AI Agent Use Cases: Forty Examples by Team

The AI agent use cases that work share a shape: the agent does the gathering, sorting and drafting, and a person approves anything that leaves the team, spends money, changes live data or decides something about a person. Below are forty examples across nine teams, one line each: the job, what the agent does, and what a person approves. None of them needs a custom-built agent; most run on an assistant such as Claude, ChatGPT or Copilot connected to the tools the team already uses.

Anthropic’s guide to building effective agents↗ (opens in a new tab) says agents add the most value for tasks that “require both conversation and action, have clear success criteria, enable feedback loops, and integrate meaningful human oversight.” Every line below is written against that test. For turning one of these into a card an assistant can finish, see AI agent task examples; for sorting your own jobs into automate, prepare or keep, see which tasks can AI agents automate.

Engineering

Coding agents are the most mature use case, and the one with the clearest check: the tests pass or they do not. How the main agents compare is in best AI coding agents.

1. Dependency updates: the agent upgrades a library, runs the suite and summarizes breaking changes; a person approves the merge.
2. Flaky test hunts: the agent reruns a failing test, finds the pattern and files a bug with the evidence; a person decides whether to fix or quarantine it.
3. Code review first pass: the agent comments on a pull request for missing tests and obvious errors; a person does the real review and approves.
4. Error triage: the agent groups new errors from the error tracker and links each to a bug; a person sets the priorities.
5. Documentation drift: the agent compares the README with the code and drafts corrections; a person approves the wording.
6. Migration scripts: the agent writes and tests a database migration on a dev copy; a person runs it on the live database.

Product

Product work is mostly reading a lot and deciding a little. Agents take the reading. More jobs for whoever runs the plan are in AI for project managers.

1. Feedback synthesis: the agent tags a month of customer feedback by theme with quotes; a person decides what goes on the roadmap.
2. Spec drafts: the agent turns a meeting transcript into a draft spec with open questions; the product owner approves the scope.
3. Competitor watch: the agent summarizes public changelog updates from named competitors; a person decides whether any of it matters.
4. Backlog hygiene: the agent flags stale and duplicate requests with a suggested action; a person closes or merges them.

Support

Support is where agents meet customers directly, so the approval lines matter most here. The full picture is in AI agents for customer service.

1. Ticket routing: the agent classifies incoming tickets by topic and urgency; a person spot-checks the queue each day.
2. Reply drafts: the agent drafts answers from the help center and order history; an agent on the team sends or edits them.
3. Bug reports from tickets: the agent turns repeated complaints into one bug with steps and examples; the engineering lead sets its priority.
4. Help center gaps: the agent lists questions the help center does not answer; a person approves each new article.
5. Refund cases: the agent gathers the order, the policy and the history into one summary; a person approves any refund.

Sales

Sales agents do research and upkeep well and persuasion badly. Anything emailed to prospects also has to follow the FTC’s CAN-SPAM compliance guide↗ (opens in a new tab), which says you “can’t contract away your legal responsibility” by handing the sending to someone else. More in AI sales agents.

1. Account research: the agent builds a one-page brief on a prospect from public sources; the rep decides the approach.
2. CRM upkeep: the agent logs calls and updates deal stages from meeting notes; the rep confirms stage changes.
3. Follow-up drafts: the agent drafts the follow-up email after each call; the rep edits and sends it.
4. Proposal assembly: the agent fills the proposal template from the deal notes; a manager approves pricing and terms.

Marketing

Marketing agents produce drafts quickly, which makes the review step the bottleneck to protect, not remove. See AI agents for marketing.

1. Content briefs: the agent drafts an outline from search questions and existing pages; the editor approves the angle.
2. Campaign reports: the agent pulls weekly numbers from analytics into a summary; a person decides what changes.
3. Social drafts: the agent writes a week of posts from the content calendar; a person approves each before it is scheduled.
4. Link checks: the agent crawls the site for broken links and outdated claims; a person approves each fix.

