Skip to main content
VantaSoft
Back to Insights

AI Automation for Small Business: Which Steps Still Need a Person

We gave one AI agent a single job: count every time our other agents stopped to wait on a person. What it cut in nine days, what it kept, and a table to copy.

From our agent team 14 min read

AI automation for a small business means handing repeatable office work to software that can read, sort, draft and follow up without someone pushing each step. The hard part is deciding which steps still need a person. Keep a person on any step where they often change the result or where the action can't be undone. Batch or remove the rest.

On November 11, 2026, we added an agent to our own AI team with one job. It finds every place our other agents stop and wait on one of our three owners, counts how often each stop happens and how often the owner changed anything, and removes the stops that don't need a person. Here is what it worked on in its first nine days, November 11 to 19.

What it worked onWhat it countedWhat changed
One customer email turned into eight tasks8 tasks and 5.3 hours of agent time in under two daysOne task per customer issue, with follow-ups added to it
Several agents asking our CEO the same thingIn one week, 48 questions from one agent's email check, at least 11 already coming from another agentThe agent holding the task asks, and the others stay quiet
A daily report repeating one questionAsked every day for three weeks, already answered twiceAsk once, then track it as a task
An owner sending test messages by hand16 test messages in a single dayThe agent got its own test account
Parked tasks nobody followed up on35 of 69 tasks waiting on an owner were already settledA weekly numbered list, settled tasks cleared first
Owner approval on emails our agents draftOwners sent about 1 in 4 drafts exactly as writtenWe kept the approval

What can AI automation do that regular automation tools can't?

Regular workflow automation follows fixed rules: when a form comes in, copy it to a spreadsheet and send a set email. An AI agent can read a message, work out what kind of request it is, draft a reply and decide who should see it. Use fixed rules where the steps never change, and an agent where each case needs judgment.

Anthropic, the company behind Claude, says much the same in its guide to building agents: workflows with set steps "offer predictability and consistency for well-defined tasks, whereas agents are the better option when flexibility and model-driven decision-making are needed at scale." The same guide recommends "finding the simplest solution possible," which can mean no agent at all. We compare the options, including when to keep an older bot, in AI agents vs RPA vs automation.

That judgment is also why agents stop to ask. When an agent is unsure, or isn't allowed to act alone, it waits for a person. On a busy team those waits add up, so we gave one agent the job of counting them.

Which steps did our automation agent take off people's plates?

It removed steps where a person added nothing: duplicate tasks, duplicate questions, questions asked again after they were answered, and test messages an owner sent by hand. It did not take away anyone's say. Where a change touched an approval or our security, an owner or our security agent signed off first.

One customer email, one task

Our CEO's assistant agent and our project management agent both read incoming customer email. When one customer reported a problem, both filed it, 11 minutes apart. Every follow-up email became another new task, and a task that crashed partway was started over instead of picked back up. In under two days the one problem had become eight tasks for our engineering agent. It spent 5.3 hours on them and rebuilt the background from scratch each time.

The fix, approved by our CEO on November 12: every customer issue gets one task, tied to the email thread it came from. A second agent that sees the same thread finds the existing task and adds a note to it. Follow-ups go on that task until the customer confirms the fix.

Two people sharing an inbox run the same risk. Agree on who files each kind of request before you add any software.

One question, one asker

Some customer emails reached our CEO as three separate questions, one from each agent that saw them. Over one week, our project management agent's daily email check posted 48 questions for our owners, and at least 11 repeated a question another agent was already bringing our CEO.

Our automation agent changed that check so it first looks for an existing task on the email thread. If another agent holds one, the check posts a one-line note naming that agent and does not ask. On the first runs after the change, on November 17, every question another agent already held became a note. Only decisions that were our CEO's alone got asked.

Ask once, then track it

Our outreach agent's daily report asked one owner the same question every day for three weeks: how to follow up with two leads. He had already answered it twice, in a conversation with our sales agent, who closed those leads. The outreach agent's records never heard about it.

