How to Document an SOP So AI Can Actually Follow It

Most Standard Operating Procedures are written for humans who already know the job. They read like compliance wallpaper: long narrative, implicit assumptions, and a photo of a screen from 2019. That is unusable for AI assistants and agents, which do not share your tacit context -- they only execute what is explicit, ordered, and testable.

An agent-ready SOP is a process document with a clear trigger, required inputs, numbered steps, explicit edge-case branches, and a verification check at the end. When those elements are present, an AI agent or assistant can follow the procedure without improvising your finance rules or your brand voice.

This guide shows how to rewrite (or write) an SOP so a model can follow it reliably. It pairs with how to build a company knowledge base for AI for where SOPs live, AI operations assistant for who runs them day to day, and how to automate sales with AI when playbooks span CRM and email. For "rules vs. judgment," see AI agents vs. automation and AI agents vs. chatbots.

Why "more detail" is not the same as "agent-ready"

Teams confuse length with clarity. A forty-page PDF can still omit:

• What triggers the process (event vs schedule vs threshold)

• What "done" looks like (artifacts, states, approvals)

• What to do when data is missing (pause, escalate, default -- which default?)

• What must never happen (double pay, public reply, delete production)

Agents need decision boundaries, not office memoirs.

The anatomy of an agent-ready SOP

If you cannot fill Verification, you do not have an SOP -- you have a vibe.

Step 1: Start from one real run, not from imagination

Pick the last time someone did this process correctly. Reconstruct:

1. Trigger -- what exact signal started work?

2. Artifacts -- PDF, row in sheet, ticket ID, email thread?

3. Decisions -- where did they branch (amount thresholds, geography, plan tier)?

4. Handoffs -- who received output and did they accept it?

Now write the SOP backward from that run. Fiction SOPs are the number-one cause of "the agent did something crazy." If you are new to Agently's model of work, core concepts explains how tasks, context, and tools fit together.

Step 2: Write steps as imperative, testable commands

Humans tolerate "handle appropriately." Models will guess.

Each step should answer: What do I click, write, or query? If the answer is "it depends," that is an edge case subsection, not a shrug in step 4.

Step 3: Encode edge cases as explicit branches

Use a simple pattern models parse well:

IF [condition] THEN [action]
ELSE IF [condition] THEN [action]
ELSE [default action + who to notify]

IF [condition] THEN [action]
ELSE IF [condition] THEN [action]
ELSE [default action + who to notify]

IF [condition] THEN [action]
ELSE IF [condition] THEN [action]
ELSE [default action + who to notify]

Defaults must be safe. Example: "If tax jurisdiction unknown, do not post; assign to tax queue." Unsafe defaults (assume US, assume standard VAT) create silent wrongness.

Step 4: Separate "policy" from "procedure"

Mixing them causes stale procedures: people skip the doc because half is obvious law and half is wrong screenshots. Cross-link instead: "Follow payment policy POL-FIN-02; this SOP is execution only."

Step 5: Add verification and rollback

For anything that touches money, access, or customers:

• Verification: "Balance matches bank feed within threshold X; exceptions listed in report R-AP-01."

• Rollback: "If batch fails after step 7, run reverse_batch job per RUNBOOK-14; never delete rows manually."

Agents (and junior humans) need panic behavior defined. Customer-facing SOPs often sit next to AI customer support agent workflows; internal finance SOPs may never touch a customer channel -- say so explicitly.

What to cut from legacy SOPs before giving them to AI

How Nova uses this in Agently

Nova is Agently's operations AI employee: it can execute recurring workflows across tools when your SOPs are clear enough to map to actions -- and when your Brain holds the policies and templates humans already trust. Integrations show which systems you can connect; Getting started walks through first setup. Try Agently free.

Frequently asked questions

How long should an SOP be?

As short as possible while preserving branching and verification. Many operational SOPs land at 1 to 3 pages when rewritten. If yours is longer, split it by trigger so each document covers one clear starting event.

Should SOPs be Markdown, Notion, or PDF?

Editable text (Markdown, Notion, Google Doc) beats PDF for maintenance. PDF is fine as a signed output customers receive -- not as your internal source of truth that needs weekly updates.

Who writes the first draft?

The person who last did the job correctly -- not the manager who has not clicked the tool in two years. Subject-matter expertise beats seniority for SOP accuracy.

How do we prevent agents from bypassing approvals?

Encode approvals in the system (permissions, dual control) where stakes are high. Do not rely on prompt text alone -- system-level controls are harder to skip than instructions in a document.

Can one SOP cover "everything in customer success"?

No. Split by trigger: onboarding complete, renewal N days out, churn risk score above threshold. Monolith SOPs become unreachable for both humans and models.

Agently turns clear SOPs into repeatable AI-assisted operations -- without losing human oversight. Try it free.