The frontline job used to be processing: work the queue, mostly routine, occasionally hard. Once governed AI agents take the routine work, what reaches a person is the exceptions the agent held, refused or flagged. The job becomes judgement all day. Job descriptions, hiring, training and measurement should change to match, and in most organisations they have not.
What does the queue look like after agents arrive?
Before agents, a frontline adviser's queue might be mostly simple items with a few difficult ones mixed in. The simple items set the pace, built confidence and gave new starters a gentle way in.
After agents, the simple items mostly stay with the agent at the depth it has earned. What arrives for the person is:
- Held items, where the agent has drafted an action and is waiting for a named person, often because it crosses the business's limit.
- Refusals, where the request is outside the agent's bounds and needs a human route.
- Escalations, where the agent detected something, such as distress or a possible fraud signal, that calls for judgement.
- Corrections, where a reviewer's fix should teach the agent.
Every item is there because judgement is needed. There is no easy work left to set the rhythm.
Why is judging exceptions a different skill?
Processing is about accuracy and speed on known patterns. Judging exceptions is about noticing what is different and deciding what to do about it. It needs a different set of habits:
- Reading why an item was held before deciding anything. The verdict names the bound that decided it.
- Checking the source of the key fact. Is it sealed by the system of record, signed by an identifiable party, declared by the customer, or external?
- Telling refusal from fault. A refusal is the agent protecting someone; working around it is the classic mistake.
- Writing a reason, because the decision is recorded with the person's name and will be read later.
- Correcting so the agent learns, not just fixing the one case.
None of these were central to the processing job. All of them are central now.
What should change in the job description?
Most job descriptions for frontline roles still describe processing: handle a volume of cases to a quality standard within service levels. That attracts and selects people for the old job. The guide on how job descriptions change when agents do part of the work explains the shift in detail.
A job description for the new frontline should say plainly that the role reviews agent-prepared work, decides exceptions, holds approval rights subject to certification, and teaches the agent through corrections. An AI agent reviewer job description template gives wording to start from. For the person who owns the agent rather than reviewing its work, the role is different again; see how to write an AI agent owner's job description.
How do new starters learn without easy cases?
This is the hardest problem in the new model. People used to learn by doing hundreds of routine cases before meeting a hard one. That ramp is gone.
On-the-job training still matters, but it has to be designed rather than assumed. New starters can work agent drafts at depth 2, where the agent prepares and they send, before they approve anything at higher depth. They can shadow certified reviewers. They can work through labs in a sandbox, meeting each type of verdict before they meet it live. And they can complete an observed practicum, where an assessor sees their decisions before they hold approval rights alone.
Should frontline staff and managers learn the same things?
They share a base: depth, verdicts, provenance, the record. Beyond that, the jobs differ. Frontline reviewers need deep practice on individual decisions. Managers need to read patterns across many decisions, staff the queue and coach. The question of whether frontline staff and managers should take the same training is best answered as shared foundations, different practice.
What is the trade-off?
A job made of exceptions is harder and more tiring. Every decision requires attention. Without routine items to break up the day, fatigue sets in, and tired reviewers start approving without reading. That is how rubber-stamping begins.
It also changes who suits the job. Some excellent processors find exception work stressful. Some people who struggled with repetition thrive on it. Teams will change shape, and some people will not want the new role.
And the business may need to recognise the job differently. If the role now demands more judgement and carries approval rights, its grading and pay may need to reflect that.
These are reasons to design the job carefully, not to avoid the change. Rotate reviewers across types of exception and give real breaks. Watch approval patterns for signs of fatigue. Be honest with staff about how the job is changing and support those who want to move. The change is happening whether the job description admits it or not.
What should you do differently?
Rewrite frontline job descriptions around judgement, review and correction. Hire for noticing and reasoning, not just speed. Build a designed ramp for new starters: drafts, shadowing, labs, observed practicum. Tie approval rights to certification. Measure decision quality and time per held item, not just volume. And plan for fatigue. The guide to AI agent skills by role sets out what each role needs to know.
Questions people ask
Do frontline teams shrink when agents take routine work?
Often the team's size changes and its work changes more. The business still needs enough certified reviewers to handle held items promptly, and that number depends on volume and the agent's depth.
How many exceptions can one reviewer handle well in a day?
There is no universal figure. Watch for signs of fatigue, such as approval times falling and changes to drafts disappearing, and set rotas so people are not reviewing hard cases without a break.
What happens to people who preferred the processing job?
Some will adapt with training and support. Others may move to roles that suit them better. Being clear early about how the job is changing helps everyone decide.