Writing standard operating procedures with AI, and keeping them true

Why SOPs rarely get written

In a 20-person firm, the knowledge of how things are done lives in a few people's heads. The office manager knows how to set up a new client in three systems. The senior technician knows the shortcuts for a particular boiler model. The bookkeeper knows which clients need special invoicing. When one of them is out sick or leaves, the business finds out how much it depended on them.

Writing procedures is tedious, and the people who know the work best are the busiest. Documents that do get written are often too long, out of date within a year and stored somewhere nobody looks. AI changes the economics of the first draft, but not the need for ownership and upkeep.

Pick the right procedures first

Do not try to document everything. Start with procedures that are done often, done by more than one person, error-prone or risky if done wrong, and currently known by only one or two people. A short list of ten is a good start.

Examples might include onboarding a new client, processing a vendor bill, closing out a job, handling a complaint, setting up a new employee's accounts, or the month-end close checklist. Ask your team which tasks they most often have to ask someone else about. Those questions are a good map of where SOPs are missing.

There is also a quality cost. When procedures are not written down, each person does the task their own way. Clients get inconsistent experiences, errors are hard to trace and improvements made by one person never reach the rest of the team.

Capture the expert, then let AI draft

The fastest capture method is to record the expert doing the task while explaining what they are doing. Screen recording tools such as Loom, the recording features in Microsoft Teams or Zoom, or Scribe (which captures click-by-click steps) all work. For field work, a phone video with narration does the job.

Feed the transcript or recording into an AI tool, such as Microsoft 365 Copilot, Gemini, ChatGPT Team or Claude, with a clear template and instructions. Ask it to produce the SOP in your standard format, flag any steps where the explanation was unclear and list questions for the expert. This usually produces a usable draft in minutes, compared with hours writing from scratch.

Ask the expert to narrate the why as well as the what. A step that says 'check the client's prior-year file' is more useful with a short reason, such as 'because carryforward figures often differ from what the client remembers'. The reasons help new staff handle situations the SOP does not cover.

• Purpose: what the procedure achieves and when to use it
• Owner and last review date
• Inputs: what you need before starting
• Steps: numbered, one action each, with the system used
• Checks: how to confirm it was done right
• Exceptions: what to do when something unusual happens

Review for accuracy, not just readability

AI produces fluent, confident text, which makes errors easy to miss. It may fill gaps in the recording with plausible steps that your business does not actually follow, or smooth over a judgment call that matters. The expert must review every step against what they actually do.

A useful test is to give the draft to someone who has never done the task and watch them follow it. Wherever they hesitate or ask a question, the SOP needs work. For procedures with regulatory weight, such as handling client tax documents, patient records or trust account transactions, have the responsible professional sign off, and make sure the SOP is consistent with your written policies and obligations.

Keep each SOP short. If a procedure runs past two pages, it is probably several procedures, or it includes background that belongs in a separate training document. Short, focused SOPs are easier to review, easier to update and far more likely to be opened in the middle of a busy day. Ask the AI to split long drafts and to cut explanatory text that does not change what the reader does.

Mind the data you feed in

Recordings of real work often show client names, account numbers, health details or other sensitive information on screen. Before uploading recordings to an AI tool, check that you are using a business account with appropriate data protection terms, that the tool does not train on your data, and that you are complying with obligations such as HIPAA, the FTC Safeguards Rule or client confidentiality rules.

Where possible, record using test data or a dummy client. It is a small extra effort that removes most of the risk.

Store, own and maintain

An SOP that cannot be found does not exist. Store procedures in one place your team already uses, such as SharePoint, Google Drive, Notion, Confluence or the knowledge base in your practice management system. Use consistent names and a simple index by department or workflow.

Give every SOP an owner and a review date, typically every 6 or 12 months, and whenever a related system changes. AI can help with maintenance too: when a process changes, record the new version and ask the AI to compare it with the existing SOP and highlight differences for the owner to approve. Keep a version history so you can see what changed and when.

Make sure people use them

Adoption is where most SOP projects fail. Link procedures from the places people work: a link in the CRM record, a pinned message in Teams, a QR code on the shop wall. Use them in training new staff and in reviewing mistakes. When someone asks a senior colleague how to do something covered by an SOP, the answer should start with the link.

Measure whether it is working: time to train a new hire on a task, error rates on documented procedures and how often senior staff get interrupted with routine questions. InTheory treats SOPs as part of team adoption rather than a documentation exercise, because a procedure only has value if it changes how the work gets done.

Celebrate improvements to procedures, not just their existence. When a technician or coordinator finds a better way to do a documented task, update the SOP and tell the team. It signals that procedures are living documents owned by the people doing the work, which is what keeps them accurate over time.