← Back to all field notes

An SOP Is Not a Document Pile. It Defines How Far an Agent Can Go

If an SOP is scattered across screenshots, memory, and obsolete files, a capable agent will only execute inside the wrong boundary.

When teams discuss AI agents, they often begin with models, prompts, and tool calls. In practice, the quality of the SOP often determines how far an agent can go.

For an enterprise system, an SOP is not merely an instruction manual. It is the operating boundary.

Four levels of SOP maturity

This is a working framework for deciding automation scope, not an external certification.

LevelCurrent stateAppropriate agent roleWhat is needed next
L1 — Tacit experiencerules live in people and chat historyorganize interviews and samples onlydocument the main flow, input, output, and common exceptions
L2 — Static documentsthe flow is written, but versions and owners are unclearretrieve, summarize, and prepare draftsassign an owner and add versions, effective dates, and exceptions
L3 — Maintainable rulessources, versions, owners, and exceptions are traceablerun a small, reversible pilotdefine permissions, validation, handoff, and write-back
L4 — Executable boundaryinput, output, permissions, validation, handoff, and feedback are explicitown stable nodes inside the boundarymonitor rule changes and exception patterns

Do not give the whole company one maturity score. Evaluate each workflow separately. Meeting notes may be close to L4 while pricing rules remain at L2.

A ten-point pre-launch checklist

Before connecting an agent to a workflow, ask:

CheckPassing condition
Inputrequired, optional, and missing-field actions are explicit
Outputformat, quality threshold, and consumer are defined
Sourceevery critical rule can be traced to an authoritative source
Versionthe active version and effective period are visible
Exceptionsuncovered cases have one recording mechanism
Permissionread, recommend, and write actions are separated
Validationcritical fields, rule citations, and commitments are checked
Handofftrigger, recipient, timing, and context are defined
Write-backhuman decisions and new exceptions return to the system
Ownera named role maintains rules, reviews exceptions, and changes permissions

If one item is unclear, restrict the agent to retrieval, organization, or draft preparation. Turn the missing item into a pilot task instead of leaving it for the model to infer.

SOPs define what the system may use

If a rule exists only in an experienced employee’s memory—or across screenshots, spreadsheets, and obsolete documents—the agent has no stable basis for action. A plausible answer is not the same as an authorized answer.

Ask:

  • Is the rule written?
  • Is the version current?
  • Are exceptions recorded?
  • Who maintains it?
  • Can the output cite the source?

SOPs also define what the system must not do

The most important rules are sometimes negative:

  • fields that may be suggested but not written automatically;
  • customer messages that require approval;
  • financial or contractual actions that remain human;
  • cases that must be escalated;
  • conditions that pause the workflow.

Without these boundaries, better reasoning can create more confident mistakes.

A version-conflict example

Imagine two complete rule documents. The old version remains in a shared folder; the new one was sent as a chat attachment without an effective date. Retrieval finds both, and either answer can appear well supported.

The right fix is not a cleverer prompt:

  1. pause automatic execution for the affected node;
  2. ask the rule owner to identify the active version and date;
  3. mark the old version as expired;
  4. include rule ID, version, and citation in every output;
  5. treat two active matches as a human-handoff trigger.

Executable knowledge needs a lifecycle: who publishes it, when it becomes active, when it expires, and who resolves conflict.

Knowledge maintenance is part of the system

A one-time knowledge cleanup will decay. Production requires:

  • a consistent SOP structure;
  • version and effective-date management;
  • an exception log;
  • a maintenance owner;
  • feedback from human decisions;
  • monitoring for rule drift.

Many successful “agent projects” ultimately create a clearer organizational knowledge system. That is not a side effect. It is often the most durable result.

Next, use the human handoff checklist or see how these boundaries appear in a freight inquiry workflow. The complete sequence is in the AI agent workflow topic.

Continue reading
AI Agents Owning Real Work
How AI Agents Take Ownership of Real Work—and How People Reorganize Around Them Define a deliverable unit of work, expand agent responsibility through evidence-based authorization, and move human effort toward customers, products, judgment, and growth. Turning Freight Inquiries and Quotes into a First AI Workflow Start from the inquiry desk and separate field extraction, rule checks, exception handoff, and result write-back into a measurable workflow. AI Learning Is the Entry Point; Long-Term Service Is the Larger Market Companies usually begin with executive awareness, role training, and small pilots. Learning reveals the deeper demand for workflows, knowledge, and ongoing service.