Why are most SOPs hard to follow?

Most SOPs are hard to follow because they were written from memory by the expert, so they skip the steps the expert no longer notices and say nothing about what to do when something goes wrong. AI SOP writing helps because it is good at structure and at asking what is missing, provided you give it the real process.

A fictional 25-person online homeware shop, Linden and Moss, has one person, Olu, who handles returns. When Olu is on holiday, returns pile up. His notes say: 'check order in Shopify, look at reason. if damaged need photo. refund or replace, usually refund unless they want replace. over 100 quid ask Kate. restock if ok, else to seconds shelf. email customer. update the sheet.' Everything a newcomer needs to ask is hiding in that note: which sheet, what counts as 'ok', and what happens when there is no photo.

How do you write an SOP with AI step by step?

Capture the process as it really happens, ask the AI to structure it into numbered steps with owners and checks, have it interview you about gaps and exceptions, and then test the draft with someone who has never done the task.

1. Capture a real walkthrough

Have the expert talk through one real case, or write notes while doing it. Remove customer names and any personal data before using an AI tool.

2. Ask for a fixed SOP structure

Purpose, scope, roles, numbered steps (action, owner, tool, check), exceptions and escalation, and records to keep.

3. Let the AI interview you

Ask it to list every question a new starter would have after reading the draft. Answer them; that is where the hidden knowledge comes out.

4. Write exceptions as if-then rules

For example: if the item is damaged and the customer has sent no photo, then email a photo request and pause the return for five working days.

5. Test with a newcomer

Ask someone unfamiliar with the task to follow the SOP on a real case while the expert watches silently. Fix every point where they hesitate.

What does a weak and an improved AI-generated SOP look like?

Weak prompt: 'Write an SOP for processing returns.' Typical result: a generic procedure with steps such as 'Receive the return request', 'Verify eligibility according to company policy' and 'Process the refund in a timely manner'. It could belong to any shop, contains no owners, and 'according to company policy' hides the very rules the newcomer needs.

Improved prompt, with Olu's notes and the structure above, followed by the gap interview. The AI asks eight questions, including: what makes an item fit to restock, which sheet is updated and which columns, who covers Kate's approval when she is away, and what the time limit is for returns. Olu's answers produce steps such as '4. If the refund is over GBP 100, send the order number to Kate for approval in the returns channel; if Kate is away, the operations lead approves. Check: approval noted in the returns sheet, column F, before refunding.' The exceptions section now covers missing photos, items outside the returns window and a customer who wants a replacement for an out-of-stock item.

How do you check an AI-written SOP?

Check it by having the expert confirm every step against reality and a newcomer follow it on a live case. An SOP is correct only when someone who has never done the task can complete it without asking for help.

Expert review

The person who does the task confirms each step, threshold and tool name. AI sometimes adds plausible steps that do not exist.

Newcomer test

Note every hesitation and question. Each one is a missing step or an unclear check.

Policy check

Refund limits, legal return rights and safety steps are confirmed with whoever owns those policies.

Owner and review date

Name the SOP's owner and when it will next be reviewed.

What are the common mistakes with AI SOP writing?

The commonest mistake is asking for an SOP from a topic instead of from the real process, which produces a generic document nobody follows. Others: skipping the exceptions, leaving 'according to policy' undefined, letting the AI invent legal or safety requirements, never testing with a newcomer, and storing the SOP where nobody can find it.

Where does SOP writing fit in a Tosheo.ai path?

SOP writing sits in the Tosheo.ai operations path, after lessons on structuring information and checking AI output, and is useful for team leads in any role. The hands-on task asks the learner to turn a rough walkthrough into an SOP, and the rubric feedback checks for missing owners, vague checks and absent exceptions.