AI at work
Useful automation with grounded evidence, controlled actions, and a human decision path.
Updated
An AI feature earns its place when a real responsibility becomes easier to complete and its errors remain manageable. A fast draft is only one part of that result. Somebody still has to check the source, decide whether the output is usable, and carry the work into the next system. Start with that whole path, including the reviewer, before comparing models or counting saved minutes.
These essays examine how to reserve model calls for ambiguous work, keep claims tied to evidence, and control the boundary between reading information and taking action. They also ask what happens to the time recovered. A shorter task can create useful capacity, but the business has to direct that capacity toward work that matters. Privacy choices belong in the same design: local transcription and selective model inputs solve particular problems, rather than removing every data responsibility.
The case studies cover email triage, assisted drafting, meeting support, and social content review. Their value as examples is in the workflow and its boundaries, including where approval remains required. Use this path to choose the smallest useful AI responsibility, define how quality will be checked, and make uncertain output wait for an accountable decision.
Decisions to make
Does this step need a model at all?
Test whether stable rules can handle known categories accurately before adding a model call. Measure missed matches and wrong actions as well as cost. Keep ambiguous cases visible, and compare the complete workflow, including review time, with the original process before deciding that a cheaper or faster result is better.
What evidence must accompany the output?
Identify the source records a reviewer needs to confirm each consequential claim. Preserve links, distinguish missing evidence from a negative finding, and validate fields that can be checked deterministically. Use representative difficult examples to test the workflow; a fluent answer or one successful demonstration does not establish dependable performance.
Who can authorize the next action?
Separate reading untrusted material from permissions to publish, send, or change records. Define the narrow actions that can proceed under validated rules and the conditions requiring review. Show the proposed action and its supporting evidence together, so approval concerns the actual result and uncertain work retains a safe place to wait.
Start reading
- What Happens to the Hour AI Gives Back?
AI time savings become business value when recovered hours reach useful work, improve a customer outcome, or reduce spending the business would actually incur.
- How to Stop AI From Making Things Up
Reduce AI hallucination risk with grounded sources, deterministic checks, claim-level citations, representative evaluations, and accountable human approval.
- Use Simple Rules Before You Pay for AI
Reduce LLM costs by routing predictable work through a keyword classifier first, measuring accuracy, and reserving model calls for ambiguous cases.
- Where AI Phone Agents Work and Where Humans Take Over
What an AI phone agent can and cannot do for lead qualification in 2026: the honest capability map, latency limits, and where a human takes over.
- Keep Call Audio Private With On-Device AI
Keep AI transcription private by running the speech model on the laptop, with no cloud audio upload, plus the limits, hardware cost, and accuracy tradeoffs.
- Edit AI Writing Until It Sounds Human
13.5% of 2024 biomedical abstracts showed an LLM word pattern. Here is how to make AI writing sound human with specific edits that earn reader trust.
- Remove Personal Data Before an AI Model Sees It
Leaky context mitigation needs 5 gates before a third-party model call: allow fields, detect PII, tokenize locally, enforce policy, then serialize.
- Should You Run Hermes Agent Yourself?
A free, self-hosted AI agent from Nous Research that turns repeated work into reusable skills. What it is, who it is for, and what you really need to start.
- An AI Agent Should Never Read and Act in One Step
Prompt injection breaks any AI agent that reads untrusted text and can also act. The fix is architecture: the part that reads holds no tools.
- When Claude Fable 5 Is Worth the Price
Claude Fable 5 costs $10 per million input tokens and $50 per million output tokens. See when its capability, safeguards, and 30-day retention fit.
Systems in practice
- Email Autopilot
An inbox that sorts and files itself, then drafts your replies.
Live in production
- Inbox Triage Pipeline
A full day of email, sorted into a briefing before work starts.
Live in production
- Contracts Brain
Guided contract drafting inside an editor, with controlled storage and human review.
Live in production
- Meeting Intelligence Copilot
A desktop meeting copilot with local transcription and configured cloud suggestions.
Live in production
- Social Media Content Engine
Plan, generate, review, revise, and schedule social content in one controlled workflow.
Publishing gated