The first phase of AI at work was about helping people do tasks faster.
The next phase is different.
When digital coworkers can research, draft, check, update and coordinate in parallel, the manager’s job shifts from doing more work to directing attention across more work.
That sounds like leverage. It can also create a new bottleneck.
If every agent asks the owner what to do next, waits for missing context, or escalates every uncertainty, the business has not removed the queue. It has moved the queue to the busiest person.
The emerging role of the Agent Boss is not to watch every agent. It is to design where human attention should go.
What Anthropic found inside its own workplace
Anthropic recently published research on how work is changing inside Anthropic itself. These are measurements from one unusually AI-intensive company, not a forecast for every business.
The findings are still useful as a signal.
Anthropic reports that, in interactive multi-turn sessions, agent sessions spent roughly 80% of their wall-clock time waiting for human input. It also found that meta-work—building skills, writing plans, setting up reviews, monitoring pipelines and otherwise managing agents—rose from 10% of sampled employee sessions in autumn 2025 to 22% in spring 2026.[1]
The report describes employees moving away from parts of direct implementation and toward supervising agents, setting direction and handling strategic work. It also says that decisions made in meetings or private channels can leave agents without the context needed to update their goals and assumptions.[1]
Again, this does not prove that every SME will see the same ratios or the same outcomes.
But it exposes a practical design problem: as execution capacity increases, attention, context and decision rights become scarce resources.
More agents can create a bigger management inbox
Imagine an SME uses agents across sales, finance and operations.
One agent prepares a weekly pipeline report. Another checks overdue invoices. A third classifies incoming enquiries. A fourth drafts supplier follow-ups.
Each agent is productive in isolation. Yet the owner receives a stream of questions:
- Which opportunities matter most this week?
- Is this overdue account sensitive?
- Which customer record is authoritative?
- Can this discount be offered?
- Who should review the exception?
- Is this external message safe to send?
None of these questions is unreasonable. Together, they can consume the time the system was supposed to release.
The wrong response is to tell agents to ask less. The better response is to redesign how attention is routed.
That design has four parts: priorities, context, exceptions and consequential judgment.
1. Priorities: decide what deserves attention first
Agents can produce more work than a small leadership team can absorb. A useful system therefore needs an explicit priority model.
Do not give every request the same urgency. Define what makes work important:
- customer impact;
- financial value;
- deadline proximity;
- operational dependency;
- reversibility;
- confidence in the available evidence.
For example, a missing field in an internal draft can wait. A payment discrepancy that blocks month-end close cannot. A routine customer acknowledgement may follow an approved pattern. A complaint involving a contractual commitment needs a named owner.
The Agent Boss sets those distinctions before the queue becomes noisy.
2. Context: make the right information visible at the right moment
Many agent interruptions are not judgment calls. They are requests for information the system should already have supplied.
An agent preparing a sales brief may need the latest CRM notes, approved pricing rules and account-owner information. An accounts agent may need the purchase order, delivery record, invoice and delegated approval threshold.
If those inputs are scattered across inboxes, private messages and people’s memories, the agent will stop, guess or ask.
The fix is not a longer prompt. It is a clearer context package:
- the authoritative records;
- the current policy or business rule;
- the owner of the decision;
- the freshness date;
- the evidence required for completion.
Good context reduces avoidable questions. It also makes the remaining questions more useful.
3. Exceptions: separate routine variation from real ambiguity
Not every difference is an exception.
A late invoice by one day may be routine. A changed bank account combined with an urgent payment request is not. A lead from an existing account may need reassignment. A request involving an unfamiliar contract term may need specialist review.
The Agent Boss should define the patterns that can be handled automatically and the conditions that must stop or reroute the work.
For each workflow, write down:
- What can proceed when evidence is complete?
- What can proceed within a defined tolerance?
- What must be reviewed by a domain expert?
- What must stop until a named person decides?
This keeps people out of predictable low-risk work without pretending that every edge case can be automated.
4. Consequential judgment: keep accountability named
Some decisions deserve human attention precisely because they carry consequence.
Examples include releasing money, changing permissions, deleting records, sending sensitive customer communications, making commitments or acting on incomplete evidence.
Human involvement here is not an efficiency failure. It is part of the job design.
The important questions are:
- Who has authority to decide?
- What evidence must they see?
- What choices can they make?
- What happens after approval, revision, rejection or escalation?
- Who remains accountable for the outcome?
An approval request that only says “Approve?” pushes investigation back onto the reviewer. A useful decision packet shows the proposed action, supporting evidence, uncertainty, consequence and next step.
The goal is not to remove people from work. It is to reserve their attention for the work that genuinely needs judgment.
Run an attention-bottleneck audit this week
Choose one recurring workflow with at least one digital coworker or AI-assisted step.
Review the last ten runs and record every point where the system waited for a person.
For each wait, ask:
- Was a priority unclear? Define the rule that determines urgency or sequence.
- Was context missing? Identify the record, policy or owner the agent should have received.
- Was this a true exception? Separate unusual but routine variation from material ambiguity.
- Did it require consequential judgment? Name the decision-maker and the evidence packet.
- Was authority unclear? State who can approve, revise, reject or escalate.
Then remove one avoidable interruption before adding another agent.
That is the real shift from operating to orchestrating.
An Agent Boss does not become the human API for a fleet of digital coworkers. The Agent Boss designs the flow of attention so routine work can move, exceptions reach the right expert and consequential decisions remain owned by people.