For business owners and operators exploring where AI agents and automation can improve the work they know best.
29 essays · Strategy, systems and execution
Lead essayMy Codex agent checks Garmin, Strava, and Runna every night, then compares each week with five years of training history. I always know where I left off, which makes it easier to shut the laptop.
Read the essay02
Browser pages contain prompt injections and sensitive data. Development toolchains contain extensions, actions, and secrets. Agent security needs input sanitation, narrow permissions, explicit scope, and a record of every action.
03
A lyrical prompt may produce a beautiful image and fail the brief. Production design needs controlled layout, approved colors, editable structure, repeatable assets, and a review process that can explain what changed.
04
Agents cannot depend on hardcoded service URLs, relearn the same website each morning, or accept one opaque search bundle. Reliable systems need verifiable discovery, reusable skills, and retrieval primitives they can inspect.
05
An agent can implement a ticket and miss why the team made the decision, which customer promise shaped it, or what depends on the file it changed. Fast code needs product memory and a current system map.
06
Code that runs can still leave bad tests, stray edits, weak architecture, and more cleanup than progress. Teams need supervision, real scenarios, scope checks, and a standard a maintainer would accept.
07
A billing issue that sounds like one task can involve identity checks, account history, policy exceptions, and a human approval. Workflow observation finds those steps before an agent does.
08
A working model call proves the feature. A company still needs identity, billing, taxes, support, model routing, cost controls, distribution, and an answer when the customer disputes the bill.
09
Alerts and traces tell you that something went wrong. A production learning loop groups repeated failures, prepares a fix, captures the reviewer’s correction, and turns the pattern into an evaluation.
10
A workflow can reason well and still act on yesterday’s price, policy, or documentation. Reliable automation needs a plan for freshness, meaningful change detection, and the moments when the machine should stay asleep.
11
AI can turn a brief into a dashboard, planner, or review hub before the old procurement meeting ends. Cheap creation also produces abandoned tools, unclear permissions, duplicate data, and nobody responsible for maintenance.
12
A customer searches for "that woven side panel" while the catalog expects "rattan." Visual and conversational search can bridge the language gap when product attributes, imagery, and taxonomy are clean.
13
Commerce agents attract attention on the buyer side. Sellers face the harder operating mess: listings, inventory, ad budgets, pricing, fulfillment, and support spread across systems that drift out of sync.
14
To manage approval of AI-generated content, give each item an owner, assign reviewers for the claims it makes, and record their decisions against a specific version. Let the publisher use only that approved version. Keep the brief, evidence, edits and publication result in the same record so nobody has to reconstruct permission from a chat thread.
15
A company does not need one giant memory file shared with every tool. It needs current context, clear ownership, and rules that decide which assistant can see which slice of the business.
16
Metered AI turns model choice into workflow design. Teams can route routine work to smaller or local models, reserve expensive reasoning for ambiguous cases, and place hard budgets around long agent runs.
17
A tool can save time and increase anxiety at the same time. Leaders need to explain where the saved work goes, how responsibilities change, and how early-career employees will build judgment when routine work disappears.
18
A small company does not need an enterprise committee for every workflow. It can still borrow the useful habits: narrow scope, named owners, access rules, review gates, and enough logging to explain a bad result.
19
A polished research answer can still combine stale pages, weak sources, and claims nobody can reconstruct. Business research needs a trail from each conclusion back to the pages and dates that support it.
20
An AI assistant can answer a business question in seconds. The answer still fails if finance, sales, and marketing use different definitions or if the assistant can see data the person asking should not see.
21
A persistent agent can update the CRM, open a ticket, change project state, and draft a response. The company needs one place to see its identity, authority, work state, failures, and human approvals.
22
A single prompt can hide the request, assumptions, checks, and side effects inside one polished answer. A staged workflow leaves evidence at each step so people can review the work before it reaches production.
23
Generic models know accounting, sales, and operations in the abstract. They do not know why your team treats one vendor, customer, or approval differently. That buried logic determines whether the workflow works.
24
Your team saves 25 hours a month. The software costs $125. You also paid for setup, and someone spends two hours a month keeping the workflow running. To judge the return, include every cost and establish what the recovered time is worth.
25
Month-end close concentrates repetitive work, missing documents, reconciliations, and late decisions in one miserable week. It also gives automation teams clear evidence, owners, and measurable outcomes.
26
The sale happened in email, calls, and someone’s memory. The CRM waited for a person to narrate the work afterward. A trustworthy automation captures evidence, stages updates, and shows who changed what.
27
Meeting transcripts solved the memory problem. They did not solve the handoff. The next useful layer drafts follow-up, stages CRM updates, creates tasks, and keeps a human in charge of the send button.
28
Small and midsize teams are no longer asking whether AI belongs in operations. They are asking where AI agents for small business teams remove coordination drag first. The biggest wins are showing up in follow-up, intake, reporting, and internal execution support.
29
Choose a recurring task your team can measure, define where a person needs to review it, and calculate whether the time recovered justifies the cost. A worked retail example shows how.
