Case study

The work always arrives.
Most of it never needed a person.

How an agentic harness takes over tasks, then functions, then whole loops. And what that is worth this year, not next cycle.


SUBJECT
Agentic workflows in a working company

DOMAIN
Operations, back office, reporting

SHIFT
From tools people use to work that runs itself

METHOD
One loop at a time. Owned, traced, fenced.

TaskTask automated

4 hands4 hands still needed

  1. arrives
  2. read
  3. matched
  4. approved
  5. filed

LoopLoop handed over

running closed

  1. arrives
  2. read
  3. matched
  4. approved
  5. filed

You did not hire your people to chase attachments. You hired them to decide things.

The problem

Your company runs on work nobody budgeted for.

Someone assembles the numbers before the meeting can happen. Someone chases the missing attachment. Someone reads forty pages so three people do not have to.

None of it appears on an org chart. All of it is paid for.

And because it is nobody’s job, it lands on your best people. The ones who can be trusted to be thorough get the work that needs care but no judgement.

Coordination work is invisible. It only becomes visible when a good person leaves.

The shift

A tool waits to be used. A harness goes to work.

Most companies met AI as a text box. Someone types a question, gets a good answer, everyone is impressed, nothing is different on Monday. The value stops at the edge of the conversation.

A harness is the frame around the intelligence. Reach into the systems where work already lives. Memory of your context and your last decision. Standing instructions about what to do, when, and where to stop.

Put those three in place and it stops being something you consult. It starts carrying.

You stop prompting. You start delegating.

What it takes over

Six things a harness does that a tool cannot.

  1. 1

    It shows up where the work already lives.

    No new screen. No migration. No login your team forgets by Thursday. Work arrives where it always arrived. Something is now sitting there to receive it.

  2. 2

    It remembers your house.

    Your suppliers, your naming, your exceptions, the rule you set in March and why. A tool starts from zero every morning. A harness accumulates.

  3. 3

    It turns a solved problem into standing capability.

    The first time it is a task. After that it is a function. Explained once, it holds, and it gets cheaper every run. Work does not normally behave that way.

  4. 4

    It works when nobody asked.

    The briefing finished before the first laptop opens. The month-end preparation that starts itself. The contract that auto-renews in thirty days if no one speaks. Unattended work removes the part of the job that depends on somebody remembering.

  5. 5

    It splits itself into a team.

    One brief becomes several workers at once. One reads, one checks, one drafts, one files. Volume stops being a scheduling problem, and the dull jobs stop being postponed.

  6. 6

    It works inside a fence.

    Defined access, defined limits, defined approval points. Draft but do not send. Prepare but do not pay. Every step leaves a trace, so “what exactly did it do?” always has an answer.

Point six is not the small print. It is why the other five are allowed near a real company.

Three choices underneath

Why this holds up, and where it usually fails.

01

Hand over a loop, not a task.

Automating a task saves minutes and impresses nobody by week two. A loop is the whole run, from arrival to finished. Close the loop and a function leaves your attention entirely.

02

The volume is automated. The judgement is not.

Machines are relentless and consistent. People are good at the odd, ambiguous, “that is not right” call. So people are kept for exactly that, with the context already assembled.

03

One owner, or it dies in the first busy quarter.

Every failed initiative fails the same way. It belonged to everyone, so it belonged to no one. Someone senior owns the outcome, not the technology. No name, no start.

What this is not

Four things worth saying out loud.

  • Not a headcount argument.

    The first return is capacity, not redundancy.

  • Not a moonshot.

    The first loop is weeks, not quarters. Longer than that means the loop was chosen badly.

  • Not a chatbot for staff to try.

    Nobody has to become good at prompting. The work simply comes back finished.

  • Not permission to skip governance.

    What it may see, what it may do, what it may never do alone. Leadership decisions, not installation settings.

The technology rarely disappoints. The decision about what to hand over is what never gets made.

The value

What you get, and how fast.

  • Hours back where they cost the most.

    The people you would least like to lose stop spending afternoons on assembly.

  • The standard stops depending on the day.

    Your best version becomes the only version. In August as well as in March.

  • Capacity without onboarding.

    Peaks get absorbed. The seasonal scramble stops being a hiring conversation.

  • The state of the work stops being folklore.

    You can see what is done, what is waiting, and why. That is worth something before a single hour is saved.

  • Small enough to prove, cheap enough to stop.

    One loop, one owner, one measure agreed up front. Very little in this category is that easy to falsify.

A side effect

You find out how your company actually works.

To hand a loop over, you have to say what it is. Who starts it. What finished means. Who approves.

Most companies discover the loop was never written down, that three people do it three ways, and that the step everyone works around has been there for years.

That map is worth having even if you automate nothing. It is usually the moment the room stops discussing AI and starts discussing the business.

Before you hand anything over

Five questions, in this order.

  1. Q1

    Why this loop?

    Frequent, rule-heavy and hated. Not the most visible one.

  2. Q2

    Why now?

    What waiting two quarters costs.

  3. Q3

    Who owns it?

    A name with authority and time.

  4. Q4

    What would prove it works?

    One number, agreed before you start.

  5. Q5

    What would make us stop?

    Decided while everyone is still calm.

Answer those five and the build is the easy part.

The work will keep arriving. That was never the problem. The problem is who it lands on, and what they are not doing while they handle it.

That is the difference between having AI and putting it to work. We pick the first loop, and we build it.

Reply within two working days.