SOPs and reusable templates: writing down how the office runs

Every office runs on knowledge that lives in someone's head: how to set up a new starter, what to do when the post room floods, which supplier to ring when the coffee machine breaks again. When that person is off, things stall. A standard operating procedure (SOP), a written step-by-step guide to a routine task, fixes that. Most offices know they need more of them and never find the time to write them.

AI helps here, as long as you remember where the knowledge comes from. It can interview you, structure your answers and tidy the wording. It can't know how your office does things. An SOP drafted from the tool's general knowledge will be plausible and wrong in the details that matter.

This guide follows Maya Roberts, Office Manager at Fernway Group (our fictional practice company), and Leah Bennett, her Operations Assistant.

Start from how the task is actually done

There are three good ways to get the real process out of your head:

• Talk it through. Dictate or type a rough brain dump as you'd explain it to a colleague. Don't tidy it.
• Write it down while doing it. Next time you do the task, note each step as it happens, including the bits you'd skip explaining.
• Let the tool interview you. Ask it to question you one step at a time. This is often the best method, because it asks about the things you do without thinking.

I want to write an SOP for a task I do regularly: [task, for example "setting up a new starter's desk, equipment and accounts"]. I'll be the source of every step.

Interview me one question at a time. Start by asking what triggers the task and what "done" looks like. Then work through the steps in order. For each step, ask who does it, what they need, and what commonly goes wrong. Also ask about exceptions, who to contact, and deadlines.

Don't suggest steps yourself. When I say "finished", write the SOP using the headings I'll paste below, and mark anything I didn't answer as [TO CONFIRM].

Maya answers about fifteen questions in ten minutes. Several are things she'd never have written down: the building pass form needs the starter's photo two days ahead; Ravi Menon in IT needs the start date a week before to image the laptop; the car park permit comes from the landlord's agent, not Fernway. Then she pastes the headings from the SOP template and asks for the finished draft.

The draft comes back with two [TO CONFIRM] markers and one step Maya doesn't recognise: "Add the new starter to the payroll system." Payroll is Dan Okafor's job in Finance, and Maya never touches it. The tool added it because it's a common onboarding step. She deletes it and adds a line: "Payroll: handled by Finance, not part of this SOP."

Test it on someone who didn't write it

An SOP is only proven when someone else follows it. Leah uses the draft to set up a pretend new starter, reads each step aloud and notes where she got stuck. She finds that "request a pass" doesn't say where the form is. Maya adds the link. That test does more for quality than any amount of prompting.

You can ask AI to help with a first pass:

Organising recurring admin

Some admin is a calendar of jobs that recur: fire alarm tests, first aid kit checks, the insurance renewal, the lease notice date, the annual equipment audit. Maya has these scattered across notes, emails and memory. AI is good at turning a scattered list into an organised schedule.

Below is a rough list of recurring admin tasks for our office, with whatever details I have. Organise it into a table: task, how often, next due date, owner, where the record is kept, notes.

Use only the dates and owners I've given. If a frequency or owner is missing, write [TO CONFIRM]. Then group the tasks by month for the next 12 months so I can see busy periods.

[paste your list]

The reminders themselves belong in your calendar or task tool, not in the AI chat. The tool helps you build the list once. Your normal systems do the reminding. If a recurring task is worth automating (a monthly reminder email, a form that files itself), the Power Automate lessons cover that.

Reusable templates and saved instructions

Once you find yourself writing the same prompt every week, turn it into something reusable. There are two levels.

A prompt template. A saved prompt with [placeholders] you fill in each time. Keep it in a shared document or your prompt library. This is simple, works in any tool and suits most office tasks.

A saved assistant. Most tools let you save standing instructions, and sometimes reference files, as a named assistant: Projects in ChatGPT and Claude, Gems in Gemini, agents in Microsoft 365 Copilot. Every chat with it starts with your instructions already loaded. This suits tasks you do often with the same rules, such as tidying minutes in house format, or answering "how do I…" questions from your own SOPs.

For the full side-by-side, see Custom assistants compared. If you want to build an assistant your whole team uses, the free Team Brain course takes you from choosing the tasks to keeping it honest.

Here's the kind of standing instruction Maya writes for an "office procedures helper" that answers from Fernway's SOPs:

You help staff at a UK office follow our standard operating procedures. Answer only from the SOP documents attached to this assistant.

In every answer, name the SOP and its version number. Give the steps in order, in plain English. If the question isn't covered by any SOP, say "This isn't in our SOPs" and suggest asking the office manager. Don't give general advice or fill gaps from outside the SOPs.

If two SOPs seem to disagree, say so and quote both. If an SOP's review date has passed, mention that it may be out of date.

The saved instructions template gives you a blank structure for writing your own.

Keeping it all current

An out-of-date SOP is worse than none, because people trust it. A saved assistant built on old documents keeps giving old answers, confidently. Build upkeep in from the start:

• Every SOP has an owner, a version number and a review date. The template includes these.
• Update the SOP as soon as the process changes, without waiting for the review date. If Leah finds a step has changed, she tells Maya the same day.
• Replace the files in any saved assistant when an SOP changes, and remove the old version. Two versions side by side confuse the assistant as much as people.
• Keep a short change log at the bottom of each SOP: date, what changed, who changed it.
• Test the assistant after updates. Ask it two or three questions you know the answer to.

AI can help with the review: paste an SOP and ask "list anything in this that refers to a person, system, form or date that might have changed". It can't tell you what changed. It can tell you where to look.

When not to use AI

Data safety

SOPs usually contain internal information: systems, suppliers, sometimes door codes or account names. Leave out passwords, security codes, alarm codes and personal contact details, which should never be in an AI tool or in an SOP stored somewhere widely shared. Only put SOPs into a saved assistant in a tool your organisation has approved, and be careful who you share the assistant with. Your organisation's policy and approved tools still take priority.

What to check

Common mistakes

• Letting AI write the process. It knows how tasks are typically done, not how yours are. The steps come from you.
• Skipping the test. An SOP that only its author has read is an untested draft.
• Building an assistant on stale documents. It will repeat the old process confidently for months.
• Saving confidential details in a shared assistant. Anyone with access can get them back out by asking.
• Over-engineering. Not every task needs an SOP or a custom assistant. Start with the task that causes the most trouble when you're away.
• No owner. An SOP nobody owns is out of date within a year.