How to build an AI assistant for your SOPs

What this automation does

A new hire on the second shift asks “how do I quarantine a pallet that failed inspection?” in Teams and gets the exact steps from the current SOP, with a link to the source and the date it was last reviewed. If no SOP covers the question, the assistant says so and tells them who to ask.

1. Ingest. On a schedule, the assistant pulls approved SOPs from Notion, Google Drive or SharePoint, along with owner, department, site, last-reviewed date and the document’s access groups.
2. Structure. Each SOP is split by section and numbered step. Every piece keeps the procedure title, section heading and step numbers so an answer can say “step 4 of the Receiving SOP”.
3. Index. Sections are stored with embeddings for meaning-based search and a keyword index for exact terms like part numbers, form codes and machine names.
4. Ask. Staff ask in Slack, Teams or a simple web page, including on a phone at a workstation.
5. Retrieve with permissions. Search returns only sections the user is allowed to see, filtered by site and role when the SOP applies to only some locations.
6. Answer with citations. The model gives the steps, quotes safety-critical steps exactly, links the source and shows its review date.
7. Refuse and route gaps. When nothing relevant is found, the assistant says no procedure covers it, points to the SOP owner, and logs the question as a content gap.

When it is worth building

This is worth it when you have dozens or hundreds of written procedures, frequent onboarding, several sites or shifts, and supervisors who answer the same questions over and over. Regulated operations, such as food production, healthcare support functions and quality-controlled manufacturing, benefit most, because following the exact procedure matters.

Do not start here if your SOPs are few, outdated or mostly in people’s heads. An assistant over bad documents gives confident bad answers. Write and clean the procedures first. And if your company already uses an AI assistant that searches your documents with citations, test it on 30 real questions before building anything. For assistants over broader company knowledge rather than procedures, see our general RAG assistant guide. If you are unsure whether your documentation is ready, an AI readiness assessment is a cheaper first step.

Architecture

Choosing between pgvector and a managed vector database is covered in our Pinecone vs pgvector comparison.

How to build it

1. Audit the SOPs. Find duplicates, conflicting versions (“Returns SOP v3” next to “Returns SOP FINAL (2)”), missing owners and documents nobody has reviewed in years. Archive what is dead. Expect this to take longer than the build.
2. Standardize metadata. Each SOP needs an owner, approval status, last-reviewed date, and the sites and roles it applies to.
3. Collect real questions. Ask supervisors and trainers for 50 to 100 questions staff actually ask, each paired with the SOP and step that answers it. This becomes the test set.
4. Ingest and split by step. Keep step numbers and headings with every piece. Never split a numbered step across two pieces.
5. Tune retrieval before prompts. Measure how often the right SOP section appears in the top results for each test question. Fix retrieval first, because no prompt can rescue a missing source.
6. Write the answer rules. Answer only from retrieved text, cite document and step, quote safety and compliance steps word for word, show the review date, and refuse when nothing relevant is found.
7. Connect Slack or Teams. Map chat users to directory groups so permission filters work, and add a thumbs up or down with an optional comment.
8. Pilot with one department. Review every question and answer weekly for the first month. Send unanswered questions to SOP owners as a writing backlog.

What breaks in production

• Conflicting versions. If drafts and old versions are indexed, the assistant will quote the wrong one. Index only documents marked approved and current.
• Paraphrased safety steps. A model summarizing a lockout procedure can drop a step. Mark safety and compliance sections so they are always quoted exactly.
• Permission leaks. Management-only procedures such as pricing overrides surface for everyone if access is handled in the prompt. Filter at retrieval.
• Instructions in images. A flowchart that holds the actual decision logic is invisible to text-only parsing. Describe images at ingestion and check those answers specifically.
• Improvised answers to edge cases. Staff ask about situations the SOP does not cover, and a helpful model fills the gap. The refusal rule must hold even when a plausible answer exists.
• Low adoption. People keep asking the supervisor out of habit. Put the bot in the channels where questions are already asked, and have supervisors answer with the bot’s link at first.

Timeline and cost

A single source system, one chat platform and a few hundred SOPs typically take three to six weeks, with the SOP audit running alongside the build. Lumeor builds SOP assistants in the range of $12,000 to $35,000. Several source systems, image-heavy procedures, multiple sites with different procedures and multilingual support move it toward the upper end. Ongoing model and hosting costs for a team of a few dozen active users usually stay modest. For a fuller breakdown, see what a RAG knowledge assistant costs.

How to measure success

• Share of answers that cite a source, and accuracy on the test question set
• Share of questions with no matching SOP, which should fall as gaps are filled
• Supervisor interruptions per shift for procedural questions
• Time for new hires to work a station independently
• SOPs past their review date
• Thumbs-down rate and the reasons given