Write an AI Agent SOP: The One-Page Template for Owners

An SOP, a standard operating procedure, is a written page that says exactly what an agent may do, what it must ask about, and who is responsible. Without one, the agent runs on whoever configured it remembers, which means nobody can check its work, nobody knows when it oversteps, and every fix starts from scratch. With one, a weekly review takes 30 minutes and incidents have a paper trail.

This guide gives you the one-page format, a filled example for inbox triage, and a blank template to copy for every task you delegate. One page per task, plain language, no jargon.

Key takeaways

• One SOP per task, one page each: purpose, trigger, inputs, allowed actions, approval tiers, escalation, logging, and review cadence.
• Name an owner on every SOP, because ownerless agents are a recognized risk and neglected agents drift within weeks.
• Approval tiers are the core of the page: what runs free, what needs a yes, and what is blocked entirely.
• Review weekly for the first month, then monthly, changing one thing at a time based on the error and escalation log.
• Copy the enterprise discipline at small scale: Anthropic's Cowork rollout guidance and Microsoft's Agent 365 both treat documented procedures and lifecycle controls as foundational.

Why written procedures beat verbal instructions

Verbal instructions evaporate. The person who set up the agent remembers telling it to ask before sending quotes, the agent's configuration says something slightly different, and three months later nobody can reconstruct which one governed the Tuesday incident. A written page settles all of that. It is also how you train the humans: the staff member clearing approvals reads the same page the agent was configured from, so disagreements surface as edits rather than arguments.

The enterprise world confirms the principle at scale. Anthropic's rollout guidance for Claude Cowork structures deployments around maturity levels and a months-long roadmap, and Microsoft's Agent 365 governs agents through a registry, per-agent identity, and audit. A five-person shop does not need that machinery, but it needs the same three properties: every agent is listed, every task has written rules, and someone reviews the record.

The eight blocks of the one-page SOP

Every SOP has the same eight blocks. Keep each to two to four lines. If a block overflows, split the task.

1. Purpose. One sentence saying what job this SOP covers and what good looks like. Example: triage every inbound inquiry within five minutes so nothing waits overnight.
2. Trigger. Exactly what starts the work: a new email, a missed call, a time on the schedule, a pipeline change. If the trigger is vague, the agent will either idle or fire at the wrong moments.
3. Inputs. The data the agent may read: which inbox folder, which pipeline stage, which calendar, which price sheet. Name them precisely so scope creep is visible.
4. Allowed actions. What the agent may do freely, such as drafting replies, labeling threads, and updating custom fields. Everything not listed here is not allowed without approval.
5. Approval tiers. Three lists: runs free, needs a yes, blocked entirely. Sending to customers, spending, deleting, and external sharing belong in the middle or last list by default. Our approval workflows guide shows how to design the tiers.
6. Escalation. Who the agent hands off to, under what conditions, and how fast a human must respond. Include a fallback contact for when the primary owner is unavailable.
7. Logging. What gets recorded per action, where the log lives, and how long you keep it. Minimum: timestamp, action, data touched, approval reference.
8. Review cadence. How often the owner reviews this task, which numbers they check, and what triggers an SOP edit versus a configuration fix.

Add a header line with the task name, the owner, and the last-reviewed date. That is the whole page.

Filled example: inbox-triage SOP

Below is a complete example for the most common first delegation: triaging inbound inquiries. Copy the shape, replace the contents.

Task: Inbound inquiry triage. Owner: Maria, office manager. Last reviewed: September 30, 2026.

Purpose. Every new inquiry gets acknowledged within five minutes and routed to the right next step the same business day, so no lead waits overnight.

Trigger. A new message arrives in the shared inbox, a website form submits, or a missed call creates a thread.

Inputs. Shared inbox unread folder, website form notifications, yesterday and today's calendar availability, current price sheet v4. No access to payroll, banking, or personnel files.

Allowed actions. Send the acknowledgment template immediately. Label threads as new, existing customer, or spam. Draft a reply with a proposed time slot. Update the CRM stage to contacted.

Approval tiers. Runs free: acknowledgment, labeling, drafting, CRM updates. Needs a yes: sending any non-template reply, quoting a price, booking the job. Blocked: deleting threads, refunding, sharing internal notes externally.

Escalation. Unhappy customers, requests the template cannot answer, and anything mentioning legal or safety issues go to Maria within one business hour with a drafted reply attached. If Maria is out, escalate to the owner by text.

Logging. Every action logged with timestamp, thread, action taken, and approval reference. Log exported to the shared drive every Friday.