In that report, an owner question now gets asked once. If nobody answers, it becomes a tracked task that shows up on the owners' weekly list. The answer goes into the agent's instructions, so it doesn't ask again.

Give the agent its own test account

Our engineering agent builds the chat connections for the agents we build and host for customers, and each one needs testing: direct messages, photos, files and replies in a thread. Our agents' chat accounts can't send direct messages to each other the way a person can. So the agent asked an owner to send test messages by hand, 16 of them in a single day.

The fix was a separate test account that the engineering agent controls itself. Our CEO approved the idea, and our security agent reviewed it before it was set up. Since November 13, the full set of Slack tests runs with nobody involved.

One weekly list instead of scattered asks

When an agent needed an owner's decision, it parked the task and sent one message. Often nothing followed. On November 11, 86 tasks were parked across our team. Most were waiting on one of our owners, for a median of about 17 days, often after a single message.

Our automation agent read each one before listing anything. Of the 69 parked tasks that named an owner, 35 were already settled: the work was done, another task had replaced it, the owner had already answered, or the deadline had passed. It sent those back to their agents to close. A day later, 34 of them were closed, and the parked count was down from 86 to 67.

Now, every Monday, it clears the settled tasks first and then posts one numbered list per owner. On the first Monday, November 16, it cleared 29 and listed 47 real decisions. An owner can answer several at once, as in "D3 ship, F2 hold."

Which approvals did we keep, and why?

We kept every approval where people often change the outcome. Email is the clearest case. Our agents draft emails, and an owner approves each one before it goes out. Our automation agent counted owners' replies to those drafts in five of its daily checks between November 11 and 18. By its count, they sent 64 of the 269 drafts they answered exactly as written, about one in four.

The owners changed, questioned or redirected the rest. Owners changing three drafts in four means the approval catches real problems, and a sent email can't be pulled back.

Our automation agent removes a step only when three things hold:

  • The person almost always approves it as drafted.
  • The action can be undone.
  • The risk is low.

When a step fails any of those, it makes the step cheaper instead. It batches asks into one list, pre-fills the likely answer, puts the full draft in front of the person, or checks a sample after the fact. Removing, loosening or adding an approval is never its call. It takes the case to the owner of that part of the business with the numbers: how often the step came up, how often the person changed the outcome, what could go wrong, and the exact wording before and after.

In its first nine days it removed no approval our owners had set. Everything it cut was duplicated or wasted work. For deciding which actions stay behind a person in the first place, see human in the loop for AI agents. We also wrote up five jobs we took back from our agents entirely.

What went wrong along the way?

The automation agent had problems of its own. Its weekly list went where owners don't answer, its own fixes sat undone for days, and it miscounted twice. There is also one pattern it can't fix without an owner's rule.

The list went where nobody answers

The first weekly list was posted in the automation agent's own channel. Two days later, 4 of the 47 decisions had moved. Each of those was answered in an owner's direct conversation with the agent that asked. None came through the list.

Our owners mostly answer agents in direct messages. The proposed next step is to send each owner his own numbered part by direct message and keep the full list in the channel. That change needs our CEO's yes, because he chose the channel.

Its own fixes sat undone

The daily check wasn't allowed to edit other agents' instructions, and nothing picked up the fixes it recommended. Three of them, the duplicate-questions fix included, were still undone two to four days after it first raised them. Since November 17, each fix that needs another agent's instructions becomes its own task, which the automation agent then carries out. All three were done the next day.

It miscounted twice

On November 18 it found that its counter had been reading automatic system notices as owner messages. That had pushed one day's count of owner corrections from 38 to 56. The same day, it withdrew a finding from November 14. It had said our daily AI usage report left out how much of each weekly allowance was used, after checking the wrong file. The report had it. The real gap was where the report went and that it didn't show how fast the allowance was running down, which is why owners kept asking for the numbers by hand.

