produces a redline summary and risk flags within 8 minutes. | AI generates draft only — produces structured output per PR-001 spec. | Attorney begins from the redline summary rather than blank page. |
| 3 | Attorney annotates findings in Word, emails to deal lead. | Attorney reviews AI redline in CLM. Accepts, edits, or rejects each flag. Accepted flags become the contract redline. | Suggestions only — nothing flows to counterparty without attorney commit. | Attorney holds authorship of record (unchanged). |
| 4 | Deal lead negotiates with counterparty over email. | Deal lead negotiates with counterparty over email (unchanged). | Not involved. | Deal lead owns negotiation (unchanged). |
| 5 | Final version routed for signature via DocuSign. | Final version routed for signature via DocuSign. Post-execution artifact packet auto-generated: AI redline summary, final contract, deviation log (for audit). | Auto-generates audit packet — no decisions made. | GC approves audit packet retention. Retention schedule per Artifact 2.1. |
Example 2 — Support Triage SOP Overlay
Intent — Illustrates a higher-autonomy (T3) workflow. Note that the human is off the per-case hot path but on the aggregate oversight path — a different shape of accountability, not less of it.
| # | Legacy Step | Updated Step | AI Role | Human Role |
|---|---|---|---|---|
| 1 | Ticket enters queue. Tier-1 agent picks next in line, reads, classifies by category. | Ticket enters queue. Workflow WF-03 classifies category, priority, and suggested team within 4 seconds. Regulated-account tag bypasses AI entirely. | Classification only (PR-003). Low-confidence cases route to human triage automatically. | Agent sees AI-suggested classification and can override with one click. |
| 2 | Agent searches knowledge base for related cases, drafts response from templates. | Workflow WF-03 returns a suggested first response pulled from the approved response library. Agent reviews, edits, sends. | Retrieval-based response suggestion only — no free-form generation on customer-visible content. | Agent is author of record for the outbound response. |
| 3 | Agent closes ticket, tags resolution code manually. | Agent closes ticket. AI-suggested resolution code pre-filled; agent confirms or overrides. | Resolution code suggestion from case content. | Agent confirms resolution code — remains accountable for case disposition. |
| 4 | Team lead pulls weekly QA sample for review. | Team lead reviews AI-assisted weekly QA sample dashboard: misrouted tickets, low-confidence overrides, CSAT delta on AI-assisted cases. | Aggregate analytics only — no case-level decisions. | Team lead drives corrective action + prompt refinement signal back to 4.2. |
Role Change Patterns
Intent — Four recurring patterns across clients. Name the pattern explicitly in the SOP so the training, RACI, and performance reviews all match.
A · Role unchanged
Same role, same accountability, same measure of success. AI is a tool the role now uses. Typical for T1 workflows (drafting, notes).
B · Role reshaped
Same role, same accountability, different day-to-day activity. Example: attorney now starts from AI-generated redline, not blank page. Capability requirements shift toward judgment and away from first-pass throughput.
C · Role elevated
Role moves off the per-case hot path and onto aggregate oversight. Typical for T3 workflows. Measurement shifts from case throughput to system health — misroute rate, override rate, drift detection.
D · Role retired
Role no longer needed in its previous form. Rare and sensitive. Explicitly named in the SOP. HR (Artifact 5.3) owns the redeployment path — do not let SOP updates be the first place impacted employees learn about the change.
Governance & Release
Intent — SOP updates are a release event. Treating them that way is what prevents documentation drift from becoming audit exposure.
Release bundle
An SOP update ships as a bundle including:
- Updated SOP document (this overlay).
- Updated RACI entries (1.2) if role changes occurred.
- Updated training materials (5.3) for affected roles.
- Release note published to the affected function.
- Effective date with a minimum 5-business-day notice for non-emergency changes.
Approvers
- Workflow Owner (Artifact 4.1).
- Function Head affected by the SOP.
- AI Governance Lead.
- HR + Comms if any role pattern is D, or pattern C for more than 5 FTE.
Audit trail
Every SOP change is versioned. Effective date, approver list, and rationale are captured at release. Prior versions remain retrievable for the retention period defined in Artifact 2.1.
Anti-Patterns
Intent — Six SOP update mistakes that recur across engagements. Flag any of these and push back.
- “Inserted by AI” footer. A single boilerplate footer on the old SOP does not update the SOP. Every affected step must be rewritten.
- Unnamed AI. “AI may assist with this step” is not a workflow reference. Every AI contribution names a WF-NN and PR-NNN.
- Silent role elevation. Quietly moving a role off the hot path without naming the pattern, updating the RACI, or retraining the role.
- Optimistic checkpoints. Describing a human checkpoint that nobody has time to actually perform. The SOP will be right; the practice will drift.
- No effective date. A change that “takes effect when approved” is a change no one can point to in an incident review.
- Undocumented rollback. If the AI step fails or the tool is retired, the legacy path must still be executable. If not, the SOP is single-vendor risk disguised as a process.
Regulatory Mapping
Intent — SOP coverage of AI-assisted work is a standard line-item in ISO 42001 audits and increasingly in customer security reviews.