AI Agents in Production Examples: 6 Jobs They Do Today
What are autonomous AI agents actually doing in production today? Less than the demos suggest, and more than the skeptics think. The agents that last in real teams do narrow, repeatable work, and they stop for a human before anything costly happens.
Here is the short answer. In production today, agents mostly sort, draft and watch. They triage support tickets and inbound leads, file invoices, work from on-call runbooks, track competitors, and grade other AI apps with evals. Very few run with no supervision at all.
Top search results list big-company case studies, but you cannot install one. This post maps six of those jobs to a free item on our shelf. We read every file below. We did not run them, and time saved is not measured.
Quick answer
| Job in production | What the agent does | Where a human decides | Start with |
|---|---|---|---|
| Support tickets | Drafts grounded replies and FAQ entries | Refunds, account changes, unconfirmed bugs | Customer support |
| Inbound leads | Ranks leads into four urgency tiers | Before triage starts | Inbound lead triage |
| Invoices | Renames and files receipts, writes a CSV | Before any file moves | Invoice organizer |
| On-call | Writes per-alert response steps | Escalation matrix | On-Call Runbook |
| Market watch | Weekly competitor brief, diff only | Each high signal gets an owner | Competitive Intelligence Monitor |
| Evals | Builds a pytest eval suite for your agent | You commit and rerun it | Deepeval |
How we picked
We kept items where the file defines one narrow job, a fixed output, and a clear point where a person decides. Broad "do everything" agents did not make the list.
1. Support tickets: draft, ground, escalate
Support is the most common production job for agents, and the easiest to get quietly wrong.
Customer support is a Claude Code subagent. It drafts single-ticket replies and FAQ entries, and it searches your existing help docs before writing anything new. Its rules forbid inventing product behavior, root causes or timelines. It pauses for a human on refunds, credits, account deletions, 2FA resets that need identity checks, legal statements and suspected bugs. It runs on Sonnet with Read, Write, Edit, Glob and Grep. Save it as a .md file in ~/.claude/agents/.
2. Inbound leads: who to call first
Inbound lead triage pulls demo requests, trial signups, downloads, webinars and chatbot chats from a period you pick. It merges duplicates by email and sorts every lead into four urgency tiers: a demo request is tier 1, a top-funnel ebook download is tier 4. It checks fit against your ideal customer profile, drafts replies for tiers 1 to 3, and flags competitors and existing customers instead of pitching them. It shows you the lead count and asks before it triages.
3. Back office: invoices and receipts
Finance work is a quiet production win because the rules rarely change.
Invoice organizer reads PDFs and scanned receipts and pulls vendor, date, invoice number and amount. It renames each file to YYYY-MM-DD Vendor - Invoice - Description and sorts it into folders. It shows you the plan before moving anything, copies instead of moving to keep originals, and writes an invoice-summary.csv for your accountant. Files it cannot read are flagged for review, not guessed.
4. On-call: the runbook an agent works from
An agent that touches live systems needs steps a tired human could also follow.
On-Call Runbook Skill writes that document for one service. It asks for your alerts, monitoring stack, failure modes, escalation contacts and rollback method. Then it produces steps per alert, an escalation matrix, incident update templates and a shift handoff. One of its own rules: no step may just say "check the logs". Every step names the exact command or dashboard.
5. Market watch: competitor signals every week
Competitive Intelligence Monitor Skill watches the competitors you list across five signal types: product, pricing, hiring, partnerships and messaging. The first run is a full brief. Later runs show only what changed, capped at 300 words. Every high-priority signal needs a specific action and a named owner, tied to an item on your roadmap.
6. Evals: how teams keep agents honest
Most production stories skip this part. Teams that keep agents running measure them.
Deepeval adds an eval loop to an agent, chatbot or RAG app. It inspects your code, asks intake questions, then builds a pytest eval suite you commit and rerun with deepeval test run. With no dataset, it generates about 30 to 50 goldens through deepeval generate instead of making them up. It needs Python 3.9+, pip install deepeval and model credentials; hosted reports need a Confident AI login.
The pattern behind all six
- One narrow job, not a general assistant.
- A fixed output: a reply, a queue, a CSV, a runbook, a brief, a test suite.
- A stop point where a person says yes.
That is what "autonomous" means in production right now. If your agent idea has no stop point, add one before you ship it.
How to start this week
- Pick the job you repeat most: tickets, leads, invoices or alerts.
- Open that item and run it on last week's real data.
- Add Deepeval once it works, so you know when it stops working.
Browse more on the Claude skills shelf and the agents shelf.
FAQ
What are AI agents doing in production today? Mostly triage, drafting and monitoring: support replies, lead queues, document filing, runbooks, market watch and evals.
Are these free? Yes. Five are MIT licensed and Deepeval is Apache-2.0. Deepeval needs model credentials to score results.