What loop engineering means if you run a function

A term from software engineering reached business writing this year. Loop engineering means designing the repeating cycle a piece of work runs through, rather than typing a fresh instruction every time. At your desk it does one useful thing. It tells you which of your monthly processes is worth building once, and which is just typing.

As of 13 September 2026.

What loop engineering means in plain words

Loop engineering is the design of a repeating workflow with checks inside it. The unit of work is the cycle, not the single question. IBM's definition is the design of the workflows, or loops, that carry an AI toward a goal you set. Its breakdown has four stages: a goal with a clear stopping point, an action, a look at what the action produced, and an adjustment before the cycle runs again.

Prompt engineering was about the words in one message. Loop engineering is about the process those words sit inside. Almost all of the writing on it is aimed at people who ship code, which is why it reads as somebody else's problem. The shape underneath is ordinary operating work.

Your monthly review pack is a loop

The monthly review pack is the clearest loop in a senior function. The numbers close, you pull them, you compare them to plan and to last year, you write what moved and why, you list what you want from each region, and it goes into a deck. Same trigger, same inputs, same output shape, every month. In an FMCG or pharma function that pack has looked the same for years, and the part that eats the week is assembly, not judgment.

Write the loop down before you open any tool. Four lines. The trigger is the day the numbers close. The inputs are the sales file, the plan, last year, the stock report. The steps are the comparisons you always run, in the order you always run them. The stopping rule is the line people skip: the pack is done when every figure traces to a source file, and every region past a variance threshold you set has a named cause beside it.

Without a stopping rule you keep poking at the same pack.

How do I learn AI for business if I run a function?

Start from the loop you already run, not from a course outline. My own start was one question I put to my own work each day: which business problem do I have to solve today, and can AI get me there faster. A loop is that same question arriving on a fixed date every month.

So the learning is narrow and it is yours. Take last month's pack. Write the brief that would have produced it: the role, the context, the task, the format, the constraints. Run it on the real files, not a sample. Compare the two packs line by line. Mark every place the output was wrong, vague, or unsourced. That mark-up is the check, and the check is worth more than the brief.

Engineered loops in software are built the same way round. OpenAI's agent documentation describes a run that keeps looping until it reaches a real stopping point. It treats a human approval as a paused run, resumed later from the same state: https://developers.openai.com/api/docs/guides/agents/running-agents

Your signature on the review pack is that pause. Nothing in the term removes it.

What to ignore in the loop engineering writing

Ignore the job-title part. Nothing here asks you to become an engineer or to hire one. Ignore the autonomy claims next. Pages that promise a system running with nobody in it are describing code that can test itself, and your variance commentary has no unit tests. The commentary is the product, and your name is on it.

Ignore tool shopping as well. The first version of your loop is four lines on paper. A tool you already have will run it badly, and badly is enough to show you which of the four lines you got wrong.

How do I get better with AI from a standing start?

You get better on one loop, run for one month, with every output marked up against the source file it came from.

How do I know when the output is wrong?

Pick three claims in the draft and check them against the source file before you read anything else. A figure that traces nowhere, a cause asserted without a number behind it, a region summarized with no data for it. Wrong output is usually confident and specific, so check the specifics first.

Can I put the sales file into a public AI tool?

Ask whoever owns the AI policy for the line in writing, because until somebody states it you are guessing on behalf of the company. Until you have that, practice on figures you would already email outside the building, or strip names and codes and keep the structure. The method works on anonymized data.

Do I start with myself or with my team?

Start with yourself, on one loop, for one month. A bad output that sounds sure of itself reads like a good one until you have marked up a few of your own. Until you can do that, you have no way to review what your team brings you. One loop done properly teaches more.