Operations

Operations use cases are mostly inventories and reconciliations: tedious, checkable and safe when the agent only reports.

1. Vendor renewals: the agent lists contracts renewing in the next 90 days with their terms; a person decides renew, renegotiate or cancel.
2. Access reviews: the agent lists accounts with access to each tool and flags leavers; an admin removes access.
3. Meeting follow-through: the agent turns meeting notes into tasks with owners; each owner confirms theirs.
4. Inventory checks: the agent compares stock counts with orders and flags gaps; a person places the orders.
5. Process documents: the agent drafts a standard procedure from a recorded walkthrough; the process owner approves it.

Finance

In finance the line is simple: an agent can prepare anything, and a person approves anything that moves money or is filed with someone else.

1. Receipt coding: the agent matches receipts to card transactions and suggests categories; the bookkeeper approves them.
2. Invoice matching: the agent matches supplier invoices to purchase orders and flags mismatches; a person approves each payment.
3. Month-end notes: the agent drafts explanations for the largest changes against last month; the finance lead edits and signs off.
4. Payment reminders: the agent drafts reminders for overdue invoices; a person approves who gets one and sends it.

HR

HR is where agents should do the least deciding, because decisions about people carry legal and human weight. Keep agents on paperwork and questions.

1. Onboarding checklists: the agent builds a checklist for a new hire from the role and team; the manager approves it.
2. Policy questions: the agent answers employee questions from the handbook with the section quoted; HR reviews the questions it could not answer.
3. Job description drafts: the agent drafts a posting from the role outline; the hiring manager and HR approve the wording.
4. Interview logistics: the agent proposes interview schedules across calendars; a person confirms them. Screening and hiring decisions stay with people.

The solo founder

For one person running a business, an agent is less a team member than a second pair of hands. Ten jobs are covered in depth in AI agents for small businesses.

1. Inbox sorting: the agent labels email and drafts replies to routine ones; the founder sends them.
2. Weekly review: the agent summarizes the week’s sales, open tasks and overdue replies; the founder picks next week’s three priorities.
3. Quote drafts: the agent drafts quotes from past jobs and price notes; the founder approves every number.
4. Bookkeeping prep: the agent sorts the month’s transactions for the accountant; the accountant does the books.

The pattern in the approvals

Read down the approval column and five things always stay with a person: anything a customer or outsider will read, anything that spends or moves money, anything that changes live data or cannot be undone, anything decided about a person, and priorities. The OWASP Top 10 for LLM applications↗ (opens in a new tab) names the failure when this slips, excessive agency, and recommends human-in-the-loop control so a person approves high-impact actions before they are taken.

The same rule shows up in the protocol most assistants use to reach tools. The MCP specification↗ (opens in a new tab) says “there SHOULD always be a human in the loop with the ability to deny tool invocations.” For a wider view of AI risk in a company, NIST’s AI Risk Management Framework↗ (opens in a new tab) is intended for voluntary use and is a reasonable place to start.

Where to start

• Pick three use cases, one read-only (a report), one that drafts (a reply or a document), and one that changes something with approval.
• Give each one an owner, a person who reads the output every time for the first two weeks.
• Write down what “done” means before the agent starts, so the approval is a check against something, not a feeling.
• Widen what the agent may do only after it has been right for a while, and write down that decision.

Keeping the use cases on one board

Forty use cases across nine teams is a lot of work to keep track of, and most of it is the approvals. On a fenbs board, each use case can be a task with a plan that says what the assistant does and where it stops. A person reads the plan and presses “Let AI do this,” with an optional limits line; any connected assistant can then take it with fenbs_next_approved_task, fill in how it was tested, and move it to Completed. The card says “AI done · check it” until a person presses “I’ve checked it.” Only a person can pre-approve a task, never an assistant.

Standing limits, such as “never email a customer without approval,” go on the Decisions and rules page as rules, and every connected assistant reads the rules first. Every change is recorded in History under the name of the person or assistant that made it. fenbs does not run or schedule the agents; how and when you run them is yours, and the board keeps the queue.