This article is part of our AI Agents for Business hub. Want help applying it? Talk to eSEOspace.
Automate a job first that happens often, carries little risk, and is easy to check and undo. Good first jobs prepare work for a person instead of acting directly on customers or money.
That is less exciting than a fully autonomous agent, but it is how you learn safely. This guide is part of our AI agents for business hub, and its sources were checked on 9 November 2026.
Key Takeaways
- Pick a job that is frequent, low risk, easy to check and easy to reverse.
- Anthropic recommends the simplest solution possible, so check whether plain automation is enough.
- The Census Bureau found AI at work is used for search, writing, ideas, summaries and administrative tasks.
- Anthropic's own traffic data suggests only 0.8% of agent actions appear irreversible, but that data is limited.
- Let a person approve anything that reaches a customer, moves money or cannot be undone.
The four tests
Run each candidate job through these four tests. They are our own rule of thumb, and they draw on the sources below.
| Test | Good first job | Save for later |
|---|---|---|
| Frequency | Happens daily or weekly | A few times a year |
| Risk | A mistake is an annoyance | A mistake costs money or trust |
| Checkable | A person can see if it is right | Only an expert can tell |
| Reversible | Easy to undo | Sends, pays or deletes |
The last test is the one people skip. A wrong draft can be fixed before anyone outside sees it. An agent that sends the reply cannot take it back.
Check first whether it needs an agent
Anthropic, which sells AI models, argues for restraint. Its engineering guide says: "When building applications with LLMs, we recommend finding the simplest solution possible, and only increasing complexity when needed."
Plenty of good first projects are not agents at all. If you can write the steps on one page, a plain automation is the better start. Our guide on whether you need an agent helps you decide.
Where people already use AI at work
Real usage is a useful guide to where AI fits. The Census Bureau's August 2026 story reports the top uses among workers. These are shares of US workers, from a March 2026 survey.
| Task | Share of workers |
|---|---|
| Searching for information or technical help | 37% |
| Writing communications, documentation or instructions | 32% |
| Generating ideas | 32% |
| Interpreting, translating or summarizing information | 31% |
| Administrative tasks | 27% |
This covers AI in general, not agents. It still points to work that is text-heavy and easy for a person to review.
First jobs to run through the four tests
These are examples for you to test against your own work. We have not measured results for them.
- Summarizing long email threads before someone replies.
- Drafting replies to common questions, for a person to send.
- Sorting incoming requests and suggesting who should handle them.
- Pulling facts from documents into a spreadsheet, with a person checking a sample.
- Preparing a report from data your team already trusts.
- Finding the right policy in your files for staff to read and confirm.
Notice the pattern. In each case, the agent prepares and a person decides.
What to leave for later
These can be good uses eventually, but they are poor first projects.
- Sending messages to customers without review.
- Moving money, issuing refunds or changing prices.
- Making or influencing decisions about hiring, credit, housing or health.
- Deleting or overwriting records.
- Anything where you cannot tell if the result is right.
OWASP's AI Agent Security Cheat Sheet says: "Require explicit approval for high-impact or irreversible actions." It also says: "Implement action previews before execution."
How much autonomy to give it
Anthropic studied how agents are used across its own traffic, and published the results on 18 February 2026. It found that only 0.8% of actions appear to be irreversible, and it gives sending an email to a customer as an example.
It also found that 80% of tool calls come from agents that appear to have at least one safeguard. It says 73% appear to have a human in the loop in some way. Anthropic says the human involvement figures are upper bounds.
The study covers one company's traffic, and nearly 50% of the tool calls were software engineering. It may not match your business. Its conclusion is still useful: "This reinforces a point we have made previously: effective oversight doesn't require approving every action but being in a position to intervene when it matters."
For a first project, keep approval on. You can relax it after the agent earns a track record. Our guide on human-in-the-loop agents covers the options.
Score your candidates
Choose five jobs. Score each from 1 to 3 on the four tests, where 3 is the better answer.
| Job | Frequency | Risk (3 is low) | Checkable | Reversible | Total |
|---|---|---|---|---|---|
| Example: Summarize support threads | 3 | 3 | 3 | 3 | 12 |
| Example: Send refund emails | 2 | 1 | 2 | 1 | 6 |
| Your job | |||||
| Your job | |||||
| Your job |
The scores are a way to compare your own options, not a benchmark. Start with the highest total, and break ties by choosing the job your team would gladly give up.
Then ask the person who does it today. They know the exceptions that never made it into any process document.
Set up the first project to succeed
- Collect 20 to 50 real examples before you build anything. Anthropic's guide to evals says: "In reality, 20-50 simple tasks drawn from real failures is a great start."
- Write down what a correct result looks like.
- Give the agent read-only access at first.
- Keep a person approving the output.
- Log what it does and what it costs.
- Name one owner.
Our guides on agent evals and common causes of failure go deeper. If you want to compare products, see the best AI agents for small business.
Next step
Fill in the table with your team this week, and pick one job. If you want help turning it into a working agent, see our AI agent development service.
Frequently asked questions
What is the easiest thing to automate with AI?
Text work that a person reviews before it goes anywhere, such as summaries and draft replies. It is easy to check, and a person can catch mistakes before they reach a customer. Whether it saves real time depends on your volume.
Should customer service be my first AI agent?
Start on the staff side, such as suggested replies for a person to send. Moving straight to customer-facing answers raises the risk. Add direct answers after testing.
How long should a first AI agent pilot run?
Long enough to see real variety in the work, including unusual cases. We don't have a verified benchmark for this, so set your own time limit and decide in advance what result would count as success.
Should I automate my biggest cost first?
Probably not, because large costs can carry large risks. A smaller, safer job teaches your team how to test and supervise an agent. Move to larger jobs once that works.
What if none of my jobs pass the tests?
That is a useful result. It may mean your processes need tidying first, or that a plain automation fits better. An agent added to a messy process will not fix it.
