An AI agent can finish a recurring job without you if the job has three properties: you can tell at a glance whether the output is wrong, a wrong result is cheap to undo, and one named person is waiting for it on a known day. Jobs with all three run fine on a clock. Jobs missing any of them should stop one step short and hand you something to release.
Most answers to this question are a list of things agents can do. The list was never the hard part. We run 94 recurring jobs across 12 agents inside VantaSoft, and the ones that work are not the clever ones. They are the ones built so a wrong answer shows up fast and costs almost nothing to fix.
What can AI agents do without a person watching?
An agent can read, gather, compare, sort, summarize and draft without supervision. It can open the tools you already use, pull what it needs, and leave you a finished piece of work. What it should not do alone is anything that leaves the building: send, pay, post, quote, sign or grant access. That line is about consequence, not about how capable the agent is.
Think by job rather than by skill. A weekly cash position summary and a weekly price change to a customer both pull numbers out of your accounting system. One is safe to run at 6 a.m. The other is not, and better accuracy does not move it, because the cost of being wrong in the two cases is not comparable.
If you want the catalog version first, 13 real agents and the jobs they do lists what ours actually handle, and 12 roles and 8 industries sorts the same ground by function. This article is about which of those jobs can run while nobody is looking.
Which of your recurring jobs can run on their own?
Score the job out of 12. Six questions, 0 to 2 points each, ordered by how often each one turns out to be the real problem.
| Question | 2 points | 1 point | 0 points |
|---|---|---|---|
| Can you tell it is wrong in under a minute? | One look at the output shows it | You would open one other tool to check | You find out later from a customer or a bank balance |
| Is a wrong result cheap to undo? | Delete it and run it again | Reversible with some cleanup | It already left the building |
| Is one named person waiting for it? | A person expects it on a known day | A shared channel sees it, nobody owns it | Nobody would notice for weeks |
| Do the inputs live in a tool it can open? | Everything it needs is in a system | Mostly, with occasional context from you | It needs to know how a relationship is going |
| Can you write the rule on one page? | Yes, including the odd cases | Yes, for the normal case | You would say "use your judgment" |
| Does it spend money or speak for you? | No | It drafts, a person releases | It would pay, sign, quote or send |
A score of 10 to 12 means run it unattended. Give it a fixed time, a fixed output and one named reader, then check a sample monthly instead of every run.
A score of 6 to 9 means run it on a schedule but stop it one step short. The agent does the work on the clock and leaves it for a person to release. Most good first jobs land here.
A score of 0 to 5 means keep it with a person for now. Have the agent gather and draft when asked rather than on a clock, and score the job again once the rule is written down.
Those cutoffs are a planning heuristic we use to start the argument, not a measured benchmark. The scoring is useful mostly because it forces the argument into the open. We have watched a job score well on five questions and zero on the second one, and that single zero settled it.
Question three is the one owners skip and then regret. A job nobody is waiting for can fail quietly for a month. That happened to us. One connection changed, six jobs went silent, and we found out late because nothing looked broken from the outside. The five failures we have had mostly share that shape.
What do our agents finish before the workday starts?
Of our 94 recurring jobs, 39 fire for the first time before 9 a.m. Pacific, and the earliest runs at 1 a.m. By the time someone opens a laptop, the overnight work is sitting in a channel waiting to be read.
The work splits into three kinds, and the kind matters more than the hour. There is gathering, which pulls the day's new information into one place: prospect research before a call, a calendar picture built from several calendars, new listings worth a look. There is watching, which checks something that only matters when it changes: invoices crossing 30 days, project dates slipping, a filing window about to close, disk space on a machine. And there is preparing, which builds something a person will finish: drafted follow-ups, a report ready to review, a task list for the day.
None of the three sends anything. The gathering and watching jobs produce a message somebody reads. The preparing jobs produce a draft somebody releases. For one of these written out end to end, including the rules we had to add after it got things wrong, see what our scheduling assistant does every morning.
For the week of October 13 to 19, the run record shows 3,001 runs for the nine of our agents whose records reach back a full week. Of those, 2,986 finished, 3 failed outright and 12 ended with no result recorded. The other three agents' records do not go back that far, because the record keeps only the last 1,000 runs.
Why is a finished run not the same as a correct one?
A finished run means the work ran to the end. It says nothing about whether the output was right, and confusing the two is how an unattended job goes bad without anyone noticing.
