-
When NOT to Use an AI Agent: Why Autonomy Becomes a Liability in Pipelines
AI agents are great when they can decide the order of steps — and a problem when they must not. If a feature needs the same answer every time, an agent planner's freedom is a reproducibility risk, not a feature. Here is where to pull autonomy back.
-
Don't Make the LLM Normalize Your Entities — Land Them on Master Records Instead
LLM entity normalization: when an LLM extracts company names, the same company shows up spelled five ways. The obvious fix is a normalization dictionary — usually wrong. Resolve entities to master records instead. Here is how.
-
Why AI Search Won't Replace Your Keyword Search (It Can't, Structurally)
Once you ship AI search, someone asks: can we retire the old keyword search now? The answer is no — and not for sentimental reasons. The two tools cover different ground, and here is exactly why.
-
RAG Can't Find Your Document? If You Search Summaries, the Summary Is Your Index
When your RAG searches AI summaries instead of the full text, any keyword the summary dropped is unsearchable forever. Here's why RAG over summaries misses documents — and why the fix is upstream, not in the retriever.
-
Who Writes the Prompt, Client or Vendor? How the SI's Role Shifts in AI Projects
In AI projects the client often writes the prompt, so what does the vendor or system integrator actually do? The role shifts from owning expertise to making that intent run inside a system. Here is where the new dividing line falls.
-
Why AI Agents Are Read-Only by Default (and When to Grant Write)
Mature autonomous agents search and read freely but rarely write by default. The reason is an asymmetry in reversibility and audit cost — here is how to reason about the read/write boundary.
-
What Makes an AI Agent Autonomous: Capabilities, Not Coded Tasks
What makes an AI agent autonomous? The capability it runs on — not the tasks you code. Here is which capability is focused on autonomous AI agents, and why agent tasks and agent capabilities are not the same thing.
-
How to Handle AI Agent Permissions: Why Gateway Access Isn't a Data Integration Problem
Connecting a chat interface to an AI agent backend, how do you handle user permissions? It is not a data integration problem — the real work is translating each user's permissions into the agent's tool scope. Here is the on-behalf-of pattern for AI agents.
-
What Makes an AI Agent Real vs Just Workflow Automation
Most enterprise AI agents only route intents to pre-built actions. A real agent composes tools at runtime — its capability ceiling is tools times reasoning, not the builder.
-
Build a Human-in-the-Loop Review Queue for LLM Output
How to review LLM-extracted records without checking every one: route by confidence, queue ambiguous cases worst-first, and keep every merge reversible with append-only provenance.
-
Human-in-the-Loop Traceability for AI-Generated Records
How to let humans correct AI records while preserving source traceability: model the source-to-output link as a first-class relationship object, not a hard foreign key.
-
How to Deduplicate Records with an LLM: Probabilistic Entity Resolution in Production
Using an LLM to merge or deduplicate records is never 100% accurate. Here is how to run entity resolution in production — deterministic blocking, confidence routing, and reversible merges — instead of chasing perfect accuracy.
-
How to Summarize Email Threads Without Losing Decisions
Decision-aware summarization for LLM pipelines: a two-track method that compresses long email threads while preserving every decision, exception, negation, and reversal.
-
How to Clean Enterprise Email for LLM and RAG Pipelines
A practical pipeline to preprocess enterprise email before an LLM or RAG: strip signatures, collapse quoted reply chains, dedupe recipient fan-out, and rebuild threads.
-
Why Summarize (Normalize) Email Before an LLM: 5 Reasons
Five reasons enterprise pipelines summarize and normalize email before sending it to an LLM — context limits, embedding search quality, cost amortization, recipient fan-out dedup, and security egress — plus why you should design it as normalization, not summarization.
-
Why Self-Hosted Open-Weight LLMs Feel Weaker Than ChatGPT
Why a self-hosted open-weight LLM feels weaker than ChatGPT — the open-weight ceiling (mini tier), baked-in quantization, and comparing a bare model against a finished consumer product.
-
Should You Summarize Before Extracting? The LLM Pitfall That Loses Key Details
When an LLM analyzes a summary instead of the raw text, it silently drops negations and reversals — the one line you needed. Here is when to extract from a summary vs. the source text, and how to stop the loss.