Most companies introduce AI by looking at the current organisation and asking:
Which person's tasks can we give to an agent?
That question starts too late.
The current role may be full of historical handoffs, duplicated checks, status updates, and information moving from one person to another. If you insert an agent without changing that design, you automate the clutter.
A better question is:
If we designed this work today, which steps would need human judgment, which could be executed by an agent, and who would own the outcome?
That is the difference between adding AI to the old org chart and redesigning work for a team of humans and digital coworkers.
The signal from business leaders
Fortune reported on a panel with leaders from Block, Synopsys, Cloudflare, and Amazon. Block CFO and COO Amrita Ahuja said the company had thought about fundamentally refactoring how it operates rather than putting AI on top of existing work.[1]
Her distinction was practical. People remain builders, goal-setters, and strategy deciders, while some roles that mainly pass information or perform an intermediate check may no longer need to exist in the same form.[1]
Cloudflare president Michelle Zatlyn said the company now hires “both employees and agents.” Fortune also reported Ahuja's expectation that agents will increasingly appear alongside humans in organisational charts and will need to be developed over time.[1]
These are observations from named leaders at a specific event, not proof that every company should copy their organisation or workforce decisions.
The useful takeaway is narrower: once agents can execute work, leaders need to redesign roles rather than simply redistribute tasks.
Old workflows create bad digital jobs
Consider a recurring customer enquiry workflow:
- Someone reads the enquiry.
- They copy details into a spreadsheet.
- Another person checks whether the lead already exists.
- A sales coordinator enters it into the CRM.
- A manager reviews the assignment.
- The coordinator emails the salesperson.
- The salesperson checks the original message again.
An agent can perform several of those steps. But if you preserve the entire sequence, you still have multiple copies of the same information, repeated checking, and unclear ownership.
You may even make the workflow worse. The agent works faster, so it feeds a poorly designed queue at higher speed. Staff then spend more time reviewing low-value updates and resolving contradictions between the spreadsheet, CRM, and email.
The problem is not agent capability. The work was never redesigned.
Map the work before assigning it
Choose one recurring workflow this week. Do not start with a company-wide transformation programme. Pick a queue that is frequent, visible, and annoying enough that people already know where it breaks.
Map every step into four categories.
1. Judgment
These steps require business context, trade-offs, empathy, accountability, or a decision the company is not willing to delegate.
Examples include approving a non-standard discount, deciding whether a customer complaint needs executive attention, or choosing whether to extend credit to a buyer with an unusual history.
Keep a human accountable for these decisions.
2. Information transfer
These steps move data from one place to another without adding meaningful judgment.
Examples include copying enquiry details into the CRM, attaching an invoice to a case, or notifying an owner that an exception is ready for review.
These are strong candidates for agent execution, provided the data source and destination are defined.
3. Checking
Some checks are deterministic: required fields are present, totals reconcile, the account exists, or the requested action is within a limit.
Automate those checks where possible.
Other checks are disguised judgment. “Does this explanation make sense?” or “Would this response damage the relationship?” may still need a person, especially for unusual or high-impact cases.
Separate the two. Do not keep every check human just because the old process did, and do not automate every check because a model can produce an answer.
4. Execution
These steps change the world outside the agent: updating a CRM record, creating a supplier, issuing a document, scheduling work, sending a message, or preparing a payment.
Execution needs a named owner and a defined authority boundary. The agent may perform the action, but a person still owns the business result.
Remove handoffs before you automate
After classifying the steps, look for three things to delete.
First, remove duplicate data entry. Decide which system is authoritative and update it once.
Second, remove checks that repeat a rule another system already enforces. If the CRM can prevent an invalid status change, do not pay a person or agent to inspect it manually every time.
Third, remove status messages that exist only because the next person cannot see the queue. Give the owner a shared view instead of generating another email.
Only then should you decide what work belongs to the agent.
This is where domain experts matter. The operations lead knows which exception is harmless. The finance manager knows which variance is material. The sales leader knows when a discount changes the economics of a deal.
Their job is not merely to “give feedback” after deployment. They define the work.
Write a five-field work package
Once the workflow is cleaner, give the agent a proper assignment.
Outcome
State the result, not the activity.
Bad: “Process incoming enquiries.”
Better: “Create or update a complete CRM lead record and route it to the correct owner within ten minutes.”
Inputs
Name the approved sources: inbox, web form, customer database, product table, or policy document.
Do not make the agent search everywhere by default.
Access boundary
Specify what the agent may read, create, change, send, or spend. A lead-routing agent may update ownership and status but must not change pricing or send a contract.
Acceptance test
Define what “done” means in the system of record. Required fields are present. The account is not duplicated. The owner matches the territory rule. The update can be read back from the CRM.
Escalation path
List the conditions that stop normal execution: missing consent, conflicting customer records, a high-value deal, an unsupported request, or repeated tool failure.
Name the person or queue that receives the exception.
The Agent Boss owns the quality of the assignment
A digital coworker still needs management.
Someone must own the clarity of its work package, the limits of its authority, the quality of its output, and the way it improves. That person is the Agent Boss for the workflow.
This is not necessarily an IT role. It may be the finance manager who understands close exceptions, the operations lead who sees every failed handoff, or the customer-service supervisor who knows when policy and judgment diverge.
The Agent Boss does not watch every action. They design the job, review exceptions, measure outcomes, and update the assignment when the business changes.
Humans still own the goal, strategy, judgment, and final accountability. Agents can move information and execute defined work. The advantage comes from deciding that boundary deliberately.
Do not add an agent to the old org chart.
Redraw the work. Remove the handoffs that no longer earn their place. Then give the agent a job worth doing.
Sources
[1] Block, Amazon, Synopsys and Cloudflare are rethinking how work gets done with AI