Review cadence. Weekly for the first month: check response times, sample five drafts, confirm approvals are genuine. Monthly after that. Edit the SOP when the same escalation repeats twice.

Notice what the page does not contain: model names, vendor settings, or implementation details. Those change. The rules should survive a vendor switch, which is also why the SOP lives in your own documents, not only inside the agent's configuration screen.

The blank template to copy

Copy everything between the lines for each new task. Fill every block; delete nothing.

Task name: ___. Owner: ___. Last reviewed: ___.

Purpose: ___ (one sentence plus what good looks like).

Trigger: ___ (the exact event or schedule that starts the work).

Inputs: ___ (named folders, pipelines, calendars, documents; nothing else is readable).

Allowed actions: ___ (what runs free; everything else needs approval).

Approval tiers: Runs free: ___. Needs a yes: ___. Blocked: ___.

Escalation: ___ (who, under what conditions, response time, fallback contact).

Logging: ___ (what is recorded, where it lives, retention).

Review cadence: ___ (frequency, numbers checked, what triggers an edit).

Two tips from experience. First, write the approval tiers before anything else; they are the page's load-bearing wall. Second, keep a changelog of three lines at the bottom: date, what changed, why.

Rolling SOPs out across the first 90 days

Do not write ten SOPs in week one. Write one for the first delegated task, run it for two weeks, and let the review teach you what the second SOP needs. A realistic pace is one SOP per task as each task goes live, roughly one every two to three weeks, with the weekly review covering all live SOPs in a single 30-minute session.

Tie the rollout to your broader plan. Our first-90-days rollout sequences scope, access, and review cadence week by week, and the internal assistant workflow shows how these pages map onto a live setup. If a task keeps breaking its SOP, the usual causes are a trigger that fires too broadly, inputs that include messy data, or an approval tier set to automatic that should require a yes. Fix the page first, then the configuration, and note the change in the log. The World Economic Forum found only 16 percent of organizations have fully redesigned roles and processes around AI, so expect the SOPs to evolve for several quarters; that is normal operations, not failure.

For teams running a dedicated setup, our internal assistant service ships with SOP-shaped configurations for the common workflows, which shortens the first draft of each page considerably.

Common SOP mistakes

Mistake	Why it hurts	Fix
One page per agent instead of per task	Triggers and approvals blur across jobs	Split by task before going live
Approval tiers left blank or all-automatic	First incident has no paper trail	Default sending, spending, deleting to require approval
No named owner	Reviews and offboarding never happen	Name the owner in the header
Inputs listed as everything	Scope creep becomes invisible	Name exact folders, stages, documents
No review cadence	Drift accumulates for months	Weekly for a month, then monthly
SOP lives only in the vendor tool	Lost on cancellation or switch	Keep a copy in your own documents

If you recognize your setup in this table, fix the lowest row first and work upward. Each correction takes less than an hour.

Next step

Pick the single task closest to going live and draft its SOP today using the blank template above. Time yourself: a good first draft takes under an hour. For the broader rollout sequence, read the 90-day plan alongside this template, and if you want a second pair of eyes on the approval tiers, the free six-step AI automation plan on our homepage is the fastest route: start your AI automation plan. To review your drafts with someone, book a call.

FAQ

What goes into a one-page AI agent SOP?

Eight blocks: purpose, trigger, inputs, allowed actions, approval tiers, escalation rules, logging, and review cadence, plus the owner's name. Each block is two to four lines. If a block does not fit on the page, the task is too broad and should be split before the agent goes live.

Why do AI agents need written procedures at all?

Because unwritten rules are unenforced rules. Enterprise rollouts treat documented procedures as the foundation: Anthropic's deployment guidance for Claude Cowork includes a five-level maturity model and a six-month roadmap, and Microsoft's Agent 365 governs agents through registry, identity, and lifecycle controls. Your one-page SOP is the small-business version of the same discipline.

How often should I review an agent SOP?

Weekly for the first month, monthly for the next two, then quarterly once the numbers hold steady. Review the error rate, the approval queue, and any escalations since the last check, and change one thing at a time. Only 16 percent of organizations have fully redesigned roles and processes around AI, so expect several rounds of adjustment.

Should each agent task get its own SOP?

Yes. One page per task, not one page per agent. An agent handling inbox triage, quote follow-up, and review requests needs three SOPs, because the triggers, data, approvals, and escalation paths differ. Shared pages blur the rules and make incidents harder to trace back to a cause.