--- title: "Why AI agents need SOPs" id: "23890" type: "post" slug: "why-ai-agents-need-sops" published_at: "2026-07-06T13:46:24+00:00" modified_at: "2026-07-06T13:49:34+00:00" url: "https://www.cherryleaf.com/2026/07/why-ai-agents-need-sops/" markdown_url: "https://www.cherryleaf.com/2026/07/why-ai-agents-need-sops.md" excerpt: "AI agents are starting to do real work. They can search systems, draft answers, update records, test software, summarise documents, create pull requests, and trigger other tools. In some organisations, they are already moving from experiment to everyday workflow. That..." taxonomy_category: - "AI" taxonomy_post_tag: - "AI" - "governance" - "SOP" --- AI agents are starting to do real work. They can search systems, draft answers, update records, test software, summarise documents, create pull requests, and trigger other tools. In some organisations, they are already moving from experiment to everyday workflow. That causes a documentation challenge. If an AI agent is going to act on behalf of a team, it needs a clear standard operating procedure. Recent product updates show why this matters: - [GitHub has added Copilot agent session streaming,](https://github.blog/changelog/2026-07-02-copilot-agent-session-streaming-is-now-in-public-preview/) so enterprises can capture prompts, responses and tool calls for audit and compliance. - It has also made [enterprise-managed Copilot settings generally available](https://github.blog/changelog/2026-07-01-enterprise-managed-settings-json-is-generally-available/) , and released [browser tools that let agents interact with live web applications](https://github.blog/changelog/2026-07-01-browser-tools-for-github-copilot-in-vs-code-are-generally-available/) . - [Anthropic’s Managed Agents updates](https://platform.claude.com/docs/en/release-notes/overview) point in the same direction: agents are becoming governed workplace systems. Once an agent can act inside business tools, the organisation needs more than a prompt. It needs an SOP. ## Agents need instructions they can follow Most organisations already have procedures for people. They explain who does the work, when it should happen, what information is needed, what system to use, what good looks like, and when to stop and ask for help. AI agents need the same kind of guidance. They need to know: - What task they are allowed to perform. - Which sources they should trust. - Which systems they can access. - What they must never change. - When a human must review the output. - How to record what they did. Without this, an agent is left to infer the process from a mixture of prompts, interface labels, old documents, chat history, and whatever happens to be available in its context. That is a poor basis for reliable work. The result might look confident. It might even be useful much of the time. But it will be hard to audit, hard to improve, and hard to defend when something goes wrong. ## Old procedures might not be ready for AI Many existing SOPs were written for people who already understand the workplace. They assume background knowledge. They use shorthand. A person can often fill in the gaps. An agent might not. If a procedure says “check the latest policy”, where is that policy? Which version is current? What happens if two sources disagree? If a customer record contains missing data, should the agent stop, ask a question, or continue with a note? These details matter. AI-ready procedures need to be more explicit about sources, decision points, exceptions, and review steps. They also need to separate routine work from judgement-heavy work. A distinction should be documented before the agent is put to work: - Which tasks can be automated? - Which should only be assisted? - Which should stay with people? ## SOPs reduce avoidable risk The risks are not only technical. An agent might give a customer the wrong answer because the knowledge base is out of date. And it might update a record without leaving enough evidence for a later review. These are documentation failures as much as AI failures. A good SOP gives the organisation a control point. It defines the approved way to complete the task. It gives reviewers something to test against. It gives support, compliance, operations, and product teams a shared reference. It also helps avoid a common problem: every team inventing its own way to use AI. Before long, an organisation can end up with several unofficial procedures and no clear way to know which one is correct. An SOP brings the work back into the open. ## The SOP should cover the human work too An AI-agent SOP should not only describe what the agent does. It should describe the whole workflow. For example, if an agent drafts release notes, the SOP should cover the source material, the required structure, the reviewer, the approval route, the publishing location, and the correction process. If an agent answers support questions, the SOP should cover approved knowledge sources, escalation triggers, regulated topics, logging, and how gaps in the knowledge base are reported. If an agent helps write procedures, the SOP should say how source documents are checked, how subject matter experts review changes, and how old versions are retired. The agent is only one part of the system. The procedure needs to cover the system. ## Treat prompts and agent instructions as controlled content Many organisations still treat prompts as personal notes. That is unlikely to work for long. If a prompt or agent instruction affects customer communication, product documentation, compliance work, operational decisions, or internal records, it should be managed like other business-critical content. That means ownership, version control, review, testing, and retirement. It should be clear who can change the instruction. It should be clear why it changed. It should be possible to test whether the new version still produces acceptable results. This is familiar ground for technical authors and content teams. It is content management, but applied to a new type of operational content. ## What an AI-agent SOP should include A useful SOP for an AI agent