One of our agents once reported success on work that was gone. The run finished. The record said completed. The output had been lost, and we caught it only because a person was expecting a specific thing and did not get it. The named reader in question three is doing real work there.
That is also why a high finish rate is nothing to sell you on. Ours is high because the jobs are small, the rules are written down and the risky last step is cut off. Point the same agent at a job with judgment in the middle and the finish rate stays high while the usefulness collapses.
A job with twenty steps that each go right 95 percent of the time finishes correctly about 36 percent of the time. Keeping steps per job low is how an unattended job stays honest, which is why we break long jobs into short ones with a readable output at each stop.
What should an AI agent never do on its own?
Anything hard to reverse that reaches people, money, access or legal rights. Inside our own fleet that means every email is drafted and waits for a person, no agent moves money, and no agent gets permissions beyond the one job it has. We treat those as rules rather than settings, and they do not loosen because an agent has behaved well for a month.
The published security guidance says the same thing. The OWASP GenAI Security Project lists Excessive Agency as LLM06 in its 2025 top ten for AI applications and names three root causes: "excessive functionality; excessive permissions; excessive autonomy." Its recommended fix for the third is plain enough to copy. Use "human-in-the-loop control to require a human to approve high-impact actions before they are taken."
Two practical consequences for a small business. Give the agent the narrowest access that does the job, which is what the four tier permission ladder is for. And decide the approval point per action rather than per agent, which is what the four oversight modes are for. One agent can run fully unattended on one action and be blocked on the next, inside the same job.
There is a related difference between a chat tool and an agent. A chat assistant works while you are in the room and stops when you close the tab. An agent holds a standing job and runs on a clock. We wrote about where an assistant stops being enough separately, because wanting something to happen without you asking is the point where the shape of the tool has to change.
What does this look like for a small business?
Take a plain example: a cash position summary that lands with the owner every Monday morning. The output is checkable at a glance by someone who knows her own numbers. A wrong one is cheap to redo. There is a named reader and a known day, so a missing report gets noticed the same morning. The inputs live in an accounting system. The rule fits on a page. Nothing is spent or sent. That job scores a 12.
What changes for an owner is less about the report than about not having to chase it. Michelle, in a public review of our Agent Service, put it this way:
Like many founders, I constantly have important tasks sitting on my list that I know need to get done, but there are simply never enough hours in the day. Building AI agents with VantaSoft has given me something incredibly valuable, peace of mind knowing those tasks are actually being executed without me having to constantly follow up or manage every detail.
That is a 5/5 Google review, quoted verbatim.
Running work on a clock is part of what we build: daily summaries, weekly reports and recurring checks that go out whether or not anyone asks. The build starts at $1,000 one time and hosting and maintenance start at $200 a month, and the four ways to get an AI agent compares that against doing it yourself.
Frequently asked questions
What can AI agents do that a chatbot cannot?
Run without being asked, reach your other systems, and remember what happened last time. A chatbot answers a question in a window. An agent holds a standing job, opens the tools the work lives in, and produces the same output on the same day every week.
Can an AI agent work overnight?
Yes. Of our 94 recurring jobs, 39 start before 9 a.m. Pacific and the earliest runs at 1 a.m. The hour is not the constraint. What matters is whether the output can be checked quickly and undone cheaply when it is wrong.
How do I know the agent actually did the work?
Have one named person expect a specific output on a known day. That beats most monitoring you can buy, because a missing or wrong report gets noticed by someone who knows what right looks like. A finished run is not proof of a correct result.
What is the best first job to hand an agent?
A recurring report or check that scores 10 or higher on the test above. Usually it is something a person already does every week, resents doing, and could glance at and immediately spot as wrong.
Do unattended agents replace staff?
Not in the jobs described here. The work that runs on its own is the gathering, watching and preparing that sits in front of a decision. The decision still has an owner, and in our own fleet that owner is a person on every action that costs money or reaches a customer.
How many systems should one unattended job touch?
As few as will do the work. Every extra connection is one more thing that can change quietly and take the job down with it, and every extra permission widens what a wrong answer can reach.
Where to start
Pick the three recurring jobs that annoy you most and score each one out of 12. Anything at 10 or above can go on a clock as it is. The ones in the middle are usually better first candidates than they look, because cutting off the last step turns a risky job into a safe one without giving up the hours it saves.
If you would rather work through that list with someone, start with 30 minutes on your workflows. We will tell you which of those jobs we would put on a schedule, which we would stop one step short, and which should stay with a person.
