Your AI agent needs a manager, not just a prompt

You're reading The AI Directive - practical AI intelligence for
business leaders. Subscribe here if
someone forwarded this.

AI agents are very easy to demonstrate.

Give one a goal, a few tools and a clean example, and it can look
astonishingly capable. It can read, summarise, draft a note, update a
file, create a task and keep going.

That is the exciting bit.

It is also not the bit I would worry about first.

Last issue was about choosing the right AI mode for the work: quick
tidy-up, deeper reasoning, or the approved corporate route. Agents raise
the next question. Not just which mode fits the task, but who owns the
work once the AI starts acting.

That sounds obvious until you watch the real workflows.

Someone asks an AI agent to sort the pipeline, prepare the supplier
review, clean the customer data, triage the inbox or keep a project
moving. It does five useful things and one questionable thing. It
updates a tracker. It drafts a message. It misses an exception. It
follows an old instruction. It treats a stale source as current. It
continues confidently.

Now ask the awkward question:

Who was managing that work?

Not who typed the prompt.

Who owned the process?

In the human world this is familiar. A junior analyst can do
excellent work, but somebody still checks the brief, agrees the scope,
reviews exceptions, signs off the output and decides what happens when
the work affects a customer, supplier, employee or financial
decision.

We do not call that bureaucracy.

We call it management.

AI agents need the same operating common sense. Not because they are
junior employees. They are not people. But because they are starting to
sit in the same messy space where real work happens: systems, emails,
spreadsheets, approvals, deadlines, handovers and consequences.

The trap is treating an agent like a clever prompt that happens to
run for longer.

That is too small.

An agent is closer to a managed workflow. It needs a job, boundaries,
escalation rules and a clear definition of done.

Managing the AI flow

Scope

Boundaries

Escalation

Proof

Owner

Brief

→

Agent work

→

Exception check

→

Human sign-off

The questions are basic:

What is it allowed to read?

What is it allowed to change?

When must it stop?

What needs human approval?

Where does it record proof?

Who checks that the output was actually useful?

These are not glamorous questions. That is partly why they get
skipped. But they are the questions that turn an agent from a demo into
a business process.

The organisations that get value from agents will not be the ones
with the most impressive demo library. They will be the ones that turn
agent work into managed work.

That means giving agents narrower jobs before broad missions.

Start with "compare these two records and flag conflicts", not
"manage supplier onboarding".

Start with "draft the weekly project risk summary from these approved
sources", not "keep me updated on everything".

Start with "prepare the first pass and show what you used", not
"decide what we should do".

There is a place for more autonomous agents. But autonomy should be
earned by evidence, not assumed because the demo looked good.

The useful model is not:

The useful model is:

That is less exciting than the sales deck version. It is also much
more likely to survive contact with a real organisation.

If an AI agent is acting inside your business, ask who manages
it.

If the answer is "nobody, it just follows the prompt", you do not yet
have an operating model.

You have a clever unmanaged process.

And business has had enough of those already.

Before putting an agent near real work, I would write a one-page
brief.

Not a policy. A brief.

It should cover: purpose, approved sources, allowed actions,
forbidden actions, escalation triggers, proof required, owner, review
rhythm and stop condition.

For example:

That is not over-engineering.

It is the minimum viable line manager.

1. Agents need narrower jobs than people think.
Broad goals create movement. Narrow jobs create evidence.

2. "Human in the loop" is too vague. Say which
human, at which decision point, with what authority.

3. The best early agent workflows are structured,
not fully autonomous: first pass, exception detection, proof trail,
human sign-off.

If you only do one thing this week, pick one AI agent or automation
idea in your organisation and write the stop rules before the
prompt.

What can it not do?

When must it ask?

Who owns the result?

That is where trust starts.

See you Tuesday.

@TobySmithFirth on
YouTube

These are tools I use personally. Affiliate links marked - I earn
a small commission if you sign up, at no extra cost to you.

Some links in this issue are affiliate links. I only recommend
tools I actually use.