INSIGHTS
AI & Automation Insights for Operators
Practical guides on AI assistants, workflow automation, lead intelligence, and field operations. Written for business owners and operators who want results, not hype.
RAG Document Permissions: Where ACL Filters Leak
In a RAG system the permission check runs at query time against a copy of the document's access list that was written into the index at ingest time, not against the source system. That copy fails in three ways a correct filter expression cannot fix: chunks that never inherited the parent document's access list, permission changes made at a parent folder or site that never re-sync, and an identity string the application supplies instead of a validated token.
Why Your AI Agent Calls the Wrong Tool
An AI agent calls the wrong tool mostly because two or more of your tools plausibly match the same request, not because you have too many tools. The published thresholds (Claude's tool selection degrades past 30 to 50 tools, OpenAI advises staying under 20) describe context dilution at scale, but the wrong-tool calls most operators hit happen at six to twelve tools and are overlap failures. The fix is an explicit do-not-use boundary in every tool description plus an escape-hatch tool for requests nothing covers.
RAG Reranking: Fix the Wrong-Chunk Problem
A reranker is a second-stage model that re-scores the passages your vector search returned, reading each one against the full query to push the truly relevant chunk to the top. It fixes precision, so reach for it when your RAG finds the right document but ranks the wrong chunk first, not when it misses the document entirely.
How to Delete Customer Data From Your AI Stack
Deleting a customer from an AI system is not one DELETE statement, because the stack has already made derived copies in a vector index, an execution log, a provider-side file store, and sometimes a fine-tuned model. Sort every copy into one of three groups by the verb it supports (delete now, expire on a schedule, or cannot be removed at all), then confirm the expiry schedules are shorter than your statutory deadline before a request ever arrives.
When to Use an AI Browser Agent Instead of an API
Use an AI browser agent when the target system has no API and the task only reads data. Anything that writes should go through an API or a person, because a browser agent finishes a run with a screenshot and a model's opinion rather than a confirmation ID, so you cannot tell a completed action from a half-finished one and cannot safely retry.
Voice AI Latency: What Fits in a One-Second Budget
An AI phone agent has roughly one second between the caller finishing and the agent needing to speak, and 400 to 800 milliseconds of that is the endpointing silence window that runs before any AI does. What is left decides which lookups and tool calls the agent can afford during a live call.
Does Your Chatbot Have to Say It's AI?
In most US states there is no general law requiring a business chatbot to announce it is AI. Four states regulate it, and each uses a different trigger: Maine when a reasonable consumer could be fooled, Utah when the customer asks directly, California only when the bot is used to deceive for a sale or a vote, and Colorado from January 2027 when automated tech drives a consequential decision. A disclosure banner satisfies Maine. The other three are runtime behaviors.
Prompt Versioning: Pin the Model, Version the Prompt
In a production automation, treat the prompt as a versioned artifact, not a string typed into a node field. Store it in a versioned source you own, reference it by version, and log that version on every run, because the prompt is the one reproducibility input that changes only when you change it, so it needs versioning rather than pinning.
Stop a Runaway AI Agent From Burning Your Budget
A runaway AI agent is a loop that never reaches its stopping condition, and a monthly provider spending cap is too coarse and too late to stop one run. Bound each run inside your own loop with two limits at once: a step cap (max_turns or max_iterations) and a per-run token or dollar budget you accumulate every turn. They catch opposite failures, a fast tool ping-pong versus a slow context-growth bleed, so you need both, plus an explicit give-up tool so the agent has an exit besides looping.
Vision Model or OCR? How to Read a Scanned Document
To read a scanned or photographed document with AI, choose the ingestion method per field by how each one fails. A vision LLM returns a confident wrong value with no per-field confidence signal, while OCR returns a per-field confidence score you can threshold and route to review. Use OCR-first for money and ID fields where a silent error is expensive, and a vision model for messy layout, mixed formats, and handwriting.
Why Your AI Automation's Answer Gets Cut Off
An AI automation's answer gets cut off when the response hits the model's maximum output token limit, which is a separate, much smaller cap than the context window. The failure is silent because the API still returns success; the only signal is a stop reason of max_tokens on Anthropic or length on OpenAI, a field most no-code platforms hide. The fix is to read that stop reason, treat it as a failed run, and size the output budget to the task instead of a template default.
Stale RAG Data: When AI Answers From Deleted Docs
A RAG system keeps answering from documents you edited or deleted because most pipelines sync new content into the vector store but never remove the old, leaving orphan vectors that still match queries and get cited as current. The fix is to treat the index as a mirror of your source of record: give every chunk a deterministic ID derived from its source document, delete by that ID before re-upserting on any change, and run a periodic reconciliation sweep to catch what the update events miss.
Want a system like the ones we write about?
We build practical AI and automation systems for businesses that need better workflows, connected tools, and cleaner operations.
Request a Workflow Review