Support ticket triage
Read every new ticket, route it by topic, and draft replies to the easy ones.
Inbox→Agent→Drafts
Business agents,
in plain English.
What they are, how they work, what they actually look like in production, and how to pick the one you should build first. Skim the pictures. Read the parts that catch you.
Chapter 01
The anatomy every agent shares
Strip away the marketing language and every working business agent has the same five parts. If you can describe these five things for one of your own tasks, you have an agent spec.
A worked example
Daily revenue snapshot agent
Trigger
Every weekday at 07:30
Read
sales.daily_agg, sales.hourly_live, same weekday last year
Guard
Skip if rows(yesterday) < 70% of expected
Compute
Revenue Δ, orders Δ, AOV Δ vs LY
Post
→ Slack #finance with a formatted snapshot
Chapter 02
Pick the right first agent
Most projects fail because the first agent was the wrong agent. Plot your candidates on effort vs value and start in the top-right quadrant.
Repetitive
It happens at least daily and the answer is roughly similar each time.
Clean data
The inputs already live somewhere accessible. Don't fix data pipelines and build agents in the same project.
Safe to fail
A wrong output causes embarrassment, not financial or legal harm.
Measurable
You can count the hours per week it saves, in numbers, by the end of month one.
Chapter 03
Twelve patterns, one anatomy
The same five-part shape, applied to the jobs that actually move the needle in a small business. Pick the row that sounds most like your week.
Returns & refund classifier
Sort returns by reason, flag the suspicious ones, surface trends to the buyer.
Returns→Agent→Tags + alert
Review & NPS responder
Watch reviews, draft a calm reply to negatives, summarise themes weekly.
Reviews→Agent→Drafts
Daily revenue snapshot
Yesterday's numbers, vs last year, with one clear sentence on why.
Sales DB→Agent→Slack
Inventory watchdog
Spot SKUs about to stock out, surface a tight shortlist to purchasing.
Stock + sales→Agent→Shortlist
AR / collections chaser
Draft polite reminders for overdue invoices on a quiet schedule.
Invoices→Agent→Email drafts
Metric anomaly detector
Notice when a number drifts outside its normal band, name it in plain English.
Warehouse→Agent→Alert
Ad ops monitor
Watch ROAS and spend, flag the campaigns that drifted overnight.
Ads API→Agent→Slack
Competitor watcher
Notice price, copy, and PDP changes on a curated list of rivals.
Scrape→Agent→Digest
Meeting to action items
Turn a transcript into a short shared note: decisions, owners, deadlines.
Transcript→Agent→Doc
Compliance reviewer
Pre-flight check a document against a policy and flag the gaps.
Doc→Agent→Review
Daily briefing
One short morning note: yesterday's numbers, today's calendar, this week's risks.
All of it→Agent→Email
Chapter 04
Five questions before you build
Run your shortlist through these, in order. Stop at the first "no" and pick a different candidate.
1
Does this task happen at least 3 times a week?
If no → reconsider. The win compounds with frequency. Once-a-month tasks rarely justify the build.
2
Can I describe the data sources in one sentence each?
If no → the data isn't ready. Fix the data plumbing first as its own project.
3
If the agent gets it wrong, is the fallout recoverable in under an hour?
If no → skip for the first agent. Build something where embarrassment is the worst case.
4
Can I write the exact format of what the agent outputs?
If no → you're not ready. If you can't describe the output, you don't know the task well enough to delegate it.
5
Will I know in week one whether it's working?
If no → measure first, build later. Most agent failures are actually measurement failures.
Chapter 05
Four ways agents fail
The same four mistakes account for most failed agent projects. Knowing them up front is worth more than another tutorial.
Pitfall 01
Building "an AI for the business"
Pick a narrow, scheduled, boring task first. The owner who tries to build one intelligent assistant that does everything ships nothing.
Pitfall 02
Skipping the guard
The agent will run on a day when its inputs are missing. Without a guard it publishes garbage with full confidence.
Pitfall 03
No measurement plan
"It feels useful" is not a result. Decide on day one how you'll measure hours saved or money moved by week four.
Pitfall 04
Letting it autonomously act
For agent #1, the agent always drafts and a human always approves. Earn the trust over weeks, not hours.
A working agent is a scheduled task with a clear input, a guard, and one output channel.
Everything else is decoration. If you can describe one of those for your business by tonight, you're closer to shipping than 90% of operators talking about AI.