Examples of Automation at Work: What AI Actually Automates

Automation covers a wide range of things — a factory robot welding a car frame and a chatbot drafting a customer reply are both automation, and they have almost nothing else in common. This is about the second kind: the software and AI automation showing up in ordinary desk jobs right now, with concrete examples of what it actually does, a worked one in full, and where it needs a person watching it.

Worth a moment of scale first, because the physical kind sets a useful contrast. The International Federation of Robotics recorded more than 4 million industrial robots working in factories worldwide, with over 500,000 new units installed in 2024 alone. Almost all of that concentrates in manufacturing tasks like welding, assembly and material handling — a small, specific slice of the economy. The automation reshaping office work is a different animal: less visible, spread across far more job titles, and — unlike a welding robot — usually a draft a person still checks rather than a finished action.

Five real examples

• —Ticket triage. Reading a raw IT support description and routing it to the right queue with a priority label, instead of a person reading every ticket cold. IT process automation covers this one in full, including the categories worth automating and the ones to leave alone.
• —Lead scoring and follow-up drafting. Turning a lead’s account and engagement data into a scoring rationale, and a generic sales inquiry into a first-draft, on-brand reply. B2B marketing automation works through a real before-and-after version of this.
• —Financial reconciliation and categorisation. Xero’s JAX agent, announced 3 September 2025, is pitched at automating bank reconciliations and data entry. Intuit’s own agent line, announced 1 July 2025, claims an Accounting Agent that assists with bookkeeping and reconciliation. Read the specific time-saved figures from either announcement as vendor marketing about its own product, not an independent study — but the shape of the task being targeted is real and consistent across both.
• —Meeting and call summarising. Turning a transcript into a short, structured list of decisions and next steps, rather than a full recording nobody replays.
• —Report drafting from numbers someone already produced. Turning a finished set of figures into the narrative paragraph a manager expects, the same task how to create a report in Excel covers from the spreadsheet side.

The pattern across every one of these: a model produces a draft, a category, or a summary, and a person still decides whether it is right before it goes anywhere that matters. None of them involve the model taking the final action on its own.

The test for picking a good candidate

Find the repetitive part sets out four questions worth running before automating anything: does it happen at least weekly, is the input predictable in shape, is a mistake cheap and visible, and could you explain the steps to a new hire in five minutes? Ticket triage and reply drafting clear all four easily. "Decide whether to approve this refund" usually does not — the input varies too much and a wrong call costs real money, so that one stays with a person regardless of how good the drafting gets elsewhere.

A worked example: turning meeting notes into next steps

Weak: “Summarise this meeting.”

Better: “Here is a raw transcript of a 30-minute client check-in. Produce three sections: Decisions made, Open questions, and Next steps with an owner and a rough date for each. Only include something under Decisions if it was explicitly agreed in the transcript — if it was discussed but not settled, put it under Open questions instead.”

The second version does two things the first does not: it gives the model a fixed structure to fill rather than an open-ended request, and it draws an explicit line between something agreed and something merely discussed. That line is the one a model will blur on its own if you do not name it, because a decided point and a debated point can read equally confident in the output. How to write a prompt that works on the first try covers the general version of specifying a task this precisely.

Where it goes wrong

The failure mode is consistent across all five examples above: the output looks equally confident whether it is right or not. What AI is actually bad at covers the general pattern, and it is exactly why a rushed rollout of any of these tends to go quiet before anyone notices it degraded — a summarised meeting with an invented decision reads no differently from one that is accurate, until someone acts on the wrong one.

The NIST AI Risk Management Framework frames the fix as ongoing monitoring rather than a one-time check: decide in advance what you would look for if the automation started quietly failing — a spike in a particular ticket category, a client who says a summary missed something they clearly said, a reconciliation exception nobody is reviewing — and check for it on a schedule, not only when something visibly breaks.

Automate the draft. Keep the decision that actually closes something.

What the job market says about it

That pattern shows up in the labour-market data too. The International Labour Organization’s global analysis and separate job-posting data from Indeed’s Hiring Lab both find demand shifting toward people who can run and check these systems, rather than disappearing outright — which matches every example above being a draft a person reviews, not a role removed entirely.

What to do Monday

Pick one task from the list above that happens at least weekly on your own desk, and write the specific, structured version of the prompt the way the meeting-notes example does — a fixed set of sections or categories, and an explicit instruction about what not to assume. Run it against a case you already know the right answer for before pointing it at anything live, and decide up front what you would check weekly to notice if it started quietly getting things wrong.

Coursium teaches exactly this kind of practical skill — writing a request specifically enough that the output is actually usable, and knowing which decisions stay with a person. Stay ahead of AI by learning the tools on your phone.