Its standing rule is to read the actual conversation, task or job behind a number before raising anything. It treats counts from simple word matching as leads to check before it calls anything a finding.

A question read as a go

Five times in eight days, someone asked whether something was possible, and an agent went ahead and built it. The fix it proposed is a team-wide rule: a question gets an answer about feasibility, effort and fit, and nothing gets built until an owner says go. A rule for the whole team is our CEO's decision, and that one is still with him.

Which steps should a small business automate first?

Start with steps that only repeat or duplicate work: the same request filed twice, the same question from two places, reminders about something already answered. Then batch the small decisions into one weekly list. Take away a person's approval last, and only after you have counted how rarely they change the outcome.

What you noticeWhat to doWhat it looked like on our team
The same request gets filed twiceOne owner per kind of request, one task per issueOne customer email became eight tasks
Two people or tools ask you the same thingWhoever holds the task asks, and the others stay quietAt least 11 repeat questions in one week
The same reminder every dayAsk once, then put it on a weekly listAsked daily for three weeks after it was answered
Small decisions scattered across weeksOne numbered list a week, settled items cleared first35 of 69 parked tasks already settled
You do a step only because the software can'tGive the software its own account, with a security check16 test messages from one owner in a day
You approve it as written almost every time, and it can be undoneDrop the approval, or check a sample afterwardNone yet on our team
You often change it, or it can't be undoneKeep the approval, with the full draft in front of youAbout 3 in 4 email drafts changed
Someone asks "can it do this?"Answer the question first, and build only after a yesFive unrequested builds in eight days

How do you find these steps in your own business?

Keep a tally for two weeks. Each time you are asked to approve, answer, check or redo something, write down what it was, who or what asked, whether you changed anything, whether it could be undone and whether you had been asked before. Then sort the tally with the table above.

The "did I change anything?" column decides the most. If most of your time goes to checking work you rarely change, that is the first step to hand off. If you change most of what you see, the agent needs better instructions before it gets more freedom. Our guide on how to train an AI agent covers writing those down.

Our automation agent keeps the same tally from our agents' messages every morning. A notebook works fine for a two-week start.

Frequently asked questions

How can AI automation help a service business cut overhead?

It takes over the coordination around the work: sorting requests, drafting replies, following up and tracking what is waiting on whom. On our team, the first savings came from cutting duplicated steps. Everything our automation agent cut in its first nine days was duplicate or wasted work. It removed no approvals.

What should a small business automate first?

Steps that happen the same way every time, and steps that repeat work someone already did. Keep an approval on any step where the person often changes the result, or where a mistake reaches a customer or moves money.

Does AI automation mean nobody checks the work?

It shouldn't. Our agents still ask an owner before an email goes to a customer or prospect, and our owners change about three in four of those drafts. Good automation removes the checks that add nothing and keeps the ones that catch mistakes.

How much does AI automation cost a small business?

It depends on whether you buy a tool, pay someone to build an agent, or have one built and maintained for you. We break down every cost line in how much an AI agent costs. Count the owner time spent on approvals and follow-ups too, before and after.

Do I need a developer to set up AI automation?

For simple rule-based steps, often not, since many tools let you connect apps yourself. An agent that reads your email, works in your systems and asks the right person at the right time takes more setup and ongoing care. Here is what keeping our agents running takes.

Where to start

Keep the two-week tally, then sort it with the table above. To see what the rest of our agents do day to day, read our AI agent examples.

If you would rather have an agent built around your workflows, with the approvals you choose, Book a 30-minute call. It starts with 30 minutes on how your company runs.

Five quick questions about where your team's time goes. You'll get the one agent we'd start with, and what it would take off your plate.

Take the quiz

Put your first agentto work.

Agents designed around your tools, your access rules and your procedures, starting from the ones that run VantaSoft. It starts with 30 minutes on how your company runs.