AI agents in the workplace do the preparation and the follow-through, and people make the calls. So the work keeps landing back on a person's desk, as a draft to approve or a login to grant. On our own task board, 37 of 101 such handoffs got an answer within a day, and 50 are still waiting.
We run VantaSoft with a team of 13 AI agents that work alongside the company's three owners. Since July 8, 2026, the agents have stopped and handed a task to a person 101 times on our shared task board, because they needed an approval, a decision, access or information they could not get on their own. This is where those handoffs stood on October 27.
| What happened to the handoff | Handoffs |
|---|---|
| Answered within an hour | 22 |
| Answered in one to 24 hours | 15 |
| Answered after more than a day | 14 |
| Still waiting on October 27 | 50 |
What do AI agents in the workplace hand to people?
Mostly requests for a yes. Of the 50 handoffs still waiting, 19 ask for approval before finished work goes out: a tested change to a website or product, or a draft email, post or quote request. Another 11 need a decision. Eleven more need someone to sign in, grant access, or take a step only they can take, like posting from their own account.
| What the agent needs from a person | Waiting now |
|---|---|
| A yes before a finished change goes live | 12 |
| A yes before a draft goes out | 7 |
| A decision, or a yes or no on one action | 11 |
| A login, an invite, or a step only they can take | 11 |
| Information only that person has | 4 |
| A second look at technical work | 4 |
| A signature from someone outside the company | 1 |
The waiting piles up at the top. Of the 50, 42 name one of our three owners as the person who has to act. Once agents take over the preparation, the owner's attention becomes the thing everything else queues for. If you want to see what our agents do in the first place, our AI agent examples post walks through all 13 and their jobs.
How long does work wait when an AI agent hands it to a person?
Usually a few hours, or else a long time. The 51 handoffs that got an answer took a median of about three hours, and 22 were answered within the hour. After the first day the odds drop. Of the 64 handoffs that went a full day without an answer, 14 have been answered since and 50 are still waiting. Some of those are only a few days old, but 26 have waited more than a week and seven more than a month.
We are not alone in this. In a report on how work is changing inside the company, published October 16, 2026, Anthropic found that its employees' interactive sessions with Claude were "around 80% idly waiting for human input." It concluded that faster agents make "human attention the bottleneck for agent output." Its measure differs from ours, idle time inside a working session rather than days on a task board, but it points the same way. An agent that finishes its part in minutes can still take a week to deliver, because the last step belonged to a person.
Why does a handoff to a person stall?
Each stall below comes from our own records, along with the rule we added after it.
It was addressed to nobody in particular
The three longest waits on our board, at 95, 95 and 77 days, were waiting on a review or on access, and none of them named who should act. The 42 waiting handoffs that do name an owner have waited a median of under a week so far, and none of them has reached 77 days. When a request goes to "an owner" or "the team," each person can fairly assume someone else has it.
It went to the wrong person
In September, a decision about the tool we use to find government contracts to bid on went to our CEO. It had a deadline, and it belonged to another owner, the one who runs that part of the business. The rule we set afterwards is that a decision waits on the owner of the department it belongs to, not on whoever sits at the top.
It arrived where the person does not look
Also in September, an internal heads-up went out as an email. Since then, when an agent needs to tell one of us something, it sends a direct message in Slack from its own account, and email is kept for people outside the company.
The next step belonged to a customer
In early September, five setup tasks for a customer's agents sat marked "in progress" while we waited for the customer to open their install links. The agent on that job checked every 30 minutes and reported each time that nothing had changed. Now a task that waits on a customer is parked on the person who owns that customer relationship, with the customer's step written down as the thing it is waiting for. A small automatic check watches for that step and hands the work back within minutes of it happening, so the only thing left for the person to do is nudge the customer.
One wait held up the work behind it
On October 13, a finished article for this blog sat on our test site for about six hours. It was ready, but two unrelated website changes ahead of it were waiting for an owner's go-ahead, and everything in that line went live together or not at all. The rule now is that finished work goes live on its own. A change that is waiting stays parked and blocks nothing else.
