# The steps that stay human

> Walk an eight-step hiring process twice, once for what an agent could do and once for what stays with a person, and learn the three reasons a step stays human and why the list is decided early.

Page: https://schubergphilis.github.io/ai-training/using-agents/steps-that-stay-human/
License: CC BY-SA 4.0, https://creativecommons.org/licenses/by-sa/4.0/

In this lesson we read the eight steps of one everyday process, hiring a
contractor, twice. The first reading asks which steps an agent could do,
and the answer is most of them. The second reading asks which steps stay with
a person anyway, and gives the reason per step. The second list is much
shorter than the first, and the gap between them is the subject of the
lesson. You don't run anything here. The process is supplied, and the work is reading
it and deciding.

The safety course showed where to put the approval before an irreversible
step. This lesson is about a wider set of steps: the ones a person keeps
even when nothing irreversible happens in them, because of who has to
answer for the step, where its information is, or what doing it does for
your own judgment.

## What an agent could do

The process below is the one the operations lead at Brindlewick
Instruments, the fictional company from the earlier lessons, follows when
the team needs an outside specialist for a few months. Read it once as a
list of steps, without deciding anything yet.

```text
Hiring a contractor, in eight steps

1. Write the job description from the project notes.
2. Post it on two job boards and collect the applications.
3. Screen the applications against the requirements into a shortlist.
4. Interview the shortlisted candidates.
5. Check the references of the two finalists.
6. Choose the contractor.
7. Agree on the rate and sign the contract.
8. Set up accounts, access and a first-week plan.
```

Now ask, step by step, whether an agent with a browser, your email and
your calendar could do it. Step 1 is a drafting task from notes it can
read. Step 2 is form filling and collecting mail. Step 3 is reading
applications against a written list of requirements and sorting them.
Step 4 is scheduling and, with a meeting tool, even taking notes. Step 5 is
sending the reference emails and summarizing the answers. Step 6 is a
comparison of the notes and the summaries. Step 7 is filling in a contract
template. Step 8 is creating accounts and writing a plan from the job
description. The agent could take part in every step and could carry out
seven of the eight on its own. The only step where a person has to be
present is the interview, and even there the agent could book the room
and write the notes.

A person who stops at this point concludes that the whole process can be
handed over with an approval before the contract is signed. That
conclusion mistakes "could do" for "should do", and the next section is
about the difference.

## Why a person keeps a step anyway

Set the process aside for a moment, because the reasons are easier to see
in smaller examples. A step stays with a person because someone has to
answer for it, because its information is in no system the agent can
reach, or because doing it keeps the person able to judge the agent's
work.

**Someone has to answer for it.** A manager's performance review has a
rating in it, and an agent can draft a fair rating from the year's notes.
The rating stays with the manager. The person who has to explain it to
the employee later is the one who must have decided it. When a colleague asks "why a three?", the answer
"the tool said so" ends the manager's part in the process. The process
was there to have a manager in it. An agent can prepare a step it
can't be accountable for, and accountability is the reason the step
stays with the person who holds it
(Building effective human-agent teams, Claude Academy).

**The information is in no system the agent can reach.** Picking a time
for a difficult meeting looks like a calendar task. The agent reads every
calendar and finds the free slot on Thursday afternoon. You know that the
colleague you are meeting has had a hard week and that Thursday afternoon
is the last thing they need, and you know it because they told you in the
corridor. Nothing the agent can read holds that fact. A step whose input
is in your memory, in a conversation, or in the look on someone's face
stays with you because the agent would do it without the input that
matters.

**Doing it is what keeps you able to judge the agent's work.** Reading the
first few applications yourself is slower than reading a shortlist. It is
also how you learn what this year's applications look like and
whether the shortlist the agent made left the right people out. A person
who has read none of the applications for two years can approve any
shortlist and reject none. They no longer know what a good one contains. A person keeps some steps as the price of being
qualified to check the rest, in the same way a pilot flies some landings by hand to stay
able to take over from the autopilot.

A step can have more than one of these reasons, and none of them is about
what the agent can do. The agent can draft the rating, find the slot, and
read the applications. The person keeps the step for a reason about the
person, and naming that reason is what tells you the step is a choice
rather than a limitation that the next version of the tool removes.

#### Checkpoint: Why does each step stay with a person?

Match each step to the reason it stays with a person.

Statements:

- Deciding the rating in a performance review
- Picking the time for a difficult meeting with a colleague who told you in person about their week
- Reading a sample of the applications before looking at the shortlist

Options:

- Someone has to answer for it
- The information is in no system the agent can reach
- Doing it keeps you able to judge the agent's work

#### Checkpoint: Does a better agent change the list?

A team's rule is that the agent drafts each performance review and the
manager decides the rating. A year later the agent's drafts are much
better. What happens to the rule?

- The rating moves to the agent once its drafts match the managers' ratings often enough.
- The rating stays with the manager, and the better agent changes the quality of the draft the manager starts from.
- The rating moves to the agent, and the manager keeps an approval before the review is sent.