will often include: - The purpose of the workflow. - The systems and sources the agent may use. - The inputs the agent needs before starting. - The steps the agent should follow. - The outputs it should produce. - The checks a human must perform. - The situations where the agent must stop. - The record that should be kept. - The owner of the procedure. - The review date. For higher-risk work, it should also include approval rules, data handling limits, exception handling, and audit requirements. This does not need to become a heavy document. In many cases, a short, clear procedure is better than a long one. The point is to remove ambiguity where ambiguity creates risk. ## This is a documentation requirement Documentation teams can help organisations decide what content agents should use, how workflows should be described, how outputs should be reviewed, and how knowledge gaps should be fixed. They can also help turn vague AI policies into practical working procedures. That is where many organisations will struggle. They might have an AI policy saying staff must use tools responsibly. They might have enthusiastic teams building agent workflows. What they often lack is the connective tissue between the two: clear operating instructions. And this is where SOPs matter. If an AI agent is going to act, the organisation needs to describe the action. If it is going to make recommendations, the organisation needs to define the evidence it should use. If it is going to produce content, the organisation needs to define the standard that content must meet. AI agents need SOPs because work needs structure. The more capable the agent, the more that structure matters. — Cherryleaf helps organisations create procedures, policies, knowledge bases and AI-aware documentation workflows that people and AI-enabled systems can use with confidence. — We used ChatGPT to generate the image for this post. We also used ChatGPT to search the web for new developments in AI and technical communication. ### Share this: - [Share on Facebook (Opens in new window) Facebook](https://www.cherryleaf.com/2026/07/why-ai-agents-need-sops/?share=facebook) - [Share on LinkedIn (Opens in new window) LinkedIn](https://www.cherryleaf.com/2026/07/why-ai-agents-need-sops/?share=linkedin) - [Share on WhatsApp (Opens in new window) WhatsApp](https://www.cherryleaf.com/2026/07/why-ai-agents-need-sops/?share=jetpack-whatsapp) - [Share on Bluesky (Opens in new window) Bluesky](https://www.cherryleaf.com/2026/07/why-ai-agents-need-sops/?share=bluesky) - ### *Related* ### Leave a Reply[Cancel reply](/2026/07/why-ai-agents-need-sops/#respond) This site uses Akismet to reduce spam. [Learn how your comment data is processed.](https://akismet.com/privacy/) #### Previous post [Claude Tag and technical writing: useful assistant or knowledge trap?](https://www.cherryleaf.com/2026/06/claude-tag-technical-writing/) #### Browse posts by month Browse posts by month Select Month July 2026 (1) June 2026 (14) May 2026 (8) April 2026 (6) March 2026 (14) February 2026 (9) January 2026 (9) December 2025 (7) November 2025 (9) October 2025 (10) September 2025 (7) August 2025 (6) July 2025 (1) June 2025 (3) May 2025 (2) April 2025 (5) March 2025 (3) February 2025 (3) January 2025 (2) November 2024 (8) October 2024 (3) September 2024 (2) August 2024 (3) July 2024 (4) June 2024 (1) May 2024 (2) April 2024 (4) March 2024 (1) February 2024 (2) January 2024 (1) December 2023 (2) November 2023 (9) October 2023 (8) September 2023 (5) August 2023 (3) July 2023 (1) June 2023 (1) May 2023 (2) April 2023 (3) March 2023 (2) January 2023 (2) December 2022 (4) November 2022 (1) October 2022 (1) September 2022 (1) August 2022 (1) July 2022 (1) June 2022 (1) May 2022 (2) April 2022 (1) March 2022 (1) February 2022 (1) January 2022 (2) December 2021 (1) November 2021 (2) October 2021 (1) August 2021 (1) July 2021 (2) June 2021 (3) May 2021 (5) April 2021 (3) March 2021 (3) February 2021 (1) January 2021 (2) December 2020 (4) November 2020 (4) August 2020 (3) February 2020 (3) January 2020 (1) November 2019 (3) October 2019 (1) September 2019 (1) August 2019 (3) July 2019 (2) June 2019 (3) May 2019 (4) April 2019 (1) November 2018 (2) October 2018 (1) July 2018 (2) June 2018 (4) May 2018 (2) April 2018 (3) March 2018 (4) February 2018 (2) January 2018 (3) December 2017 (2) October 2017 (3) September 2017 (1) August 2017 (3) July 2017 (2) June 2017 (7) May 2017 (4) April 2017 (5) March 2017 (1) February 2017 (7) January 2017 (1) December 2016 (3) November 2016 (4) October 2016 (1) September 2016 (5) July 2016 (1) June 2016 (3) May 2016 (1) April 2016 (2) February 2016 (2) January 2016 (5) December 2015 (1) November 2015 (4) October 2015 (2) August 2015 (1) July 2015 (1) June 2015 (2) May 2015 (3) February 2015 (2) January 2015 (6) December 2014 (1) October 2014 (1) September 2014 (4) June 2014 (4) May 2014 (1) April 2014 (2) March 2014 (3) February 2014 (4) January 2014 (4) November 2013 (2) October 2013 (5) September 2013 (3) July 2013 (3) June 2013 (11) May 2013 (4) April 2013 (6) March 2013 (7) February 2013 (4) January 2013 (4) December 2012 (9) November 2012 (23) October 2012 (14) September 2012 (8) August 2012 (4) July 2012 (12) June 2012 (12) May 2012 (10) April 2012 (8) March 2012 (12) February 2012 (3) January 2012 (6) December 2011 (2) October 2011 (3) August 2011 (1) June 2011 (5) May 2011 (2) April 2011 (1) January 2011 (1) December 2010 (1) November 2010 (3) October 2010 (4) September 2010 (8) August 2010 (5) May 2010 (2) April 2010 (3) March 2010 (3) February 2010 (2) January 2010 (3) December 2009 (2) November 2009 (6) September 2009 (1) July 2009 (6) June 2009 (7) April 2009 (8) March 2009 (3) February 2009 (3) January 2009 (3) December 2008 (1) November 2008 (3) October 2008 (3) September 2008 (4) August 2008 (1) July 2008 (3) June 2008 (5) May 2008 (2) April 2008 (2) March 2008 (3) February 2008 (2) January 2008 (1) November 2007 (3) October 2007 (2) September 2007 (1) July 2007 (3) November 2006 (1) October 2006 (2) September 2006 (3) August 2006 (1) April 2006 (1) March 2006 (3) #### Browse posts by category