Which rules fixed our stalled handoffs?
| What stalled | The rule now | Since |
|---|---|---|
| A decision sent to the CEO by default | It waits on the owner of that department | September 2026 |
| A heads-up sent by email | Internal notices go by direct message in Slack | September 16, 2026 |
| A task waiting on a customer's click | It waits on whoever owns the customer, and a check hands it back | September 5, 2026 |
| Finished work stuck behind a waiting change | Finished work goes live on its own | October 13, 2026 |
| An agent missing a login | It asks the one owner who grants access and says what that unblocks | October 15, 2026 |
| Routine calls waiting on the CEO | A standing yes, with spending and customers still held back | October 17, 2026 |
The last two take the handoff out altogether. The access rule tells an agent that lacks a login to ask the one owner who grants access, name the exact login, and say what it unblocks, instead of waiting quietly or working around it. And in mid October our CEO told the agent that writes this blog to stop asking before it acts on its own content and search suggestions. Paid spending, anything that asks a customer for something, and changes to how we position the company still wait for an owner.
Not every rule covers every place work can wait. Our general task board still lets an agent ask for "a review" without naming anyone, and the oldest waits on it show what that costs.
How do you set up handoffs for AI agents in your own workplace?
Ask your agent to write every handoff as a short card. If it cannot fill in a line, the handoff is not ready to send. Here is the card filled in for an example 15-person business.
| Line | What it says | Example |
|---|---|---|
| Who | One named person, never "the team" | Maria, operations lead |
| The one thing | One verb: approve, choose, sign in, answer | Approve this reply to a refund request |
| By when | A date, and why that date | Friday, before the customer's renewal |
| If nobody answers | What the agent does next | Asks again Thursday, then goes to Dan |
| Where | The place that person already reads | A direct message in Teams |
| How it comes back | How the agent learns it is done | Maria replies, or the agent sees the reply sent |
Then give one person a short, daily look at everything still waiting, at the same time each day. On our board, a handoff that went a day without an answer was more likely to still be waiting than to get one, so a daily look catches requests while an answer is still likely. Anything that comes up week after week is a candidate for a standing yes. Which actions need a person's approval in the first place is its own question, and our guide to human in the loop for AI agents covers it action by action.
What changes for the people who work alongside AI agents?
More of their day goes to answering the agents. Anthropic saw the same shift. Written standing instructions for its agents grew "from a few hundred in July 2025 to 8,500 a year later," and one engineer "set up an agent chief of staff to triage agent requests." In a small business that load falls on the owner and one or two leads, because they are the people who can say yes.
So plan for it before you add an agent. Decide who answers each kind of request, how fast you expect an answer, and what the agent does while it waits. Count the time spent answering the agent as part of what it costs, the same way you would count the time it takes to manage a new hire. Some jobs need no handoff at all, and our six-point test for unattended work helps you find them. For the access side, our guide to what access an AI agent should have shows how to keep logins narrow so that fewer requests have to reach you.
Frequently asked questions
How are companies using AI agents in the workplace?
At VantaSoft, 13 agents handle content and search, outreach, finance, engineering, testing, project delivery, sales research, proposals and scheduling for the company's three owners. The agents prepare the work and follow it through. People decide, approve and sign. Since July, that split has produced 101 handoffs from an agent to a person on our task board alone.
Do AI agents replace the people they work with?
In our experience they change what those people spend the day on. The work that reaches a person on our board is approvals, decisions and access. Anthropic's report describes an executive assistant who built an agent for scheduling and email triage and moved on to higher-order "chief of staff" work.
Who should an AI agent report to?
One named person for each job, plus a named backup for when that person is out. In our records, the requests that named nobody in particular are the ones that waited longest.
Why do requests from AI agents go unanswered?
In our records, the usual cause is the request itself. It named nobody in particular, went to the wrong person, or arrived somewhere that person does not check. Requests that depend on a customer stall for a different reason, and they need the person who owns that customer to nudge them while a check watches for the customer's step.
How do agents and people pass work back and forth?
Through a shared task list, where each task has one owner at a time and a note saying what it is waiting for. Handoffs between agents are their own topic, and our post on multi-agent AI covers how our agents pass work to each other.
Where to start
When we build agents for a business, discovery covers every point where an agent will need a person: who approves what, where each request goes, and what happens if nobody answers. If your team is already the bottleneck for the tools it has, that map is a good first step. It starts with 30 minutes on your workflows. Start automating.