## Naming the step is a design choice

The list of steps a person keeps is a decision you make, and the time to
make it is when the process is set up (AI Fluency for builders, Claude Academy).
The reason is how automation spreads through a process that nobody has
mapped this way. Each step the agent can do gets handed over when
someone is busy and the tool looks ready. Nobody decides to automate the
rating in the review or the choice of contractor. One month the draft has
a rating in it, the manager is short of time, and the rating goes out as
drafted. The step nobody named as human is automated by default, one busy
week at a time, and it happens sooner as tools improve.

A named step behaves differently. "Step 6, choosing the contractor, is
done by the hiring manager" is a line in the process description, and
when the agent's summary comes with a recommendation, the line is there
to say that the recommendation is input and the choice is still a
person's. A team that writes an AI policy names these steps in it as the
approvals a task needs, which is the same list seen from the
organization's side. The rule about which step a person keeps and who
oversees it is written down in both places, and for some steps the law
writes it down as well. Under the EU AI Act, an AI system that screens
or ranks job applicants is a high-risk use (Annex III, point 4(a)). From
December 2, 2027, the organization that deploys it must give human
oversight to named people who have the competence and the authority to
act (Article 26(2)) (Regulation (EU) 2024/1689 (Artificial Intelligence Act), consolidated text of 27 July 2026, EUR-Lex).
So step 3 of the hiring process, the screening, has a person in it in
any case. That person reads a sample, owns the shortlist, and is named in
the process description.

Naming the human steps has a second effect. A process with named human
steps keeps working when the tool is down, because the person who owns a
step knows how to do it, has done it recently, and can do the neighboring
steps by hand for a day. A process where every step was handed over as
soon as it could be has nobody who remembers how step 3 was done. The
outage of a tool becomes the outage of the process. The human path is the
one you keep on purpose, the same way an office keeps a paper form for
the day the system is unavailable.

#### Checkpoint: The summary ends with a recommendation

The agent's summary of the two finalists arrives, and its last line is
"Recommendation: candidate B". The process description says nothing about
who chooses. What do you do?

- Accept the recommendation and move on to the contract, since the agent has read every application, every interview note and every reference, and you have read its summary.
- Read the two interview notes and the reference summaries yourself, make the choice, and add a line to the process that names the hiring manager as the one who chooses.
- Ask the agent to explain its recommendation in more detail, with the interview notes and references it weighed, and accept it once the explanation reads well.
- Reject the recommendation and redo the screening of every application yourself, since the agent has gone one step further than you intended it to.

#### Pitfall: Being careful as the boundary

A team agrees that the agent may do the whole hiring process "as long as
we are careful", and everyone means it. Six months later the agent is
choosing contractors, because careful has no step number, so no step ever
came up for a decision. "Be careful" is a mood, and a process can't hold
a mood. The boundary that holds is a named step with a named owner: "Step
6 is done by the hiring manager." Write the step, and the care has a
place to happen.

## Exercise

Go back to the eight-step process at the top of the lesson. For each
step, write down whether it stays with a person or goes to the agent,
and for each step that stays, name the reason from the lesson that
holds it there. Ten minutes is enough, and the
answer is a list of eight lines.

A good answer keeps step 6, the choice, because the hiring manager has to
answer for it. The decision in step 7, the rate and the signature, stays
for the same reason, and the agent may fill in the contract template.
Step 4, the interviews, stays because what the candidate is like in the
room is in no system the agent can reach, even if the agent books the
rooms and writes the notes. Part of step 3 stays too: the manager reads
a sample of the applications and decides the shortlist. That keeps the
manager able to judge the agent's screening, and screening applicants is
a high-risk use under the EU AI Act that needs named human oversight. Steps 1, 2, 5 and 8 go to the agent,
with the manager reading the results. Where your answer differs, write
down the reason you weighed differently, because a different reason is a
fine answer and a missing one is the thing to fix. Which step in your
list would be automated by default if nobody wrote it down?

Stretch: Take a process from your own work with five or more steps, write the steps out, and mark the ones that stay with a person, with the reason per step. Then check whether your team's written process says so.

## Recap

1. An agent could take part in almost every step of an everyday process, and "could do" is a different question from "should do".
2. A step stays with a person when someone has to answer for it (Building effective human-agent teams, Claude Academy). It also stays when the information it needs is in no system the agent can reach. And it stays when doing it keeps the person able to judge the agent's work.
3. The reasons are about the person and the process. A better agent changes the draft the person starts from and leaves the list of human steps as it was.
4. A step nobody named is automated by default, one busy week at a time, so the list is written when the process is set up (AI Fluency for builders, Claude Academy).
5. Named human steps keep the process running when the tool is down, because someone still knows how to do them.
6. The boundary that holds is a named step with a named owner, and "be careful" names no step.

## References

- Building effective human-agent teams, Claude Academy, https://academy.claude.com/courses/building-effective-human-agent-teams
- AI Fluency for builders, Claude Academy, https://academy.claude.com/courses/ai-fluency-for-builders
- Regulation (EU) 2024/1689 (Artificial Intelligence Act), consolidated text of 27 July 2026, EUR-Lex, https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=CELEX:02024R1689-20260727